ci: update path-filtering orb to 1.3.0 (#3588)

Signed-off-by: Jared Van Bortel <jared@nomic.ai>

.circleci/config.yml

@@ -1,13 +1,17 @@
 version: 2.1
 setup: true
 orbs:
-  path-filtering: circleci/path-filtering@1.1.0
+  path-filtering: circleci/path-filtering@1.3.0
 
 workflows:
   version: 2.1
   generate-config:
     jobs:
       - path-filtering/filter:
+          filters:
+            tags:
+              only:
+                - /.*/
           base-revision: main
           config-path: .circleci/continue_config.yml
           mapping: |

.gitmodules

@@ -17,9 +17,9 @@
 [submodule "gpt4all-chat/deps/QXlsx"]
 	path = gpt4all-chat/deps/QXlsx
 	url = https://github.com/nomic-ai/QXlsx.git
-[submodule "gpt4all-chat/deps/Jinja2Cpp"]
-	path = gpt4all-chat/deps/Jinja2Cpp
-	url = https://github.com/nomic-ai/jinja2cpp.git
-[submodule "gpt4all-chat/deps/rapidjson"]
-	path = gpt4all-chat/deps/rapidjson
-	url = https://github.com/nomic-ai/rapidjson.git
+[submodule "gpt4all-chat/deps/minja"]
+	path = gpt4all-chat/deps/minja
+	url = https://github.com/nomic-ai/minja.git
+[submodule "gpt4all-chat/deps/json"]
+	path = gpt4all-chat/deps/json
+	url = https://github.com/nlohmann/json.git

MAINTAINERS.md

@@ -72,6 +72,6 @@ Discord: `@Tim453`
 - Flatpak
 
 Jack ([@wuodoo](https://github.com/wuodoo))<br/>
-E-mail: 2296103047@qq.com><br/>
+E-mail: 2296103047@qq.com<br/>
 Discord: `@mikage`
 - zh\_CN translation

README.md

@@ -1,5 +1,9 @@
 <h1 align="center">GPT4All</h1>
 
+<p align="center">
+  Now with support for DeepSeek R1 Distillations
+</p>
+
 <p align="center">
   <a href="https://www.nomic.ai/gpt4all">Website</a> &bull; <a href="https://docs.gpt4all.io">Documentation</a> &bull; <a href="https://discord.gg/mGZE39AS3e">Discord</a> &bull; <a href="https://www.youtube.com/watch?v=gQcZDXRVJok">YouTube Tutorial</a>
 </p>
@@ -31,6 +35,11 @@ GPT4All is made possible by our compute partner <a href="https://www.paperspace.
     <img src="gpt4all-bindings/python/docs/assets/windows.png" style="height: 1em; width: auto" /> Windows Installer
   </a> &mdash;
 </p>
+<p>
+  &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-win64-arm.exe">
+    <img src="gpt4all-bindings/python/docs/assets/windows.png" style="height: 1em; width: auto" /> Windows ARM Installer
+  </a> &mdash;
+</p>
 <p>
   &mdash; <a href="https://gpt4all.io/installers/gpt4all-installer-darwin.dmg">
     <img src="gpt4all-bindings/python/docs/assets/mac.png" style="height: 1em; width: auto" /> macOS Installer
@@ -42,10 +51,16 @@ GPT4All is made possible by our compute partner <a href="https://www.paperspace.
   </a> &mdash;
 </p>
 <p>
-  Windows and Linux require Intel Core i3 2nd Gen / AMD Bulldozer, or better. x86-64 only, no ARM.
+  The Windows and Linux builds require Intel Core i3 2nd Gen / AMD Bulldozer, or better.
 </p>
 <p>
-  macOS requires Monterey 12.6 or newer. Best results with Apple Silicon M-series processors.
+  The Windows ARM build supports Qualcomm Snapdragon and Microsoft SQ1/SQ2 processors.
+</p>
+<p>
+  The Linux build is x86-64 only (no ARM).
+</p>
+<p>
+  The macOS build requires Monterey 12.6 or newer. Best results with Apple Silicon M-series processors.
 </p>
 
 See the full [System Requirements](gpt4all-chat/system_requirements.md) for more details.

gpt4all-backend/CMakeLists.txt

@@ -69,7 +69,7 @@ if (LLMODEL_CUDA)
     cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
 
     # Defaults must be set before enable_language(CUDA).
-    # Keep this in sync with the arch list in ggml/src/CMakeLists.txt.
+    # Keep this in sync with the arch list in ggml/src/CMakeLists.txt (plus 5.0 for non-F16 branch).
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
         # 52 == lowest CUDA 12 standard
         # 60 == f16 CUDA intrinsics
@@ -78,7 +78,7 @@ if (LLMODEL_CUDA)
         if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75") # needed for f16 CUDA intrinsics
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75") # lowest CUDA 12 standard + lowest for integer intrinsics
+            set(CMAKE_CUDA_ARCHITECTURES "50;52;61;70;75") # lowest CUDA 12 standard + lowest for integer intrinsics
             #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
         endif()
     endif()

gpt4all-backend/deps/llama.cpp-mainline

@@ -1 +1 @@
-Subproject commit 58a55efc4ae5dd3bc12887d47981faa7136027af
+Subproject commit 11f734c3b0334dbae4823b4a7467764e447fc6d6

gpt4all-backend/src/llamamodel.cpp

@@ -53,6 +53,8 @@ static const std::vector<const char *> KNOWN_ARCHES {
     "gpt2",
     // "gptj", -- no inference code
     "gptneox",
+    "granite",
+    "granitemoe",
     "mpt",
     "baichuan",
     "starcoder",
@@ -80,6 +82,7 @@ static const std::vector<const char *> KNOWN_ARCHES {
     "command-r",
     // "dbrx", -- 16x12B parameters
     "olmo",
+    "olmoe",
     "openelm",
     // "arctic", -- 10B+128x3.66B parameters
     "deepseek2",

gpt4all-backend/src/llmodel.cpp

@@ -140,9 +140,14 @@ const std::vector<LLModel::Implementation> &LLModel::Implementation::implementat
             std::string path;
             // Split the paths string by the delimiter and process each path.
             while (std::getline(ss, path, ';')) {
-                std::u8string u8_path(path.begin(), path.end());
+                fs::directory_iterator iter;
+                try {
+                    iter = fs::directory_iterator(std::u8string(path.begin(), path.end()));
+                } catch (const fs::filesystem_error &) {
+                    continue; // skip nonexistent path
+                }
                 // Iterate over all libraries
-                for (const auto &f : fs::directory_iterator(u8_path)) {
+                for (const auto &f : iter) {
                     const fs::path &p = f.path();
 
                     if (p.extension() != LIB_FILE_EXT) continue;

gpt4all-chat/CHANGELOG.md

@@ -4,11 +4,62 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
+## [3.10.0] - 2025-02-24
+
+### Added
+- Whitelist Granite (non-MoE) model architecture (by [@ThiloteE](https://github.com/ThiloteE) in [#3487](https://github.com/nomic-ai/gpt4all/pull/3487))
+- Add support for CUDA compute 5.0 GPUs such as the GTX 750 ([#3499](https://github.com/nomic-ai/gpt4all/pull/3499))
+- Add a Remote Providers tab to the Add Model page ([#3506](https://github.com/nomic-ai/gpt4all/pull/3506))
+
+### Changed
+- Substitute prettier default templates for OLMoE 7B 0924/0125 and Granite 3.1 3B/8B (by [@ThiloteE](https://github.com/ThiloteE) in [#3471](https://github.com/nomic-ai/gpt4all/pull/3471))
+- Build with LLVM Clang 19 on macOS and Ubuntu ([#3500](https://github.com/nomic-ai/gpt4all/pull/3500))
+
+### Fixed
+- Fix several potential crashes ([#3465](https://github.com/nomic-ai/gpt4all/pull/3465))
+- Fix visual spacing issues with deepseek models ([#3470](https://github.com/nomic-ai/gpt4all/pull/3470))
+- Add missing strings to Italian translation (by [@Harvester62](https://github.com/Harvester62) in [#3496](https://github.com/nomic-ai/gpt4all/pull/3496))
+- Update Simplified Chinese translation (by [@Junior2Ran](https://github.com/Junior2Ran) in [#3467](https://github.com/nomic-ai/pull/3467))
+
+## [3.9.0] - 2025-02-04
+
+### Added
+- Whitelist OLMoE and Granite MoE model architectures (no Vulkan) (by [@ThiloteE](https://github.com/ThiloteE) in [#3449](https://github.com/nomic-ai/gpt4all/pull/3449))
+
+### Fixed
+- Fix "index N is not a prompt" when using LocalDocs with reasoning ([#3451](https://github.com/nomic-ai/gpt4all/pull/3451))
+- Work around rendering artifacts on Snapdragon SoCs with Windows ([#3450](https://github.com/nomic-ai/gpt4all/pull/3450))
+- Prevent DeepSeek-R1 reasoning from appearing in chat names and follow-up questions ([#3458](https://github.com/nomic-ai/gpt4all/pull/3458))
+- Fix LocalDocs crash on Windows ARM when reading PDFs ([#3460](https://github.com/nomic-ai/gpt4all/pull/3460))
+- Fix UI freeze when chat template is `{#` ([#3446](https://github.com/nomic-ai/gpt4all/pull/3446))
+
+## [3.8.0] - 2025-01-30
+
+### Added
+- Support DeepSeek-R1 Qwen models ([#3431](https://github.com/nomic-ai/gpt4all/pull/3431))
+- Support for think tags in the GUI ([#3440](https://github.com/nomic-ai/gpt4all/pull/3440))
+- Support specifying SHA256 hash in models3.json instead of MD5 ([#3437](https://github.com/nomic-ai/gpt4all/pull/3437))
+
+### Changed
+- Use minja instead of Jinja2Cpp for significantly improved template compatibility ([#3433](https://github.com/nomic-ai/gpt4all/pull/3433))
+
+### Fixed
+- Fix regression while using localdocs with server API ([#3410](https://github.com/nomic-ai/gpt4all/pull/3410))
+- Don't show system messages in server chat view ([#3411](https://github.com/nomic-ai/gpt4all/pull/3411))
+- Fix `codesign --verify` failure on macOS ([#3413](https://github.com/nomic-ai/gpt4all/pull/3413))
+- Code Interpreter: Fix console.log not accepting a single string after v3.7.0 ([#3426](https://github.com/nomic-ai/gpt4all/pull/3426))
+- Fix Phi 3.1 Mini 128K Instruct template (by [@ThiloteE](https://github.com/ThiloteE) in [#3412](https://github.com/nomic-ai/gpt4all/pull/3412))
+- Don't block the gui thread for reasoning ([#3435](https://github.com/nomic-ai/gpt4all/pull/3435))
+- Fix corruption of unicode in output of reasoning models ([#3443](https://github.com/nomic-ai/gpt4all/pull/3443))
+
 ## [3.7.0] - 2025-01-21
 
 ### Added
 - Add support for the Windows ARM64 target platform (CPU-only) ([#3385](https://github.com/nomic-ai/gpt4all/pull/3385))
 
+### Changed
+- Update from Qt 6.5.1 to 6.8.1 ([#3386](https://github.com/nomic-ai/gpt4all/pull/3386))
+
 ### Fixed
 - Fix the timeout error in code interpreter ([#3369](https://github.com/nomic-ai/gpt4all/pull/3369))
 - Fix code interpreter console.log not accepting multiple arguments ([#3371](https://github.com/nomic-ai/gpt4all/pull/3371))
@@ -17,7 +68,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - Save chats on quit, even if the window isn't closed first ([#3387](https://github.com/nomic-ai/gpt4all/pull/3387))
 - Add chat template replacements for five new models and fix EM German Mistral ([#3393](https://github.com/nomic-ai/gpt4all/pull/3393))
 - Fix crash when entering `{{ a["foo"(` as chat template ([#3394](https://github.com/nomic-ai/gpt4all/pull/3394))
-- Sign the maintenance tool on macOS to prevent crash on Sonoma ([#3391](https://github.com/nomic-ai/gpt4all/pull/3391))
+- Sign the maintenance tool on macOS to prevent crash on Sequoia ([#3391](https://github.com/nomic-ai/gpt4all/pull/3391))
 - Jinja2Cpp: Fix operator precedence in 'not X is defined' ([#3402](https://github.com/nomic-ai/gpt4all/pull/3402))
 
 ## [3.6.1] - 2024-12-20
@@ -261,6 +312,10 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - Fix several Vulkan resource management issues ([#2694](https://github.com/nomic-ai/gpt4all/pull/2694))
 - Fix crash/hang when some models stop generating, by showing special tokens ([#2701](https://github.com/nomic-ai/gpt4all/pull/2701))
 
+[3.10.0]: https://github.com/nomic-ai/gpt4all/compare/v3.9.0...v3.10.0
+[3.9.0]: https://github.com/nomic-ai/gpt4all/compare/v3.8.0...v3.9.0
+[3.8.0]: https://github.com/nomic-ai/gpt4all/compare/v3.7.0...v3.8.0
+[3.7.0]: https://github.com/nomic-ai/gpt4all/compare/v3.6.1...v3.7.0
 [3.6.1]: https://github.com/nomic-ai/gpt4all/compare/v3.6.0...v3.6.1
 [3.6.0]: https://github.com/nomic-ai/gpt4all/compare/v3.5.3...v3.6.0
 [3.5.3]: https://github.com/nomic-ai/gpt4all/compare/v3.5.2...v3.5.3

gpt4all-chat/CMakeLists.txt

@@ -3,13 +3,17 @@ cmake_minimum_required(VERSION 3.25)  # for try_compile SOURCE_FROM_VAR
 include(../common/common.cmake)
 
 set(APP_VERSION_MAJOR 3)
-set(APP_VERSION_MINOR 7)
-set(APP_VERSION_PATCH 0)
+set(APP_VERSION_MINOR 10)
+set(APP_VERSION_PATCH 1)
 set(APP_VERSION_BASE "${APP_VERSION_MAJOR}.${APP_VERSION_MINOR}.${APP_VERSION_PATCH}")
-set(APP_VERSION "${APP_VERSION_BASE}")
+set(APP_VERSION "${APP_VERSION_BASE}-dev0")
 
 project(gpt4all VERSION ${APP_VERSION_BASE} LANGUAGES CXX C)
 
+if (CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
+  set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR}/install CACHE PATH "..." FORCE)
+endif()
+
 if(APPLE)
   option(BUILD_UNIVERSAL "Build a Universal binary on macOS" OFF)
   if(BUILD_UNIVERSAL)
@@ -31,6 +35,8 @@ option(GPT4ALL_SIGN_INSTALL "Sign installed binaries and installers (requires si
 option(GPT4ALL_GEN_CPACK_CONFIG "Generate the CPack config.xml in the package step and nothing else." OFF)
 set(GPT4ALL_USE_QTPDF "AUTO" CACHE STRING "Whether to Use QtPDF for LocalDocs. If OFF or not available on this platform, PDFium is used.")
 set_property(CACHE GPT4ALL_USE_QTPDF PROPERTY STRINGS AUTO ON OFF)
+set(GPT4ALL_FORCE_D3D12 "AUTO" CACHE STRING "Whether to use Direct3D 12 as the Qt scene graph backend. Defaults to ON on Windows ARM.")
+set_property(CACHE GPT4ALL_FORCE_D3D12 PROPERTY STRINGS AUTO ON OFF)
 
 include(cmake/cpack_config.cmake)
 
@@ -47,7 +53,7 @@ set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 set(CMAKE_CXX_STANDARD 23)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 if (MSVC)
-    # Enable accurate __cplusplus macro to fix errors in Jinja2Cpp
+    # Enable accurate __cplusplus macro
     add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/Zc:__cplusplus>)
 endif()
 
@@ -86,12 +92,6 @@ include_directories("${CMAKE_CURRENT_BINARY_DIR}")
 set(CMAKE_AUTOMOC ON)
 set(CMAKE_AUTORCC ON)
 
-# Generate a header file with the version number
-configure_file(
-  "${CMAKE_CURRENT_SOURCE_DIR}/cmake/config.h.in"
-  "${CMAKE_CURRENT_BINARY_DIR}/config.h"
-)
-
 set(CMAKE_FIND_PACKAGE_TARGETS_GLOBAL ON)
 set(GPT4ALL_QT_COMPONENTS Core HttpServer LinguistTools Quick QuickDialogs2 Sql Svg)
 set(GPT4ALL_USING_QTPDF OFF)
@@ -104,7 +104,7 @@ elseif (GPT4ALL_USE_QTPDF MATCHES "^(ON|AUTO)$")
     set(GPT4ALL_USING_QTPDF ON)
     list(APPEND GPT4ALL_QT_COMPONENTS Pdf)
 endif()
-find_package(Qt6 6.5 COMPONENTS ${GPT4ALL_QT_COMPONENTS} REQUIRED)
+find_package(Qt6 6.8 COMPONENTS ${GPT4ALL_QT_COMPONENTS} REQUIRED)
 
 if (QT_KNOWN_POLICY_QTP0004)
     qt_policy(SET QTP0004 NEW)  # generate extra qmldir files on Qt 6.8+
@@ -126,10 +126,24 @@ message(STATUS "Qt 6 root directory: ${Qt6_ROOT_DIR}")
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-if (CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
-  set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR}/install CACHE PATH "..." FORCE)
+set(GPT4ALL_CONFIG_FORCE_D3D12 -1)
+if (NOT CMAKE_SYSTEM_NAME MATCHES Windows OR Qt6_VERSION VERSION_LESS "6.6")
+    # Direct3D 12 is not available.
+    if (GPT4ALL_FORCE_D3D12 STREQUAL "ON")
+        message(FATAL_ERROR "Cannot use Direct3D 12 on this platform.")
+    endif()
+elseif (GPT4ALL_FORCE_D3D12 MATCHES "^(ON|AUTO)$")
+    if (GPT4ALL_FORCE_D3D12 STREQUAL "ON" OR CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|AARCH64|arm64|ARM64)$")
+        set(GPT4ALL_CONFIG_FORCE_D3D12 1)
+    endif()
 endif()
 
+# Generate a header file for configuration
+configure_file(
+    "${CMAKE_CURRENT_SOURCE_DIR}/src/config.h.in"
+    "${CMAKE_CURRENT_BINARY_DIR}/config.h"
+)
+
 add_subdirectory(deps)
 add_subdirectory(../gpt4all-backend llmodel)
 
@@ -252,6 +266,7 @@ qt_add_qml_module(chat
       qml/AddModelView.qml
       qml/AddGPT4AllModelView.qml
       qml/AddHFModelView.qml
+      qml/AddRemoteModelView.qml
       qml/ApplicationSettings.qml
       qml/ChatDrawer.qml
       qml/ChatCollapsibleItem.qml
@@ -300,6 +315,7 @@ qt_add_qml_module(chat
       qml/MyTextField.qml
       qml/MyToolButton.qml
       qml/MyWelcomeButton.qml
+      qml/RemoteModelCard.qml
     RESOURCES
       icons/antenna_1.svg
       icons/antenna_2.svg
@@ -330,6 +346,7 @@ qt_add_qml_module(chat
       icons/gpt4all-48.png
       icons/gpt4all.svg
       icons/gpt4all_transparent.svg
+      icons/groq.svg
       icons/home.svg
       icons/image.svg
       icons/info.svg
@@ -337,12 +354,14 @@ qt_add_qml_module(chat
       icons/left_panel_open.svg
       icons/local-docs.svg
       icons/models.svg
+      icons/mistral.svg
       icons/network.svg
       icons/nomic_logo.svg
       icons/notes.svg
       icons/paperclip.svg
       icons/plus.svg
       icons/plus_circle.svg
+      icons/openai.svg
       icons/recycle.svg
       icons/regenerate.svg
       icons/search.svg
@@ -437,7 +456,10 @@ else()
     target_link_libraries(chat PRIVATE pdfium)
 endif()
 target_link_libraries(chat
-    PRIVATE llmodel SingleApplication fmt::fmt duckx::duckx QXlsx jinja2cpp)
+    PRIVATE llmodel SingleApplication fmt::fmt duckx::duckx QXlsx)
+target_include_directories(chat PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/deps/json/include)
+target_include_directories(chat PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/deps/json/include/nlohmann)
+target_include_directories(chat PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/deps/minja/include)
 
 if (APPLE)
     target_link_libraries(chat PRIVATE ${COCOA_LIBRARY})
@@ -445,12 +467,18 @@ endif()
 
 # -- install --
 
+if (APPLE)
+    set(GPT4ALL_LIB_DEST bin/gpt4all.app/Contents/Frameworks)
+else()
+    set(GPT4ALL_LIB_DEST lib)
+endif()
+
 install(TARGETS chat DESTINATION bin COMPONENT ${COMPONENT_NAME_MAIN})
 
 install(
     TARGETS llmodel
-    LIBRARY DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
-    RUNTIME DESTINATION bin COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
+    LIBRARY DESTINATION ${GPT4ALL_LIB_DEST} COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
+    RUNTIME DESTINATION bin                 COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
 )
 
 # We should probably iterate through the list of the cmake for backend, but these need to be installed
@@ -473,8 +501,8 @@ endif()
 
 install(
     TARGETS ${MODEL_IMPL_TARGETS}
-    LIBRARY DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
-    RUNTIME DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
+    LIBRARY DESTINATION ${GPT4ALL_LIB_DEST} COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
+    RUNTIME DESTINATION lib                 COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
 )
 
 if(APPLE AND GPT4ALL_SIGN_INSTALL)
@@ -503,7 +531,7 @@ if (LLMODEL_CUDA)
         TARGETS llamamodel-mainline-cuda
                 llamamodel-mainline-cuda-avxonly
         RUNTIME_DEPENDENCY_SET llama-cuda-deps
-        LIBRARY DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .so/.dylib
+        LIBRARY DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .so
         RUNTIME DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN}  # .dll
     )
     if (WIN32)
@@ -520,9 +548,9 @@ endif()
 if (NOT GPT4ALL_USING_QTPDF)
     # Install PDFium
     if (WIN32)
-        install(FILES "${PDFium_LIBRARY}" DESTINATION bin COMPONENT ${COMPONENT_NAME_MAIN})
+        install(FILES ${PDFium_LIBRARY} DESTINATION bin                 COMPONENT ${COMPONENT_NAME_MAIN})  # .dll
     else()
-        install(FILES "${PDFium_LIBRARY}" DESTINATION lib COMPONENT ${COMPONENT_NAME_MAIN})
+        install(FILES ${PDFium_LIBRARY} DESTINATION ${GPT4ALL_LIB_DEST} COMPONENT ${COMPONENT_NAME_MAIN})  # .so/.dylib
     endif()
 endif()
 

gpt4all-chat/README.md

@@ -1,45 +0,0 @@
-# gpt4all-chat
-
-Cross platform Qt based GUI for GPT4All versions with GPT-J as the base
-model. NOTE: The model seen in the screenshot is actually a preview of a
-new training run for GPT4All based on GPT-J. The GPT4All project is busy
-at work getting ready to release this model including installers for all
-three major OS's. In the meantime, you can try this UI out with the original
-GPT-J model by following build instructions below.
-
-![image](https://user-images.githubusercontent.com/50458173/231464085-da9edff6-a593-410e-8f38-7513f75c8aab.png)
-
-## Install
-
-One click installers for macOS, Linux, and Windows at https://www.nomic.ai/gpt4all
-
-## Features
-
-* Cross-platform (Linux, Windows, MacOSX)
-* The UI is made to look and feel like you've come to expect from a chatty gpt
-* Check for updates so you can always stay fresh with latest models
-* Easy to install with precompiled binaries available for all three major desktop platforms
-* Multi-modal - Ability to load more than one model and switch between them
-* Multi-chat - a list of current and past chats and the ability to save/delete/export and switch between
-* Supports models that are supported by llama.cpp
-* Model downloader in GUI featuring many popular open source models
-* Settings dialog to change temp, top_p, min_p, top_k, threads, etc
-* Copy your conversation to clipboard
-* RAG via LocalDocs feature
-* Check for updates to get the very latest GUI
-
-## Building and running
-
-* Follow the visual instructions on the [build_and_run](build_and_run.md) page
-
-## Getting the latest
-
-If you've already checked out the source code and/or built the program make sure when you do a git fetch to get the latest changes and that you also do `git submodule update --init --recursive` to update the submodules. (If you ever run into trouble, deinitializing via `git submodule deinit -f .` and then initializing again via `git submodule update --init --recursive` fixes most issues)
-
-## Contributing
-
-* Pull requests welcome. See the feature wish list for ideas :)
-
-
-## License
-The source code of this chat interface is currently under a MIT license.

gpt4all-chat/cmake/config.h.in

@@ -1,6 +0,0 @@
-#ifndef CONFIG_H
-#define CONFIG_H
-
-#define APP_VERSION "@APP_VERSION@"
-
-#endif // CONFIG_H

gpt4all-chat/cmake/cpack_config.cmake

@@ -37,7 +37,6 @@ set(CPACK_PACKAGE_VERSION_PATCH ${PROJECT_VERSION_PATCH})
 set(CPACK_PACKAGE_HOMEPAGE_URL "https://www.nomic.ai/gpt4all")
 set(CPACK_PACKAGE_ICON "${CMAKE_CURRENT_SOURCE_DIR}/icons/gpt4all-48.png")
 set(CPACK_RESOURCE_FILE_LICENSE ${CMAKE_CURRENT_SOURCE_DIR}/LICENSE)
-set(CPACK_RESOURCE_FILE_README ${CMAKE_CURRENT_SOURCE_DIR}/README.md)
 set(CPACK_PACKAGE_EXECUTABLES "GPT4All")
 set(CPACK_CREATE_DESKTOP_LINKS "GPT4All")
 set(CPACK_IFW_PACKAGE_NAME "GPT4All")

gpt4all-chat/cmake/deploy-qt-mac.cmake.in

@@ -8,12 +8,6 @@ if (GPT4ALL_SIGN_INSTALL)
     set(MAC_NOTARIZE -sign-for-notarization=${GPT4ALL_SIGNING_ID})
 endif()
 execute_process(COMMAND ${MACDEPLOYQT} ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/bin/gpt4all.app -qmldir=${CMAKE_CURRENT_SOURCE_DIR} -verbose=2 ${MAC_NOTARIZE})
-file(GLOB MYLLAMALIBS ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/lib/libllama*)
-file(GLOB MYLLMODELLIBS ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/lib/libllmodel.*)
-file(COPY ${MYLLAMALIBS}
-     DESTINATION ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/bin/gpt4all.app/Contents/Frameworks)
-file(COPY ${MYLLMODELLIBS}
-     DESTINATION ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data/bin/gpt4all.app/Contents/Frameworks)
 file(COPY "${CMAKE_CURRENT_SOURCE_DIR}/icons/gpt4all-32.png"
      DESTINATION ${CPACK_TEMPORARY_INSTALL_DIRECTORY}/packages/${COMPONENT_NAME_MAIN}/data)
 file(COPY "${CMAKE_CURRENT_SOURCE_DIR}/icons/gpt4all-48.png"

gpt4all-chat/deps/CMakeLists.txt

@@ -15,43 +15,34 @@ add_subdirectory(DuckX)
 set(QT_VERSION_MAJOR 6)
 add_subdirectory(QXlsx/QXlsx)
 
-# forked dependency of Jinja2Cpp
-set(RAPIDJSON_BUILD_DOC                  OFF)
-set(RAPIDJSON_BUILD_EXAMPLES             OFF)
-set(RAPIDJSON_BUILD_TESTS                OFF)
-set(RAPIDJSON_ENABLE_INSTRUMENTATION_OPT OFF)
-add_subdirectory(rapidjson)
-
-add_subdirectory(Jinja2Cpp)
-
 if (NOT GPT4ALL_USING_QTPDF)
     # If we do not use QtPDF, we need to get PDFium.
-    set(GPT4ALL_PDFIUM_TAG "chromium/6954")
+    set(GPT4ALL_PDFIUM_TAG "chromium/6996")
     if (CMAKE_SYSTEM_NAME MATCHES Linux)
         FetchContent_Declare(
             pdfium
             URL "https://github.com/bblanchon/pdfium-binaries/releases/download/${GPT4ALL_PDFIUM_TAG}/pdfium-linux-x64.tgz"
-            URL_HASH "SHA256=69917fd9543befc6c806254aff6c8a604d9e7cd3999a3e70fc32b8690d372da2"
+            URL_HASH "SHA256=68b381b87efed539f2e33ae1e280304c9a42643a878cc296c1d66a93b0cb4335"
         )
     elseif (CMAKE_SYSTEM_NAME MATCHES Windows)
         if (CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|AMD64|amd64)$")
             FetchContent_Declare(
                 pdfium
                 URL "https://github.com/bblanchon/pdfium-binaries/releases/download/${GPT4ALL_PDFIUM_TAG}/pdfium-win-x64.tgz"
-                URL_HASH "SHA256=62ecac78fbaf658457beaffcc05eb147f493d435a2e1309e6a731808b4e80d38"
+                URL_HASH "SHA256=83e714c302ceacccf403826d5cb57ea39b77f393d83b8d5781283012774a9378"
             )
         elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|AARCH64|arm64|ARM64)$")
             FetchContent_Declare(
                 pdfium
                 URL "https://github.com/bblanchon/pdfium-binaries/releases/download/${GPT4ALL_PDFIUM_TAG}/pdfium-win-arm64.tgz"
-                URL_HASH "SHA256=a0b69014467f2b9824776c064920bc95359c9ba0d88793bdda1894a0f22206f8"
+                URL_HASH "SHA256=78e77e871453a4915cbf66fb381b951c9932f88a747c6b2b33c9f27ec2371445"
             )
         endif()
     elseif (CMAKE_SYSTEM_NAME MATCHES Darwin)
         FetchContent_Declare(
             pdfium
             URL "https://github.com/bblanchon/pdfium-binaries/releases/download/${GPT4ALL_PDFIUM_TAG}/pdfium-mac-univ.tgz"
-            URL_HASH "SHA256=7442f1dc6bef90898b2b7bd38dbec369ddd81bbf66c1c5aac3a1b60e107098f9"
+            URL_HASH "SHA256=e7577f3242ff9c1df50025f9615673a43601a201bc51ee4792975f98920793a2"
         )
     endif()
 

gpt4all-chat/deps/Jinja2Cpp

@@ -1 +0,0 @@
-Subproject commit ce10f783bae46ede6afa4b09a8a169ebe88a14d4

gpt4all-chat/deps/json

@@ -0,0 +1 @@
+Subproject commit 606b6347edf0758c531abb6c36743e09a4c48a84

gpt4all-chat/deps/minja

@@ -0,0 +1 @@
+Subproject commit e97bb2442cd6ab3d5bb5f5a3e8a1f7d6081d613b

gpt4all-chat/deps/rapidjson

@@ -1 +0,0 @@
-Subproject commit 9b547ef4bd86210ef084abc2790bd1ddfe66b592

gpt4all-chat/icons/groq.svg

@@ -0,0 +1,3 @@
+<?xml version="1.0" encoding="utf-8" ?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN" "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 26.3 26.3"><defs><style>.cls-1{fill:#f05237;}.cls-2{fill:#fff;}</style></defs><g id="Layer_2" data-name="Layer 2"><g id="Content"><circle class="cls-1" cx="13.15" cy="13.15" r="13.15"/><path class="cls-2" d="M13.17,6.88a4.43,4.43,0,0,0,0,8.85h1.45V14.07H13.17a2.77,2.77,0,1,1,2.77-2.76v4.07a2.74,2.74,0,0,1-4.67,2L10.1,18.51a4.37,4.37,0,0,0,3.07,1.29h.06a4.42,4.42,0,0,0,4.36-4.4V11.2a4.43,4.43,0,0,0-4.42-4.32"/></g></g></svg>

gpt4all-chat/icons/mistral.svg

@@ -0,0 +1 @@
+<svg viewBox="0 0 512 512" xmlns="http://www.w3.org/2000/svg" fill-rule="evenodd" clip-rule="evenodd" stroke-linejoin="round" stroke-miterlimit="2"><path d="M189.08 303.228H94.587l.044-94.446h94.497l-.048 94.446z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M283.528 397.674h-94.493l.044-94.446h94.496l-.047 94.446z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M283.575 303.228H189.08l.046-94.446h94.496l-.047 94.446z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M378.07 303.228h-94.495l.044-94.446h94.498l-.047 94.446zM189.128 208.779H94.633l.044-94.448h94.498l-.047 94.448zM378.115 208.779h-94.494l.045-94.448h94.496l-.047 94.448zM94.587 303.227H.093l.044-96.017h94.496l-.046 96.017z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M94.633 208.779H.138l.046-94.448H94.68l-.047 94.448z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M94.68 115.902H.185L.23 19.885h94.498l-.047 96.017zM472.657 114.331h-94.495l.044-94.446h94.497l-.046 94.446zM94.54 399.244H.046l.044-97.588h94.497l-.047 97.588z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M94.495 492.123H0l.044-94.446H94.54l-.045 94.446zM472.563 303.228H378.07l.044-94.446h94.496l-.047 94.446zM472.61 208.779h-94.495l.044-94.448h94.498l-.047 94.448z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M472.517 397.674h-94.494l.044-94.446h94.497l-.047 94.446z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M472.47 492.121h-94.493l.044-96.017h94.496l-.047 96.017z" fill="#1c1c1b" fill-rule="nonzero"/><path d="M228.375 303.22h-96.061l.046-94.446h96.067l-.052 94.446z" fill="#ff7000" fill-rule="nonzero"/><path d="M322.827 397.666h-94.495l.044-96.018h94.498l-.047 96.018z" fill="#ff4900" fill-rule="nonzero"/><path d="M324.444 303.22h-97.636l.046-94.446h97.638l-.048 94.446z" fill="#ff7000" fill-rule="nonzero"/><path d="M418.938 303.22h-96.064l.045-94.446h96.066l-.047 94.446z" fill="#ff7000" fill-rule="nonzero"/><path d="M228.423 208.77H132.36l.045-94.445h96.066l-.05 94.446zM418.985 208.77H322.92l.044-94.445h96.069l-.048 94.446z" fill="#ffa300" fill-rule="nonzero"/><path d="M133.883 304.79H39.392l.044-96.017h94.496l-.049 96.017z" fill="#ff7000" fill-rule="nonzero"/><path d="M133.929 208.77H39.437l.044-95.445h94.496l-.048 95.445z" fill="#ffa300" fill-rule="nonzero"/><path d="M133.976 114.325H39.484l.044-94.448h94.497l-.05 94.448zM511.954 115.325h-94.493l.044-95.448h94.497l-.048 95.448z" fill="#ffce00" fill-rule="nonzero"/><path d="M133.836 399.667H39.345l.044-96.447h94.496l-.049 96.447z" fill="#ff4900" fill-rule="nonzero"/><path d="M133.79 492.117H39.3l.044-94.448h94.496l-.049 94.448z" fill="#ff0107" fill-rule="nonzero"/><path d="M511.862 303.22h-94.495l.046-94.446h94.496l-.047 94.446z" fill="#ff7000" fill-rule="nonzero"/><path d="M511.907 208.77h-94.493l.044-94.445h94.496l-.047 94.446z" fill="#ffa300" fill-rule="nonzero"/><path d="M511.815 398.666h-94.493l.044-95.447h94.496l-.047 95.447z" fill="#ff4900" fill-rule="nonzero"/><path d="M511.77 492.117h-94.496l.046-94.448h94.496l-.047 94.448z" fill="#ff0107" fill-rule="nonzero"/></svg>

gpt4all-chat/icons/openai.svg

@@ -0,0 +1,2 @@
+<?xml version="1.0" encoding="utf-8"?><!-- Uploaded to: SVG Repo, www.svgrepo.com, Generator: SVG Repo Mixer Tools -->
+<svg fill="#000000" width="800px" height="800px" viewBox="0 0 24 24" role="img" xmlns="http://www.w3.org/2000/svg"><title>OpenAI icon</title><path d="M22.2819 9.8211a5.9847 5.9847 0 0 0-.5157-4.9108 6.0462 6.0462 0 0 0-6.5098-2.9A6.0651 6.0651 0 0 0 4.9807 4.1818a5.9847 5.9847 0 0 0-3.9977 2.9 6.0462 6.0462 0 0 0 .7427 7.0966 5.98 5.98 0 0 0 .511 4.9107 6.051 6.051 0 0 0 6.5146 2.9001A5.9847 5.9847 0 0 0 13.2599 24a6.0557 6.0557 0 0 0 5.7718-4.2058 5.9894 5.9894 0 0 0 3.9977-2.9001 6.0557 6.0557 0 0 0-.7475-7.0729zm-9.022 12.6081a4.4755 4.4755 0 0 1-2.8764-1.0408l.1419-.0804 4.7783-2.7582a.7948.7948 0 0 0 .3927-.6813v-6.7369l2.02 1.1686a.071.071 0 0 1 .038.052v5.5826a4.504 4.504 0 0 1-4.4945 4.4944zm-9.6607-4.1254a4.4708 4.4708 0 0 1-.5346-3.0137l.142.0852 4.783 2.7582a.7712.7712 0 0 0 .7806 0l5.8428-3.3685v2.3324a.0804.0804 0 0 1-.0332.0615L9.74 19.9502a4.4992 4.4992 0 0 1-6.1408-1.6464zM2.3408 7.8956a4.485 4.485 0 0 1 2.3655-1.9728V11.6a.7664.7664 0 0 0 .3879.6765l5.8144 3.3543-2.0201 1.1685a.0757.0757 0 0 1-.071 0l-4.8303-2.7865A4.504 4.504 0 0 1 2.3408 7.872zm16.5963 3.8558L13.1038 8.364 15.1192 7.2a.0757.0757 0 0 1 .071 0l4.8303 2.7913a4.4944 4.4944 0 0 1-.6765 8.1042v-5.6772a.79.79 0 0 0-.407-.667zm2.0107-3.0231l-.142-.0852-4.7735-2.7818a.7759.7759 0 0 0-.7854 0L9.409 9.2297V6.8974a.0662.0662 0 0 1 .0284-.0615l4.8303-2.7866a4.4992 4.4992 0 0 1 6.6802 4.66zM8.3065 12.863l-2.02-1.1638a.0804.0804 0 0 1-.038-.0567V6.0742a4.4992 4.4992 0 0 1 7.3757-3.4537l-.142.0805L8.704 5.459a.7948.7948 0 0 0-.3927.6813zm1.0976-2.3654l2.602-1.4998 2.6069 1.4998v2.9994l-2.5974 1.4997-2.6067-1.4997Z"/></svg>

gpt4all-chat/metadata/latestnews.md

@@ -1,16 +1,15 @@
 ## Latest News
 
-GPT4All v3.6.1 was released on December 20th and fixes issues with the stop generation and copy conversation buttons which were broken in v3.6.0.
+GPT4All v3.10.0 was released on February 24th. Changes include:
 
----
-
-GPT4All v3.6.0 was released on December 19th. Changes include:
-
-* **Reasoner v1:**
-  * Built-in javascript code interpreter tool.
-  * Custom curated model that utilizes the code interpreter to break down, analyze, perform, and verify complex reasoning tasks.
-* **Templates:** Automatically substitute chat templates that are not compatible with Jinja2Cpp in GGUFs.
-* **Fixes:**
-  * Remote model template to allow for XML in messages.
-  * Jinja2Cpp bug that broke system message detection in chat templates.
-  * LocalDocs sources displaying in unconsolidated form after v3.5.0.
+* **Remote Models:**
+  * The Add Model page now has a dedicated tab for remote model providers.
+  * Groq, OpenAI, and Mistral remote models are now easier to configure.
+* **CUDA Compatibility:** GPUs with CUDA compute capability 5.0 such as the GTX 750 are now supported by the CUDA backend.
+* **New Model:** The non-MoE Granite model is now supported.
+* **Translation Updates:**
+  * The Italian translation has been updated.
+  * The Simplified Chinese translation has been significantly improved.
+* **Better Chat Templates:** The default chat templates for OLMoE 7B 0924/0125 and Granite 3.1 3B/8B have been improved.
+* **Whitespace Fixes:** DeepSeek-R1-based models now have better whitespace behavior in their output.
+* **Crash Fixes:** Several issues that could potentially cause GPT4All to crash have been fixed.

gpt4all-chat/metadata/models3.json

@@ -32,6 +32,66 @@
     "systemPrompt": "",
     "chatTemplate": "{%- set loop_messages = messages %}\n{%- for message in loop_messages %}\n    {%- set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\\n\\n'+ message['content'] | trim + '<|eot_id|>' %}\n    {{- content }}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' }}\n{%- endif %}"
   },
+  {
+    "order": "aa1",
+    "sha256sum": "5cd4ee65211770f1d99b4f6f4951780b9ef40e29314bd6542bb5bd0ad0bc29d1",
+    "name": "DeepSeek-R1-Distill-Qwen-7B",
+    "filename": "DeepSeek-R1-Distill-Qwen-7B-Q4_0.gguf",
+    "filesize": "4444121056",
+    "requires": "3.8.0",
+    "ramrequired": "8",
+    "parameters": "7 billion",
+    "quant": "q4_0",
+    "type": "deepseek",
+    "description": "<p>The official Qwen2.5-Math-7B distillation of DeepSeek-R1.</p><ul><li>License: <a href=\"https://opensource.org/license/mit\">MIT</a></li><li>No restrictions on commercial use</li><li>#reasoning</li></ul>",
+    "url": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF/resolve/main/DeepSeek-R1-Distill-Qwen-7B-Q4_0.gguf",
+    "chatTemplate": "{%- if not add_generation_prompt is defined %}\n    {%- set add_generation_prompt = false %}\n{%- endif %}\n{%- if messages[0]['role'] == 'system' %}\n    {{- messages[0]['content'] }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '<｜User｜>' + message['content'] }}\n    {%- endif %}\n    {%- if message['role'] == 'assistant' %}\n        {%- set content = message['content'] | regex_replace('^[\\\\s\\\\S]*</think>', '') %}\n        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}\n    {%- endif %}\n{%- endfor -%}\n{%- if add_generation_prompt %}\n    {{- '<｜Assistant｜>' }}\n{%- endif %}"
+  },
+  {
+    "order": "aa2",
+    "sha256sum": "906b3382f2680f4ce845459b4a122e904002b075238080307586bcffcde49eef",
+    "name": "DeepSeek-R1-Distill-Qwen-14B",
+    "filename": "DeepSeek-R1-Distill-Qwen-14B-Q4_0.gguf",
+    "filesize": "8544267680",
+    "requires": "3.8.0",
+    "ramrequired": "16",
+    "parameters": "14 billion",
+    "quant": "q4_0",
+    "type": "deepseek",
+    "description": "<p>The official Qwen2.5-14B distillation of DeepSeek-R1.</p><ul><li>License: <a href=\"https://opensource.org/license/mit\">MIT</a></li><li>No restrictions on commercial use</li><li>#reasoning</li></ul>",
+    "url": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF/resolve/main/DeepSeek-R1-Distill-Qwen-14B-Q4_0.gguf",
+    "chatTemplate": "{%- if not add_generation_prompt is defined %}\n    {%- set add_generation_prompt = false %}\n{%- endif %}\n{%- if messages[0]['role'] == 'system' %}\n    {{- messages[0]['content'] }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '<｜User｜>' + message['content'] }}\n    {%- endif %}\n    {%- if message['role'] == 'assistant' %}\n        {%- set content = message['content'] | regex_replace('^[\\\\s\\\\S]*</think>', '') %}\n        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}\n    {%- endif %}\n{%- endfor -%}\n{%- if add_generation_prompt %}\n    {{- '<｜Assistant｜>' }}\n{%- endif %}"
+  },
+  {
+    "order": "aa3",
+    "sha256sum": "0eb93e436ac8beec18aceb958c120d282cb2cf5451b23185e7be268fe9d375cc",
+    "name": "DeepSeek-R1-Distill-Llama-8B",
+    "filename": "DeepSeek-R1-Distill-Llama-8B-Q4_0.gguf",
+    "filesize": "4675894112",
+    "requires": "3.8.0",
+    "ramrequired": "8",
+    "parameters": "8 billion",
+    "quant": "q4_0",
+    "type": "deepseek",
+    "description": "<p>The official Llama-3.1-8B distillation of DeepSeek-R1.</p><ul><li>License: <a href=\"https://opensource.org/license/mit\">MIT</a></li><li>No restrictions on commercial use</li><li>#reasoning</li></ul>",
+    "url": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF/resolve/main/DeepSeek-R1-Distill-Llama-8B-Q4_0.gguf",
+    "chatTemplate": "{%- if not add_generation_prompt is defined %}\n    {%- set add_generation_prompt = false %}\n{%- endif %}\n{%- if messages[0]['role'] == 'system' %}\n    {{- messages[0]['content'] }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '<｜User｜>' + message['content'] }}\n    {%- endif %}\n    {%- if message['role'] == 'assistant' %}\n        {%- set content = message['content'] | regex_replace('^[\\\\s\\\\S]*</think>', '') %}\n        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}\n    {%- endif %}\n{%- endfor -%}\n{%- if add_generation_prompt %}\n    {{- '<｜Assistant｜>' }}\n{%- endif %}"
+  },
+  {
+    "order": "aa4",
+    "sha256sum": "b3af887d0a015b39fab2395e4faf682c1a81a6a3fd09a43f0d4292f7d94bf4d0",
+    "name": "DeepSeek-R1-Distill-Qwen-1.5B",
+    "filename": "DeepSeek-R1-Distill-Qwen-1.5B-Q4_0.gguf",
+    "filesize": "1068807776",
+    "requires": "3.8.0",
+    "ramrequired": "3",
+    "parameters": "1.5 billion",
+    "quant": "q4_0",
+    "type": "deepseek",
+    "description": "<p>The official Qwen2.5-Math-1.5B distillation of DeepSeek-R1.</p><ul><li>License: <a href=\"https://opensource.org/license/mit\">MIT</a></li><li>No restrictions on commercial use</li><li>#reasoning</li></ul>",
+    "url": "https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-1.5B-GGUF/resolve/main/DeepSeek-R1-Distill-Qwen-1.5B-Q4_0.gguf",
+    "chatTemplate": "{%- if not add_generation_prompt is defined %}\n    {%- set add_generation_prompt = false %}\n{%- endif %}\n{%- if messages[0]['role'] == 'system' %}\n    {{- messages[0]['content'] }}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '<｜User｜>' + message['content'] }}\n    {%- endif %}\n    {%- if message['role'] == 'assistant' %}\n        {%- set content = message['content'] | regex_replace('^[\\\\s\\\\S]*</think>', '') %}\n        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}\n    {%- endif %}\n{%- endfor -%}\n{%- if add_generation_prompt %}\n    {{- '<｜Assistant｜>' }}\n{%- endif %}"
+  },
   {
     "order": "b",
     "md5sum": "27b44e8ae1817525164ddf4f8dae8af4",
@@ -472,7 +532,7 @@
     "filename": "qwen2-1_5b-instruct-q4_0.gguf",
     "filesize": "937532800",
     "requires": "3.0",
-    "ramrequired": "4",
+    "ramrequired": "3",
     "parameters": "1.5 billion",
     "quant": "q4_0",
     "type": "qwen2",

gpt4all-chat/metadata/release.json

@@ -251,12 +251,32 @@
   },
   {
     "version": "3.6.0",
-    "notes": "* **Reasoner v1:**\n  * Built-in javascript code interpreter tool.\n  * Custom curated model that utilizes the code interpreter to break down, analyze, perform, and verify complex reasoning tasks.\n* **Templates:** Automatically substitute chat templates that are not compatible with Jinja2Cpp in GGUFs.\n* **Fixes:**\n  * Remote model template to allow for XML in messages.\n  * Jinja2Cpp bug that broke system message detection in chat templates.\n  * LocalDocs sources displaying in unconsolidated form after v3.5.0.",
+    "notes": "* **Reasoner v1:**\n  * Built-in javascript code interpreter tool.\n  * Custom curated model that utilizes the code interpreter to break down, analyze, perform, and verify complex reasoning tasks.\n* **Templates:** Automatically substitute chat templates that are not compatible with Jinja2Cpp in GGUFs.\n* **Fixes:**\n  * Remote model template to allow for XML in messages.\n  * Jinja2Cpp bug that broke system message detection in chat templates.\n  * LocalDocs sources displaying in unconsolidated form after v3.5.0.\n",
     "contributors": "* Adam Treat (Nomic AI)\n* Jared Van Bortel (Nomic AI)"
   },
   {
     "version": "3.6.1",
-    "notes": "* **Fixes:**\n  * The stop generation button no longer working in v3.6.0.\n  * The copy entire conversation button no longer working in v3.6.0.",
+    "notes": "* **Fixes:**\n  * The stop generation button no longer working in v3.6.0.\n  * The copy entire conversation button no longer working in v3.6.0.\n",
     "contributors": "* Adam Treat (Nomic AI)"
+  },
+  {
+    "version": "3.7.0",
+    "notes": "* **Windows ARM Support:** GPT4All now supports the Windows ARM platform, ensuring compatibility with devices powered by Qualcomm Snapdragon and Microsoft SQ-series processors.\n  * **NOTE:** Support for GPU and/or NPU acceleration is not available at this time. Only the CPU will be used to run LLMs.\n  * **NOTE:** You must install the new *Windows ARM* version of GPT4All from the website. The standard *Windows* version will not work due to emulation limitations.\n* **Fixed Updating on macOS:** The maintenance tool no longer crashes when attempting to update or uninstall GPT4All on Sequoia.\n  * **NOTE:** If you have installed the version from the GitHub releases as a workaround for this issue, you can safely uninstall it and switch back to the version from the website.\n* **Fixed Chat Saving on macOS:** Chats now save as expected when the application is quit with Command-Q.\n* **Code Interpreter Improvements:**\n  * The behavior when the code takes too long to execute and times out has been improved.\n  * console.log now accepts multiple arguments for better compatibility with native JavaScript.\n* **Chat Templating Improvements:**\n  * Two crashes and one compatibility issue have been fixed in the chat template parser.\n  * The default chat template for EM German Mistral has been fixed.\n  * Automatic replacements have been added for five new models as we continue to improve compatibility with common chat templates.\n",
+    "contributors": "* Jared Van Bortel (Nomic AI)\n* Adam Treat (Nomic AI)\n* Riccardo Giovanetti (`@Harvester62`)"
+  },
+  {
+    "version": "3.8.0",
+    "notes": "* **Native DeepSeek-R1-Distill Support:** GPT4All now has robust support for the DeepSeek-R1 family of distillations.\n  * Several model variants are now available on the downloads page.\n  * Reasoning (wrapped in \"think\" tags) is displayed similarly to the Reasoner model.\n  * The DeepSeek-R1 Qwen pretokenizer is now supported, resolving the loading failure in previous versions.\n  * The model is now configured with a GPT4All-compatible prompt template by default.\n* **Chat Templating Overhaul:** The template parser has been *completely* replaced with one that has much better compatibility with common models.\n* **Code Interpreter Fixes:**\n  * An issue preventing the code interpreter from logging a single string in v3.7.0 has been fixed.\n  * The UI no longer freezes while the code interpreter is running a computation.\n* **Local Server Fixes:**\n  * An issue preventing the server from using LocalDocs after the first request since v3.5.0 has been fixed.\n  * System messages are now correctly hidden from the message history.\n",
+    "contributors": "* Jared Van Bortel (Nomic AI)\n* Adam Treat (Nomic AI)\n* ThiloteE (`@ThiloteE`)"
+  },
+  {
+    "version": "3.9.0",
+    "notes": "* **LocalDocs Fix:** LocalDocs no longer shows an error on later messages with reasoning models.\n* **DeepSeek Fix:** DeepSeek-R1 reasoning (in 'think' tags) no longer appears in chat names and follow-up questions.\n* **Windows ARM Improvements:**\n  * Graphical artifacts on some SoCs have been fixed.\n  * A crash when adding a collection of PDFs to LocalDocs has been fixed.\n* **Template Parser Fixes:** Chat templates containing an unclosed comment no longer freeze GPT4All.\n* **New Models:** OLMoE and Granite MoE models are now supported.\n",
+    "contributors": "* Jared Van Bortel (Nomic AI)\n* Adam Treat (Nomic AI)\n* ThiloteE (`@ThiloteE`)"
+  },
+  {
+    "version": "3.10.0",
+    "notes": "* **Remote Models:**\n  * The Add Model page now has a dedicated tab for remote model providers.\n  * Groq, OpenAI, and Mistral remote models are now easier to configure.\n* **CUDA Compatibility:** GPUs with CUDA compute capability 5.0 such as the GTX 750 are now supported by the CUDA backend.\n* **New Model:** The non-MoE Granite model is now supported.\n* **Translation Updates:**\n  * The Italian translation has been updated.\n  * The Simplified Chinese translation has been significantly improved.\n* **Better Chat Templates:** The default chat templates for OLMoE 7B 0924/0125 and Granite 3.1 3B/8B have been improved.\n* **Whitespace Fixes:** DeepSeek-R1-based models now have better whitespace behavior in their output.\n* **Crash Fixes:** Several issues that could potentially cause GPT4All to crash have been fixed.\n",
+    "contributors": "* Jared Van Bortel (Nomic AI)\n* Adam Treat (Nomic AI)\n* ThiloteE (`@ThiloteE`)\n* Lil Bob (`@Junior2Ran`)\n* Riccardo Giovanetti (`@Harvester62`)"
   }
 ]

gpt4all-chat/qml/AddGPT4AllModelView.qml

@@ -204,7 +204,7 @@ ColumnLayout {
                                     Layout.minimumWidth: 200
                                     Layout.fillWidth: true
                                     Layout.alignment: Qt.AlignTop | Qt.AlignHCenter
-                                    visible: !isOnline && !installed && !calcHash && downloadError === ""
+                                    visible: !installed && !calcHash && downloadError === ""
                                     Accessible.description: qsTr("Stop/restart/start the download")
                                     onClicked: {
                                         if (!isDownloading) {
@@ -230,52 +230,6 @@ ColumnLayout {
                                     }
                                 }
 
-                                MySettingsButton {
-                                    id: installButton
-                                    visible: !installed && isOnline
-                                    Layout.topMargin: 20
-                                    Layout.leftMargin: 20
-                                    Layout.minimumWidth: 200
-                                    Layout.fillWidth: true
-                                    Layout.alignment: Qt.AlignTop | Qt.AlignHCenter
-                                    text: qsTr("Install")
-                                    font.pixelSize: theme.fontSizeLarge
-                                    onClicked: {
-                                        var apiKeyText = apiKey.text.trim(),
-                                        baseUrlText = baseUrl.text.trim(),
-                                        modelNameText = modelName.text.trim();
-
-                                        var apiKeyOk = apiKeyText !== "",
-                                        baseUrlOk = !isCompatibleApi || baseUrlText !== "",
-                                        modelNameOk = !isCompatibleApi || modelNameText !== "";
-
-                                        if (!apiKeyOk)
-                                            apiKey.showError();
-                                        if (!baseUrlOk)
-                                            baseUrl.showError();
-                                        if (!modelNameOk)
-                                            modelName.showError();
-
-                                        if (!apiKeyOk || !baseUrlOk || !modelNameOk)
-                                            return;
-
-                                        if (!isCompatibleApi)
-                                            Download.installModel(
-                                                        filename,
-                                                        apiKeyText,
-                                                        );
-                                        else
-                                            Download.installCompatibleModel(
-                                                        modelNameText,
-                                                        apiKeyText,
-                                                        baseUrlText,
-                                                        );
-                                    }
-                                    Accessible.role: Accessible.Button
-                                    Accessible.name: qsTr("Install")
-                                    Accessible.description: qsTr("Install online model")
-                                }
-
                                 ColumnLayout {
                                     spacing: 0
                                     Label {
@@ -390,69 +344,6 @@ ColumnLayout {
                                         Accessible.description: qsTr("Displayed when the file hash is being calculated")
                                     }
                                 }
-
-                                MyTextField {
-                                    id: apiKey
-                                    visible: !installed && isOnline
-                                    Layout.topMargin: 20
-                                    Layout.leftMargin: 20
-                                    Layout.minimumWidth: 200
-                                    Layout.alignment: Qt.AlignTop | Qt.AlignHCenter
-                                    wrapMode: Text.WrapAnywhere
-                                    function showError() {
-                                        messageToast.show(qsTr("ERROR: $API_KEY is empty."));
-                                        apiKey.placeholderTextColor = theme.textErrorColor;
-                                    }
-                                    onTextChanged: {
-                                        apiKey.placeholderTextColor = theme.mutedTextColor;
-                                    }
-                                    placeholderText: qsTr("enter $API_KEY")
-                                    Accessible.role: Accessible.EditableText
-                                    Accessible.name: placeholderText
-                                    Accessible.description: qsTr("Whether the file hash is being calculated")
-                                }
-
-                                MyTextField {
-                                    id: baseUrl
-                                    visible: !installed && isOnline && isCompatibleApi
-                                    Layout.topMargin: 20
-                                    Layout.leftMargin: 20
-                                    Layout.minimumWidth: 200
-                                    Layout.alignment: Qt.AlignTop | Qt.AlignHCenter
-                                    wrapMode: Text.WrapAnywhere
-                                    function showError() {
-                                        messageToast.show(qsTr("ERROR: $BASE_URL is empty."));
-                                        baseUrl.placeholderTextColor = theme.textErrorColor;
-                                    }
-                                    onTextChanged: {
-                                        baseUrl.placeholderTextColor = theme.mutedTextColor;
-                                    }
-                                    placeholderText: qsTr("enter $BASE_URL")
-                                    Accessible.role: Accessible.EditableText
-                                    Accessible.name: placeholderText
-                                    Accessible.description: qsTr("Whether the file hash is being calculated")
-                                }
-
-                                MyTextField {
-                                    id: modelName
-                                    visible: !installed && isOnline && isCompatibleApi
-                                    Layout.topMargin: 20
-                                    Layout.leftMargin: 20
-                                    Layout.minimumWidth: 200
-                                    Layout.alignment: Qt.AlignTop | Qt.AlignHCenter
-                                    wrapMode: Text.WrapAnywhere
-                                    function showError() {
-                                        messageToast.show(qsTr("ERROR: $MODEL_NAME is empty."))
-                                        modelName.placeholderTextColor = theme.textErrorColor;
-                                    }
-                                    onTextChanged: {
-                                        modelName.placeholderTextColor = theme.mutedTextColor;
-                                    }
-                                    placeholderText: qsTr("enter $MODEL_NAME")
-                                    Accessible.role: Accessible.EditableText
-                                    Accessible.name: placeholderText
-                                    Accessible.description: qsTr("Whether the file hash is being calculated")
-                                }
                             }
                         }
                     }

gpt4all-chat/qml/AddModelView.qml

@@ -89,6 +89,13 @@ Rectangle {
                     gpt4AllModelView.show();
                 }
             }
+            MyTabButton {
+                text: qsTr("Remote Providers")
+                isSelected: remoteModelView.isShown()
+                onPressed: {
+                    remoteModelView.show();
+                }
+            }
             MyTabButton {
                 text: qsTr("HuggingFace")
                 isSelected: huggingfaceModelView.isShown()
@@ -112,7 +119,20 @@ Rectangle {
                     stackLayout.currentIndex = 0;
                 }
                 function isShown() {
-                    return stackLayout.currentIndex === 0
+                    return stackLayout.currentIndex === 0;
+                }
+            }
+
+            AddRemoteModelView {
+                id: remoteModelView
+                Layout.fillWidth: true
+                Layout.fillHeight: true
+
+                function show() {
+                    stackLayout.currentIndex = 1;
+                }
+                function isShown() {
+                    return stackLayout.currentIndex === 1;
                 }
             }
 
@@ -126,10 +146,10 @@ Rectangle {
                 anchors.fill: parent
 
                 function show() {
-                    stackLayout.currentIndex = 1;
+                    stackLayout.currentIndex = 2;
                 }
                 function isShown() {
-                    return stackLayout.currentIndex === 1
+                    return stackLayout.currentIndex === 2;
                 }
             }
         }

gpt4all-chat/qml/AddRemoteModelView.qml

@@ -0,0 +1,147 @@
+import QtCore
+import QtQuick
+import QtQuick.Controls
+import QtQuick.Controls.Basic
+import QtQuick.Layouts
+import QtQuick.Dialogs
+import Qt.labs.folderlistmodel
+import Qt5Compat.GraphicalEffects
+
+import llm
+import chatlistmodel
+import download
+import modellist
+import network
+import gpt4all
+import mysettings
+import localdocs
+
+ColumnLayout {
+    Layout.fillWidth: true
+    Layout.alignment: Qt.AlignTop
+    spacing: 5
+
+    Label {
+        Layout.topMargin: 0
+        Layout.bottomMargin: 25
+        Layout.rightMargin: 150 * theme.fontScale
+        Layout.alignment: Qt.AlignTop
+        Layout.fillWidth: true
+        verticalAlignment: Text.AlignTop
+        text: qsTr("Various remote model providers that use network resources for inference.")
+        font.pixelSize: theme.fontSizeLarger
+        color: theme.textColor
+        wrapMode: Text.WordWrap
+    }
+
+    ScrollView {
+        id: scrollView
+        ScrollBar.vertical.policy: ScrollBar.AsNeeded
+        Layout.fillWidth: true
+        Layout.fillHeight: true
+        contentWidth: availableWidth
+        clip: true
+        Flow {
+            anchors.left: parent.left
+            anchors.right: parent.right
+            spacing: 20
+            bottomPadding: 20
+            property int childWidth: 330 * theme.fontScale
+            property int childHeight: 400 + 166 * theme.fontScale
+            RemoteModelCard {
+                width: parent.childWidth
+                height: parent.childHeight
+                providerBaseUrl: "https://api.groq.com/openai/v1/"
+                providerName: qsTr("Groq")
+                providerImage: "qrc:/gpt4all/icons/groq.svg"
+                providerDesc: qsTr('Groq offers a high-performance AI inference engine designed for low-latency and efficient processing. Optimized for real-time applications, Groq’s technology is ideal for users who need fast responses from open large language models and other AI workloads.<br><br>Get your API key: <a href="https://console.groq.com/keys">https://groq.com/</a>')
+                modelWhitelist: [
+                    // last updated 2025-02-24
+                    "deepseek-r1-distill-llama-70b",
+                    "deepseek-r1-distill-qwen-32b",
+                    "gemma2-9b-it",
+                    "llama-3.1-8b-instant",
+                    "llama-3.2-1b-preview",
+                    "llama-3.2-3b-preview",
+                    "llama-3.3-70b-specdec",
+                    "llama-3.3-70b-versatile",
+                    "llama3-70b-8192",
+                    "llama3-8b-8192",
+                    "mixtral-8x7b-32768",
+                    "qwen-2.5-32b",
+                    "qwen-2.5-coder-32b",
+                ]
+            }
+            RemoteModelCard {
+                width: parent.childWidth
+                height: parent.childHeight
+                providerBaseUrl: "https://api.openai.com/v1/"
+                providerName: qsTr("OpenAI")
+                providerImage: "qrc:/gpt4all/icons/openai.svg"
+                providerDesc: qsTr('OpenAI provides access to advanced AI models, including GPT-4 supporting a wide range of applications, from conversational AI to content generation and code completion.<br><br>Get your API key: <a href="https://platform.openai.com/signup">https://openai.com/</a>')
+                modelWhitelist: [
+                    // last updated 2025-02-24
+                    "gpt-3.5-turbo",
+                    "gpt-3.5-turbo-16k",
+                    "gpt-4",
+                    "gpt-4-32k",
+                    "gpt-4-turbo",
+                    "gpt-4o",
+                ]
+            }
+            RemoteModelCard {
+                width: parent.childWidth
+                height: parent.childHeight
+                providerBaseUrl: "https://api.mistral.ai/v1/"
+                providerName: qsTr("Mistral")
+                providerImage: "qrc:/gpt4all/icons/mistral.svg"
+                providerDesc: qsTr('Mistral AI specializes in efficient, open-weight language models optimized for various natural language processing tasks. Their models are designed for flexibility and performance, making them a solid option for applications requiring scalable AI solutions.<br><br>Get your API key: <a href="https://mistral.ai/">https://mistral.ai/</a>')
+                modelWhitelist: [
+                    // last updated 2025-02-24
+                    "codestral-2405",
+                    "codestral-2411-rc5",
+                    "codestral-2412",
+                    "codestral-2501",
+                    "codestral-latest",
+                    "codestral-mamba-2407",
+                    "codestral-mamba-latest",
+                    "ministral-3b-2410",
+                    "ministral-3b-latest",
+                    "ministral-8b-2410",
+                    "ministral-8b-latest",
+                    "mistral-large-2402",
+                    "mistral-large-2407",
+                    "mistral-large-2411",
+                    "mistral-large-latest",
+                    "mistral-medium-2312",
+                    "mistral-medium-latest",
+                    "mistral-saba-2502",
+                    "mistral-saba-latest",
+                    "mistral-small-2312",
+                    "mistral-small-2402",
+                    "mistral-small-2409",
+                    "mistral-small-2501",
+                    "mistral-small-latest",
+                    "mistral-tiny-2312",
+                    "mistral-tiny-2407",
+                    "mistral-tiny-latest",
+                    "open-codestral-mamba",
+                    "open-mistral-7b",
+                    "open-mistral-nemo",
+                    "open-mistral-nemo-2407",
+                    "open-mixtral-8x22b",
+                    "open-mixtral-8x22b-2404",
+                    "open-mixtral-8x7b",
+                ]
+            }
+            RemoteModelCard {
+                width: parent.childWidth
+                height: parent.childHeight
+                providerIsCustom: true
+                providerName: qsTr("Custom")
+                providerImage: "qrc:/gpt4all/icons/antenna_3.svg"
+                providerDesc: qsTr("The custom provider option allows users to connect their own OpenAI-compatible AI models or third-party inference services. This is useful for organizations with proprietary models or those leveraging niche AI providers not listed here.")
+            }
+        }
+    }
+}

gpt4all-chat/qml/ChatCollapsibleItem.qml

@@ -13,6 +13,8 @@ ColumnLayout {
     property alias textContent: innerTextItem.textContent
     property bool isCurrent: false
     property bool isError: false
+    property bool isThinking: false
+    property int  thinkingTime: 0
 
     Layout.topMargin: 10
     Layout.bottomMargin: 10
@@ -26,16 +28,20 @@ ColumnLayout {
             anchors.bottom: parent.bottom
 
             Item {
-                width: myTextArea.width
-                height: myTextArea.height
+                Layout.preferredWidth: myTextArea.implicitWidth
+                Layout.preferredHeight: myTextArea.implicitHeight
                 TextArea {
                     id: myTextArea
                     text: {
                         if (isError)
                             return qsTr("Analysis encountered error");
                         if (isCurrent)
-                            return qsTr("Analyzing");
-                        return qsTr("Analyzed");
+                            return isThinking ? qsTr("Thinking") : qsTr("Analyzing");
+                        return isThinking
+                            ? qsTr("Thought for %1 %2")
+                                  .arg(Math.ceil(thinkingTime / 1000.0))
+                                  .arg(Math.ceil(thinkingTime / 1000.0) === 1 ? qsTr("second") : qsTr("seconds"))
+                            : qsTr("Analyzed");
                     }
                     padding: 0
                     font.pixelSize: theme.fontSizeLarger

gpt4all-chat/qml/ChatItemView.qml

@@ -110,6 +110,7 @@ GridLayout {
                         case Chat.PromptProcessing: return qsTr("processing ...")
                         case Chat.ResponseGeneration: return qsTr("generating response ...");
                         case Chat.GeneratingQuestions: return qsTr("generating questions ...");
+                        case Chat.ToolCallGeneration: return qsTr("generating toolcall ...");
                         default: return ""; // handle unexpected values
                         }
                     }
@@ -188,6 +189,18 @@ GridLayout {
                         isError: modelData.isToolCallError
                     }
                 }
+                DelegateChoice {
+                    roleValue: "Think: ";
+                    ChatCollapsibleItem {
+                        Layout.fillWidth: true
+                        textContent: modelData.content
+                        isCurrent: modelData.isCurrentResponse
+                        isError: false
+                        isThinking: true
+                        thinkingTime: modelData.thinkingTime
+                        visible: modelData.content !== ""
+                    }
+                }
             }
 
             delegate: chooser

gpt4all-chat/qml/ChatView.qml

@@ -828,7 +828,7 @@ Rectangle {
                                     textInput.cursorPosition = text.length;
                                 }
                                 height: visible ? implicitHeight : 0
-                                visible: name !== "ToolResponse: "
+                                visible: name !== "ToolResponse: " && name !== "System: "
                             }
 
                             remove: Transition {

gpt4all-chat/qml/MyComboBox.qml

@@ -60,27 +60,28 @@ ComboBox {
         highlighted: comboBox.highlightedIndex === index
     }
     popup: Popup {
-        // FIXME This should be made much nicer to take into account lists that are very long so
-        // that it is scrollable and also sized optimally taking into account the x,y and the content
-        // width and height as well as the window width and height
         y: comboBox.height - 1
         width: comboBox.width
-        implicitHeight: contentItem.implicitHeight + 20
+        implicitHeight: Math.min(window.height - y, contentItem.implicitHeight + 20)
         padding: 0
-
         contentItem: Rectangle {
-            implicitWidth: myListView.contentWidth
+            implicitWidth: comboBox.width
             implicitHeight: myListView.contentHeight
             color: "transparent"
-            ListView {
-                id: myListView
+            radius: 10
+            ScrollView {
                 anchors.fill: parent
                 anchors.margins: 10
                 clip: true
-                implicitHeight: contentHeight
-                model: comboBox.popup.visible ? comboBox.delegateModel : null
-                currentIndex: comboBox.highlightedIndex
-                ScrollIndicator.vertical: ScrollIndicator { }
+                ScrollBar.vertical.policy: ScrollBar.AsNeeded
+                ScrollBar.horizontal.policy: ScrollBar.AlwaysOff
+                ListView {
+                    id: myListView
+                    implicitHeight: contentHeight
+                    model: comboBox.popup.visible ? comboBox.delegateModel : null
+                    currentIndex: comboBox.highlightedIndex
+                    ScrollIndicator.vertical: ScrollIndicator { }
+                }
             }
         }
 

gpt4all-chat/qml/RemoteModelCard.qml

@@ -0,0 +1,221 @@
+import QtCore
+import QtQuick
+import QtQuick.Controls
+import QtQuick.Controls.Basic
+import QtQuick.Layouts
+import QtQuick.Dialogs
+import Qt.labs.folderlistmodel
+import Qt5Compat.GraphicalEffects
+
+import llm
+import chatlistmodel
+import download
+import modellist
+import network
+import gpt4all
+import mysettings
+import localdocs
+
+
+Rectangle {
+    property alias providerName: providerNameLabel.text
+    property alias providerImage: myimage.source
+    property alias providerDesc: providerDescLabel.text
+    property string providerBaseUrl: ""
+    property bool providerIsCustom: false
+    property var modelWhitelist: null
+
+    color: theme.conversationBackground
+    radius: 10
+    border.width: 1
+    border.color: theme.controlBorder
+    implicitHeight: topColumn.height + bottomColumn.height + 33 * theme.fontScale
+
+    ColumnLayout {
+        id: topColumn
+        anchors.left: parent.left
+        anchors.right: parent.right
+        anchors.top: parent.top
+        anchors.margins: 20
+        spacing: 15 * theme.fontScale
+        RowLayout {
+            Layout.alignment: Qt.AlignTop
+            spacing: 10
+            Item {
+                Layout.preferredWidth: 27 * theme.fontScale
+                Layout.preferredHeight: 27 * theme.fontScale
+                Layout.alignment: Qt.AlignLeft
+
+                Image {
+                    id: myimage
+                    anchors.centerIn: parent
+                    sourceSize.width: parent.width
+                    sourceSize.height: parent.height
+                    mipmap: true
+                    fillMode: Image.PreserveAspectFit
+                }
+            }
+
+            Label {
+                id: providerNameLabel
+                color: theme.textColor
+                font.pixelSize: theme.fontSizeBanner
+            }
+        }
+
+        Label {
+            id: providerDescLabel
+            Layout.fillWidth: true
+            wrapMode: Text.Wrap
+            color: theme.settingsTitleTextColor
+            font.pixelSize: theme.fontSizeLarge
+            onLinkActivated: function(link) { Qt.openUrlExternally(link); }
+
+            MouseArea {
+                anchors.fill: parent
+                acceptedButtons: Qt.NoButton // pass clicks to parent
+                cursorShape: parent.hoveredLink ? Qt.PointingHandCursor : Qt.ArrowCursor
+            }
+        }
+    }
+
+    ColumnLayout {
+        id: bottomColumn
+        anchors.left: parent.left
+        anchors.right: parent.right
+        anchors.bottom: parent.bottom
+        anchors.margins: 20
+        spacing: 30
+
+        ColumnLayout {
+            MySettingsLabel {
+                text: qsTr("API Key")
+                font.bold: true
+                font.pixelSize: theme.fontSizeLarge
+                color: theme.settingsTitleTextColor
+            }
+
+            MyTextField {
+                id: apiKeyField
+                Layout.fillWidth: true
+                font.pixelSize: theme.fontSizeLarge
+                wrapMode: Text.WrapAnywhere
+                function showError() {
+                    messageToast.show(qsTr("ERROR: $API_KEY is empty."));
+                    apiKeyField.placeholderTextColor = theme.textErrorColor;
+                }
+                onTextChanged: {
+                    apiKeyField.placeholderTextColor = theme.mutedTextColor;
+                    if (!providerIsCustom) {
+                        let models = ModelList.remoteModelList(apiKeyField.text, providerBaseUrl);
+                        if (modelWhitelist !== null)
+                            models = models.filter(m => modelWhitelist.includes(m));
+                        myModelList.model = models;
+                        myModelList.currentIndex = -1;
+                    }
+                }
+                placeholderText: qsTr("enter $API_KEY")
+                Accessible.role: Accessible.EditableText
+                Accessible.name: placeholderText
+                Accessible.description: qsTr("Whether the file hash is being calculated")
+            }
+        }
+
+        ColumnLayout {
+            visible: providerIsCustom
+            MySettingsLabel {
+                text: qsTr("Base Url")
+                font.bold: true
+                font.pixelSize: theme.fontSizeLarge
+                color: theme.settingsTitleTextColor
+            }
+            MyTextField {
+                id: baseUrlField
+                Layout.fillWidth: true
+                font.pixelSize: theme.fontSizeLarge
+                wrapMode: Text.WrapAnywhere
+                function showError() {
+                    messageToast.show(qsTr("ERROR: $BASE_URL is empty."));
+                    baseUrlField.placeholderTextColor = theme.textErrorColor;
+                }
+                onTextChanged: {
+                    baseUrlField.placeholderTextColor = theme.mutedTextColor;
+                }
+                placeholderText: qsTr("enter $BASE_URL")
+                Accessible.role: Accessible.EditableText
+                Accessible.name: placeholderText
+            }
+        }
+        ColumnLayout {
+            visible: providerIsCustom
+            MySettingsLabel {
+                text: qsTr("Model Name")
+                font.bold: true
+                font.pixelSize: theme.fontSizeLarge
+                color: theme.settingsTitleTextColor
+            }
+            MyTextField {
+                id: modelNameField
+                Layout.fillWidth: true
+                font.pixelSize: theme.fontSizeLarge
+                wrapMode: Text.WrapAnywhere
+                function showError() {
+                    messageToast.show(qsTr("ERROR: $MODEL_NAME is empty."))
+                    modelNameField.placeholderTextColor = theme.textErrorColor;
+                }
+                onTextChanged: {
+                    modelNameField.placeholderTextColor = theme.mutedTextColor;
+                }
+                placeholderText: qsTr("enter $MODEL_NAME")
+                Accessible.role: Accessible.EditableText
+                Accessible.name: placeholderText
+            }
+        }
+
+        ColumnLayout {
+            visible: myModelList.count > 0 && !providerIsCustom
+
+            MySettingsLabel {
+                text: qsTr("Models")
+                font.bold: true
+                font.pixelSize: theme.fontSizeLarge
+                color: theme.settingsTitleTextColor
+            }
+
+            RowLayout {
+                spacing: 10
+
+                MyComboBox {
+                    Layout.fillWidth: true
+                    id: myModelList
+                    currentIndex: -1;
+                }
+            }
+        }
+
+        MySettingsButton {
+            id: installButton
+            Layout.alignment: Qt.AlignRight
+            text: qsTr("Install")
+            font.pixelSize: theme.fontSizeLarge
+
+            property string apiKeyText: apiKeyField.text.trim()
+            property string baseUrlText: providerIsCustom ? baseUrlField.text.trim() : providerBaseUrl.trim()
+            property string modelNameText: providerIsCustom ? modelNameField.text.trim() : myModelList.currentText.trim()
+
+            enabled: apiKeyText !== "" && baseUrlText !== "" && modelNameText !== ""
+
+            onClicked: {
+                Download.installCompatibleModel(
+                            modelNameText,
+                            apiKeyText,
+                            baseUrlText,
+                            );
+                myModelList.currentIndex = -1;
+            }
+            Accessible.role: Accessible.Button
+            Accessible.name: qsTr("Install")
+            Accessible.description: qsTr("Install remote model")
+        }
+    }
+}

gpt4all-chat/src/chat.cpp

@@ -7,24 +7,26 @@
 #include "toolcallparser.h"
 #include "toolmodel.h"
 
-#include <QBuffer>
+#include <QByteArray>
 #include <QDataStream>
 #include <QDebug>
-#include <QJsonDocument>
-#include <QJsonObject>
-#include <QJsonValue>
+#include <QFile>
+#include <QFileInfo>
+#include <QIODevice>
 #include <QLatin1String>
 #include <QMap>
 #include <QRegularExpression>
 #include <QString>
-#include <QVariant>
 #include <Qt>
+#include <QtAssert>
 #include <QtLogging>
 
+#include <optional>
 #include <utility>
 
 using namespace ToolEnums;
 
+
 Chat::Chat(QObject *parent)
     : QObject(parent)
     , m_id(Network::globalInstance()->generateUniqueId())
@@ -181,6 +183,11 @@ QVariant Chat::popPrompt(int index)
 
 void Chat::stopGenerating()
 {
+    // In future if we have more than one tool we'll have to keep track of which tools are possibly
+    // running, but for now we only have one
+    Tool *toolInstance = ToolModel::globalInstance()->get(ToolCallConstants::CodeInterpreterFunction);
+    Q_ASSERT(toolInstance);
+    toolInstance->interrupt();
     m_llmodel->stopGenerating();
 }
 
@@ -242,56 +249,71 @@ void Chat::responseStopped(qint64 promptResponseMs)
 
     const QString possibleToolcall = m_chatModel->possibleToolcall();
 
-    ToolCallParser parser;
-    parser.update(possibleToolcall);
-
-    if (parser.state() == ToolEnums::ParseState::Complete) {
-        const QString toolCall = parser.toolCall();
-
-        // Regex to remove the formatting around the code
-        static const QRegularExpression regex("^\\s*```javascript\\s*|\\s*```\\s*$");
-        QString code = toolCall;
-        code.remove(regex);
-        code = code.trimmed();
-
-        // Right now the code interpreter is the only available tool
-        Tool *toolInstance = ToolModel::globalInstance()->get(ToolCallConstants::CodeInterpreterFunction);
-        Q_ASSERT(toolInstance);
-
-        // The param is the code
-        const ToolParam param = { "code", ToolEnums::ParamType::String, code };
-        const QString result = toolInstance->run({param}, 10000 /*msecs to timeout*/);
-        const ToolEnums::Error error = toolInstance->error();
-        const QString errorString = toolInstance->errorString();
-
-        // Update the current response with meta information about toolcall and re-parent
-        m_chatModel->updateToolCall({
-            ToolCallConstants::CodeInterpreterFunction,
-            { param },
-            result,
-            error,
-            errorString
-        });
-
-        ++m_consecutiveToolCalls;
-
-        // We limit the number of consecutive toolcalls otherwise we get into a potentially endless loop
-        if (m_consecutiveToolCalls < 3 || error == ToolEnums::Error::NoError) {
-            resetResponseState();
-            emit promptRequested(m_collections); // triggers a new response
-            return;
-        }
-    }
-
-    if (m_generatedName.isEmpty())
-        emit generateNameRequested();
-
-    m_consecutiveToolCalls = 0;
-    Network::globalInstance()->trackChatEvent("response_complete", {
+    Network::globalInstance()->trackChatEvent("response_stopped", {
         {"first", m_firstResponse},
         {"message_count", chatModel()->count()},
         {"$duration", promptResponseMs / 1000.},
     });
+
+    ToolCallParser parser;
+    parser.update(possibleToolcall.toUtf8());
+    if (parser.state() == ToolEnums::ParseState::Complete && parser.startTag() != ToolCallConstants::ThinkStartTag)
+        processToolCall(parser.toolCall());
+    else
+        responseComplete();
+}
+
+void Chat::processToolCall(const QString &toolCall)
+{
+    m_responseState = Chat::ToolCallGeneration;
+    emit responseStateChanged();
+    // Regex to remove the formatting around the code
+    static const QRegularExpression regex("^\\s*```javascript\\s*|\\s*```\\s*$");
+    QString code = toolCall;
+    code.remove(regex);
+    code = code.trimmed();
+
+    // Right now the code interpreter is the only available tool
+    Tool *toolInstance = ToolModel::globalInstance()->get(ToolCallConstants::CodeInterpreterFunction);
+    Q_ASSERT(toolInstance);
+    connect(toolInstance, &Tool::runComplete, this, &Chat::toolCallComplete, Qt::SingleShotConnection);
+
+    // The param is the code
+    const ToolParam param = { "code", ToolEnums::ParamType::String, code };
+    m_responseInProgress = true;
+    emit responseInProgressChanged();
+    toolInstance->run({param});
+}
+
+void Chat::toolCallComplete(const ToolCallInfo &info)
+{
+    // Update the current response with meta information about toolcall and re-parent
+    m_chatModel->updateToolCall(info);
+
+    ++m_consecutiveToolCalls;
+
+    m_responseInProgress = false;
+    emit responseInProgressChanged();
+
+    // We limit the number of consecutive toolcalls otherwise we get into a potentially endless loop
+    if (m_consecutiveToolCalls < 3 || info.error == ToolEnums::Error::NoError) {
+        resetResponseState();
+        emit promptRequested(m_collections); // triggers a new response
+        return;
+    }
+
+    responseComplete();
+}
+
+void Chat::responseComplete()
+{
+    if (m_generatedName.isEmpty())
+        emit generateNameRequested();
+
+    m_responseState = Chat::ResponseStopped;
+    emit responseStateChanged();
+
+    m_consecutiveToolCalls = 0;
     m_firstResponse = false;
 }
 
@@ -361,11 +383,8 @@ void Chat::trySwitchContextOfLoadedModel()
 
 void Chat::generatedNameChanged(const QString &name)
 {
-    // Only use the first three words maximum and remove newlines and extra spaces
-    m_generatedName = name.simplified();
-    QStringList words = m_generatedName.split(' ', Qt::SkipEmptyParts);
-    int wordCount = qMin(7, words.size());
-    m_name = words.mid(0, wordCount).join(' ');
+    m_generatedName = name;
+    m_name = name;
     emit nameChanged();
     m_needsSave = true;
 }

gpt4all-chat/src/chat.h

@@ -3,21 +3,26 @@
 
 #include "chatllm.h"
 #include "chatmodel.h"
-#include "database.h" // IWYU pragma: keep
-#include "localdocsmodel.h" // IWYU pragma: keep
+#include "database.h"
+#include "localdocsmodel.h"
 #include "modellist.h"
+#include "tool.h"
 
 #include <QDateTime>
 #include <QList>
 #include <QObject>
-#include <QQmlEngine>
+#include <QQmlEngine> // IWYU pragma: keep
 #include <QString>
 #include <QStringList> // IWYU pragma: keep
-#include <QStringView>
-#include <QtGlobal>
+#include <QUrl>
+#include <QVariant>
+#include <QtTypes>
 
+// IWYU pragma: no_forward_declare LocalDocsCollectionsModel
+// IWYU pragma: no_forward_declare ToolCallInfo
 class QDataStream;
 
+
 class Chat : public QObject
 {
     Q_OBJECT
@@ -55,7 +60,8 @@ public:
         LocalDocsProcessing,
         PromptProcessing,
         GeneratingQuestions,
-        ResponseGeneration
+        ResponseGeneration,
+        ToolCallGeneration
     };
     Q_ENUM(ResponseState)
 
@@ -166,6 +172,9 @@ private Q_SLOTS:
     void promptProcessing();
     void generatingQuestions();
     void responseStopped(qint64 promptResponseMs);
+    void processToolCall(const QString &toolCall);
+    void toolCallComplete(const ToolCallInfo &info);
+    void responseComplete();
     void generatedNameChanged(const QString &name);
     void generatedQuestionFinished(const QString &question);
     void handleModelLoadingError(const QString &error);

gpt4all-chat/src/chatapi.cpp

@@ -2,6 +2,9 @@
 
 #include "utils.h"
 
+#include <fmt/format.h>
+
+#include <QAnyStringView>
 #include <QCoreApplication>
 #include <QDebug>
 #include <QGuiApplication>
@@ -9,15 +12,17 @@
 #include <QJsonDocument>
 #include <QJsonObject>
 #include <QJsonValue>
+#include <QLatin1String>
 #include <QNetworkAccessManager>
 #include <QNetworkRequest>
+#include <QStringView>
 #include <QThread>
 #include <QUrl>
-#include <QUtf8StringView>
+#include <QUtf8StringView> // IWYU pragma: keep
 #include <QVariant>
 #include <QXmlStreamReader>
 #include <Qt>
-#include <QtGlobal>
+#include <QtAssert>
 #include <QtLogging>
 
 #include <expected>
@@ -29,6 +34,7 @@ using namespace Qt::Literals::StringLiterals;
 
 //#define DEBUG
 
+
 ChatAPI::ChatAPI()
     : QObject(nullptr)
     , m_modelName("gpt-3.5-turbo")

gpt4all-chat/src/chatapi.h

@@ -3,10 +3,11 @@
 
 #include <gpt4all-backend/llmodel.h>
 
-#include <QByteArray> // IWYU pragma: keep
+#include <QByteArray>
 #include <QNetworkReply>
 #include <QObject>
 #include <QString>
+#include <QtPreprocessorSupport>
 
 #include <cstddef>
 #include <cstdint>
@@ -17,9 +18,11 @@
 #include <unordered_map>
 #include <vector>
 
+// IWYU pragma: no_forward_declare QByteArray
+class ChatAPI;
 class QNetworkAccessManager;
 
-class ChatAPI;
+
 class ChatAPIWorker : public QObject {
     Q_OBJECT
 public:

gpt4all-chat/src/chatlistmodel.cpp

@@ -1,23 +1,24 @@
 #include "chatlistmodel.h"
 
-#include "database.h" // IWYU pragma: keep
 #include "mysettings.h"
 
+#include <QCoreApplication>
 #include <QDataStream>
 #include <QDir>
 #include <QElapsedTimer>
+#include <QEvent>
 #include <QFile>
 #include <QFileInfo>
 #include <QGlobalStatic>
 #include <QGuiApplication>
 #include <QIODevice>
 #include <QSettings>
-#include <QString>
-#include <QStringList>
+#include <QStringList> // IWYU pragma: keep
 #include <Qt>
+#include <QtTypes>
 
 #include <algorithm>
-#include <memory>
+
 
 static constexpr quint32 CHAT_FORMAT_MAGIC   = 0xF5D553CC;
 static constexpr qint32  CHAT_FORMAT_VERSION = 12;

gpt4all-chat/src/chatlistmodel.h

@@ -7,17 +7,20 @@
 
 #include <QAbstractListModel>
 #include <QByteArray>
+#include <QDate>
 #include <QDebug>
 #include <QHash>
 #include <QList>
 #include <QMutex>
 #include <QObject>
+#include <QString>
 #include <QThread>
 #include <QVariant>
-#include <QVector>
+#include <QVector> // IWYU pragma: keep
 #include <Qt>
-#include <QtGlobal>
+#include <QtAssert>
 #include <QtLogging>
+#include <QtPreprocessorSupport>
 
 #include <memory>
 

gpt4all-chat/src/chatllm.cpp

@@ -12,38 +12,43 @@
 #include "toolcallparser.h"
 
 #include <fmt/format.h>
+#include <minja/minja.hpp>
+#include <nlohmann/json.hpp>
 
-#include <jinja2cpp/error_info.h>
-#include <jinja2cpp/template.h>
-#include <jinja2cpp/template_env.h>
-#include <jinja2cpp/user_callable.h>
-#include <jinja2cpp/value.h>
-
+#include <QChar>
 #include <QDataStream>
 #include <QDebug>
 #include <QFile>
 #include <QGlobalStatic>
-#include <QIODevice>
+#include <QIODevice> // IWYU pragma: keep
 #include <QJsonDocument>
 #include <QJsonObject>
 #include <QJsonValue>
 #include <QMap>
-#include <QMutex>
+#include <QMutex> // IWYU pragma: keep
 #include <QMutexLocker> // IWYU pragma: keep
-#include <QRegularExpression>
-#include <QRegularExpressionMatch>
+#include <QRegularExpression> // IWYU pragma: keep
+#include <QRegularExpressionMatch> // IWYU pragma: keep
 #include <QSet>
+#include <QStringView>
+#include <QTextStream>
 #include <QUrl>
+#include <QVariant>
 #include <QWaitCondition>
 #include <Qt>
+#include <QtAssert>
 #include <QtLogging>
+#include <QtTypes> // IWYU pragma: keep
 
 #include <algorithm>
 #include <chrono>
 #include <cmath>
+#include <concepts>
 #include <cstddef>
+#include <cstdint>
 #include <ctime>
 #include <exception>
+#include <functional>
 #include <iomanip>
 #include <limits>
 #include <optional>
@@ -60,44 +65,104 @@
 using namespace Qt::Literals::StringLiterals;
 using namespace ToolEnums;
 namespace ranges = std::ranges;
+using json = nlohmann::ordered_json;
 
 //#define DEBUG
 //#define DEBUG_MODEL_LOADING
 
 // NOTE: not threadsafe
-static jinja2::TemplateEnv *jinjaEnv()
+static const std::shared_ptr<minja::Context> &jinjaEnv()
 {
-    static std::optional<jinja2::TemplateEnv> environment;
+    static std::shared_ptr<minja::Context> environment;
     if (!environment) {
-        auto &env = environment.emplace();
-        auto &settings = env.GetSettings();
-        settings.trimBlocks = true;
-        settings.lstripBlocks = true;
-        env.AddGlobal("raise_exception", jinja2::UserCallable(
-            /*callable*/ [](auto &params) -> jinja2::Value {
-                auto messageArg = params.args.find("message");
-                if (messageArg == params.args.end() || !messageArg->second.isString())
-                    throw std::runtime_error("'message' argument to raise_exception() must be a string");
-                throw std::runtime_error(fmt::format("Jinja template error: {}", messageArg->second.asString()));
-            },
-            /*argsInfo*/ { jinja2::ArgInfo("message", /*isMandatory*/ true) }
-        ));
-        env.AddGlobal("strftime_now", jinja2::UserCallable(
-            /*callable*/ [](auto &params) -> jinja2::Value {
+        environment = minja::Context::builtins();
+        environment->set("strftime_now", minja::simple_function(
+            "strftime_now", { "format" },
+            [](const std::shared_ptr<minja::Context> &, minja::Value &args) -> minja::Value {
+                auto format = args.at("format").get<std::string>();
                 using Clock = std::chrono::system_clock;
-                auto formatArg = params.args.find("format");
-                if (formatArg == params.args.end() || !formatArg->second.isString())
-                    throw std::runtime_error("'format' argument to strftime_now() must be a string");
                 time_t nowUnix = Clock::to_time_t(Clock::now());
                 auto localDate = *std::localtime(&nowUnix);
                 std::ostringstream ss;
-                ss << std::put_time(&localDate, formatArg->second.asString().c_str());
+                ss << std::put_time(&localDate, format.c_str());
                 return ss.str();
-            },
-            /*argsInfo*/ { jinja2::ArgInfo("format", /*isMandatory*/ true) }
+            }
+        ));
+        environment->set("regex_replace", minja::simple_function(
+            "regex_replace", { "str", "pattern", "repl" },
+            [](const std::shared_ptr<minja::Context> &, minja::Value &args) -> minja::Value {
+                auto str     = args.at("str"    ).get<std::string>();
+                auto pattern = args.at("pattern").get<std::string>();
+                auto repl    = args.at("repl"   ).get<std::string>();
+                return std::regex_replace(str, std::regex(pattern), repl);
+            }
         ));
     }
-    return &*environment;
+    return environment;
+}
+
+class BaseResponseHandler {
+public:
+    virtual void onSplitIntoTwo    (const QString &startTag, const QString &firstBuffer, const QString &secondBuffer) = 0;
+    virtual void onSplitIntoThree  (const QString &secondBuffer, const QString &thirdBuffer)                          = 0;
+    // "old-style" responses, with all of the implementation details left in
+    virtual void onOldResponseChunk(const QByteArray &chunk)                                                          = 0;
+    // notify of a "new-style" response that has been cleaned of tool calling
+    virtual bool onBufferResponse  (const QString &response, int bufferIdx)                                           = 0;
+    // notify of a "new-style" response, no tool calling applicable
+    virtual bool onRegularResponse ()                                                                                 = 0;
+    virtual bool getStopGenerating () const                                                                           = 0;
+};
+
+static auto promptModelWithTools(
+    LLModel *model, const LLModel::PromptCallback &promptCallback, BaseResponseHandler &respHandler,
+    const LLModel::PromptContext &ctx, const QByteArray &prompt, const QStringList &toolNames
+) -> std::pair<QStringList, bool>
+{
+    ToolCallParser toolCallParser(toolNames);
+    auto handleResponse = [&toolCallParser, &respHandler](LLModel::Token token, std::string_view piece) -> bool {
+        Q_UNUSED(token)
+
+        toolCallParser.update(piece.data());
+
+        // Split the response into two if needed
+        if (toolCallParser.numberOfBuffers() < 2 && toolCallParser.splitIfPossible()) {
+            const auto parseBuffers = toolCallParser.buffers();
+            Q_ASSERT(parseBuffers.size() == 2);
+            respHandler.onSplitIntoTwo(toolCallParser.startTag(), parseBuffers.at(0), parseBuffers.at(1));
+        }
+
+        // Split the response into three if needed
+        if (toolCallParser.numberOfBuffers() < 3 && toolCallParser.startTag() == ToolCallConstants::ThinkStartTag
+            && toolCallParser.splitIfPossible()) {
+            const auto parseBuffers = toolCallParser.buffers();
+            Q_ASSERT(parseBuffers.size() == 3);
+            respHandler.onSplitIntoThree(parseBuffers.at(1), parseBuffers.at(2));
+        }
+
+        respHandler.onOldResponseChunk(QByteArray::fromRawData(piece.data(), piece.size()));
+
+        bool ok;
+        const auto parseBuffers = toolCallParser.buffers();
+        if (parseBuffers.size() > 1) {
+            ok = respHandler.onBufferResponse(parseBuffers.last(), parseBuffers.size() - 1);
+        } else {
+            ok = respHandler.onRegularResponse();
+        }
+        if (!ok)
+            return false;
+
+        const bool shouldExecuteToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete
+            && toolCallParser.startTag() != ToolCallConstants::ThinkStartTag;
+
+        return !shouldExecuteToolCall && !respHandler.getStopGenerating();
+    };
+    model->prompt(std::string_view(prompt), promptCallback, handleResponse, ctx);
+
+    const bool shouldExecuteToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete
+        && toolCallParser.startTag() != ToolCallConstants::ThinkStartTag;
+
+    return { toolCallParser.buffers(), shouldExecuteToolCall };
 }
 
 class LLModelStore {
@@ -738,7 +803,8 @@ std::vector<MessageItem> ChatLLM::forkConversation(const QString &prompt) const
         conversation.reserve(items.size() + 1);
         conversation.assign(items.begin(), items.end());
     }
-    conversation.emplace_back(MessageItem::Type::Prompt, prompt.toUtf8());
+    qsizetype nextIndex = conversation.empty() ? 0 : conversation.back().index().value() + 1;
+    conversation.emplace_back(nextIndex, MessageItem::Type::Prompt, prompt.toUtf8());
     return conversation;
 }
 
@@ -757,19 +823,18 @@ static uint parseJinjaTemplateVersion(QStringView tmpl)
     return 0;
 }
 
-static auto loadJinjaTemplate(
-    std::optional<jinja2::Template> &tmpl /*out*/, const std::string &source
-) -> jinja2::Result<void>
+static std::shared_ptr<minja::TemplateNode> loadJinjaTemplate(const std::string &source)
 {
-    tmpl.emplace(jinjaEnv());
-    return tmpl->Load(source);
+    return minja::Parser::parse(source, { .trim_blocks = true, .lstrip_blocks = true, .keep_trailing_newline = false });
 }
 
 std::optional<std::string> ChatLLM::checkJinjaTemplateError(const std::string &source)
 {
-    std::optional<jinja2::Template> tmpl;
-    if (auto res = loadJinjaTemplate(tmpl, source); !res)
-        return res.error().ToString();
+    try {
+        loadJinjaTemplate(source);
+    } catch (const std::runtime_error &e) {
+        return e.what();
+    }
     return std::nullopt;
 }
 
@@ -801,29 +866,29 @@ std::string ChatLLM::applyJinjaTemplate(std::span<const MessageItem> items) cons
     uint version = parseJinjaTemplateVersion(chatTemplate);
 
     auto makeMap = [version](const MessageItem &item) {
-        return jinja2::GenericMap([msg = std::make_shared<JinjaMessage>(version, item)] { return msg.get(); });
+        return JinjaMessage(version, item).AsJson();
     };
 
     std::unique_ptr<MessageItem> systemItem;
     bool useSystem = !isAllSpace(systemMessage);
 
-    jinja2::ValuesList messages;
+    json::array_t messages;
     messages.reserve(useSystem + items.size());
     if (useSystem) {
-        systemItem = std::make_unique<MessageItem>(MessageItem::Type::System, systemMessage.toUtf8());
+        systemItem = std::make_unique<MessageItem>(MessageItem::system_tag, systemMessage.toUtf8());
         messages.emplace_back(makeMap(*systemItem));
     }
     for (auto &item : items)
         messages.emplace_back(makeMap(item));
 
-    jinja2::ValuesList toolList;
+    json::array_t toolList;
     const int toolCount = ToolModel::globalInstance()->count();
     for (int i = 0; i < toolCount; ++i) {
         Tool *t = ToolModel::globalInstance()->get(i);
         toolList.push_back(t->jinjaValue());
     }
 
-    jinja2::ValuesMap params {
+    json::object_t params {
         { "messages",              std::move(messages) },
         { "add_generation_prompt", true                },
         { "toolList",              toolList            },
@@ -831,12 +896,14 @@ std::string ChatLLM::applyJinjaTemplate(std::span<const MessageItem> items) cons
     for (auto &[name, token] : model->specialTokens())
         params.emplace(std::move(name), std::move(token));
 
-    std::optional<jinja2::Template> tmpl;
-    auto maybeRendered = loadJinjaTemplate(tmpl, chatTemplate.toStdString())
-        .and_then([&] { return tmpl->RenderAsString(params); });
-    if (!maybeRendered)
-        throw std::runtime_error(fmt::format("Failed to parse chat template: {}", maybeRendered.error().ToString()));
-    return *maybeRendered;
+    try {
+        auto tmpl = loadJinjaTemplate(chatTemplate.toStdString());
+        auto context = minja::Context::make(minja::Value(std::move(params)), jinjaEnv());
+        return tmpl->render(context);
+    } catch (const std::runtime_error &e) {
+        throw std::runtime_error(fmt::format("Failed to parse chat template: {}", e.what()));
+    }
+    Q_UNREACHABLE();
 }
 
 auto ChatLLM::promptInternalChat(const QStringList &enabledCollections, const LLModel::PromptContext &ctx,
@@ -862,7 +929,7 @@ auto ChatLLM::promptInternalChat(const QStringList &enabledCollections, const LL
             // Find the prompt that represents the query. Server chats are flexible and may not have one.
             auto items = getChat();
             if (auto peer = m_chatModel->getPeer(items, items.end() - 1)) // peer of response
-                query = { *peer - items.begin(), (*peer)->content() };
+                query = { (*peer)->index().value(), (*peer)->content() };
         }
 
         if (query) {
@@ -888,6 +955,63 @@ auto ChatLLM::promptInternalChat(const QStringList &enabledCollections, const LL
     };
 }
 
+class ChatViewResponseHandler : public BaseResponseHandler {
+public:
+    ChatViewResponseHandler(ChatLLM *cllm, QElapsedTimer *totalTime, ChatLLM::PromptResult *result)
+        : m_cllm(cllm), m_totalTime(totalTime), m_result(result) {}
+
+    void onSplitIntoTwo(const QString &startTag, const QString &firstBuffer, const QString &secondBuffer) override
+    {
+        if (startTag == ToolCallConstants::ThinkStartTag)
+            m_cllm->m_chatModel->splitThinking({ firstBuffer, secondBuffer });
+        else
+            m_cllm->m_chatModel->splitToolCall({ firstBuffer, secondBuffer });
+    }
+
+    void onSplitIntoThree(const QString &secondBuffer, const QString &thirdBuffer) override
+    {
+        m_cllm->m_chatModel->endThinking({ secondBuffer, thirdBuffer }, m_totalTime->elapsed());
+    }
+
+    void onOldResponseChunk(const QByteArray &chunk) override
+    {
+        m_result->responseTokens++;
+        m_cllm->m_timer->inc();
+        m_result->response.append(chunk);
+    }
+
+    bool onBufferResponse(const QString &response, int bufferIdx) override
+    {
+        Q_UNUSED(bufferIdx)
+        try {
+            QString r = response;
+            m_cllm->m_chatModel->setResponseValue(removeLeadingWhitespace(r));
+        } catch (const std::exception &e) {
+            // We have a try/catch here because the main thread might have removed the response from
+            // the chatmodel by erasing the conversation during the response... the main thread sets
+            // m_stopGenerating before doing so, but it doesn't wait after that to reset the chatmodel
+            Q_ASSERT(m_cllm->m_stopGenerating);
+            return false;
+        }
+        emit m_cllm->responseChanged();
+        return true;
+    }
+
+    bool onRegularResponse() override
+    {
+        auto respStr = QString::fromUtf8(m_result->response);
+        return onBufferResponse(respStr, 0);
+    }
+
+    bool getStopGenerating() const override
+    { return m_cllm->m_stopGenerating; }
+
+private:
+    ChatLLM               *m_cllm;
+    QElapsedTimer         *m_totalTime;
+    ChatLLM::PromptResult *m_result;
+};
+
 auto ChatLLM::promptInternal(
     const std::variant<std::span<const MessageItem>, std::string_view> &prompt,
     const LLModel::PromptContext &ctx,
@@ -935,53 +1059,22 @@ auto ChatLLM::promptInternal(
         return !m_stopGenerating;
     };
 
-    ToolCallParser toolCallParser;
-    auto handleResponse = [this, &result, &toolCallParser](LLModel::Token token, std::string_view piece) -> bool {
-        Q_UNUSED(token)
-        result.responseTokens++;
-        m_timer->inc();
-
-        // FIXME: This is *not* necessarily fully formed utf data because it can be partial at this point
-        // handle this like below where we have a QByteArray
-        toolCallParser.update(QString::fromStdString(piece.data()));
-
-        // Create a toolcall and split the response if needed
-        if (!toolCallParser.hasSplit() && toolCallParser.state() == ToolEnums::ParseState::Partial) {
-            const QPair<QString, QString> pair = toolCallParser.split();
-            m_chatModel->splitToolCall(pair);
-        }
-
-        result.response.append(piece.data(), piece.size());
-        auto respStr = QString::fromUtf8(result.response);
-
-        try {
-            if (toolCallParser.hasSplit())
-                m_chatModel->setResponseValue(toolCallParser.buffer());
-            else
-                m_chatModel->setResponseValue(removeLeadingWhitespace(respStr));
-        } catch (const std::exception &e) {
-            // We have a try/catch here because the main thread might have removed the response from
-            // the chatmodel by erasing the conversation during the response... the main thread sets
-            // m_stopGenerating before doing so, but it doesn't wait after that to reset the chatmodel
-            Q_ASSERT(m_stopGenerating);
-            return false;
-        }
-
-        emit responseChanged();
-
-        const bool foundToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete;
-        return !foundToolCall && !m_stopGenerating;
-    };
-
     QElapsedTimer totalTime;
     totalTime.start();
-    m_timer->start();
+    ChatViewResponseHandler respHandler(this, &totalTime, &result);
 
+    m_timer->start();
+    QStringList finalBuffers;
+    bool        shouldExecuteTool;
     try {
         emit promptProcessing();
         m_llModelInfo.model->setThreadCount(mySettings->threadCount());
         m_stopGenerating = false;
-        m_llModelInfo.model->prompt(conversation, handlePrompt, handleResponse, ctx);
+        std::tie(finalBuffers, shouldExecuteTool) = promptModelWithTools(
+            m_llModelInfo.model.get(), handlePrompt, respHandler, ctx,
+            QByteArray::fromRawData(conversation.data(), conversation.size()),
+            ToolCallConstants::AllTagNames
+        );
     } catch (...) {
         m_timer->stop();
         throw;
@@ -990,20 +1083,18 @@ auto ChatLLM::promptInternal(
     m_timer->stop();
     qint64 elapsed = totalTime.elapsed();
 
-    const bool foundToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete;
-
     // trim trailing whitespace
     auto respStr = QString::fromUtf8(result.response);
-    if (!respStr.isEmpty() && (std::as_const(respStr).back().isSpace() || foundToolCall)) {
-        if (toolCallParser.hasSplit())
-            m_chatModel->setResponseValue(toolCallParser.buffer());
+    if (!respStr.isEmpty() && (std::as_const(respStr).back().isSpace() || finalBuffers.size() > 1)) {
+        if (finalBuffers.size() > 1)
+            m_chatModel->setResponseValue(finalBuffers.last().trimmed());
         else
             m_chatModel->setResponseValue(respStr.trimmed());
         emit responseChanged();
     }
 
     bool doQuestions = false;
-    if (!m_isServer && messageItems && !foundToolCall) {
+    if (!m_isServer && messageItems && !shouldExecuteTool) {
         switch (mySettings->suggestionMode()) {
             case SuggestionMode::On:            doQuestions = true;          break;
             case SuggestionMode::LocalDocsOnly: doQuestions = usedLocalDocs; break;
@@ -1081,6 +1172,66 @@ void ChatLLM::reloadModel()
         loadModel(m);
 }
 
+// This class throws discards the text within thinking tags, for use with chat names and follow-up questions.
+class SimpleResponseHandler : public BaseResponseHandler {
+public:
+    SimpleResponseHandler(ChatLLM *cllm)
+        : m_cllm(cllm) {}
+
+    void onSplitIntoTwo(const QString &startTag, const QString &firstBuffer, const QString &secondBuffer) override
+    { /* no-op */ }
+
+    void onSplitIntoThree(const QString &secondBuffer, const QString &thirdBuffer) override
+    { /* no-op */ }
+
+    void onOldResponseChunk(const QByteArray &chunk) override
+    { m_response.append(chunk); }
+
+    bool onBufferResponse(const QString &response, int bufferIdx) override
+    {
+        if (bufferIdx == 1)
+            return true; // ignore "think" content
+        return onSimpleResponse(response);
+    }
+
+    bool onRegularResponse() override
+    { return onBufferResponse(QString::fromUtf8(m_response), 0); }
+
+    bool getStopGenerating() const override
+    { return m_cllm->m_stopGenerating; }
+
+protected:
+    virtual bool onSimpleResponse(const QString &response) = 0;
+
+protected:
+    ChatLLM    *m_cllm;
+    QByteArray  m_response;
+};
+
+class NameResponseHandler : public SimpleResponseHandler {
+private:
+    // max length of chat names, in words
+    static constexpr qsizetype MAX_WORDS = 3;
+
+public:
+    using SimpleResponseHandler::SimpleResponseHandler;
+
+protected:
+    bool onSimpleResponse(const QString &response) override
+    {
+        QTextStream stream(const_cast<QString *>(&response), QIODeviceBase::ReadOnly);
+        QStringList words;
+        while (!stream.atEnd() && words.size() < MAX_WORDS) {
+            QString word;
+            stream >> word;
+            words << word;
+        }
+
+        emit m_cllm->generatedNameChanged(words.join(u' '));
+        return words.size() < MAX_WORDS || stream.atEnd();
+    }
+};
+
 void ChatLLM::generateName()
 {
     Q_ASSERT(isModelLoaded());
@@ -1097,23 +1248,15 @@ void ChatLLM::generateName()
         return;
     }
 
-    QByteArray response; // raw UTF-8
-
-    auto handleResponse = [this, &response](LLModel::Token token, std::string_view piece) -> bool {
-        Q_UNUSED(token)
-
-        response.append(piece.data(), piece.size());
-        QStringList words = QString::fromUtf8(response).simplified().split(u' ', Qt::SkipEmptyParts);
-        emit generatedNameChanged(words.join(u' '));
-        return words.size() <= 3;
-    };
+    NameResponseHandler respHandler(this);
 
     try {
-        m_llModelInfo.model->prompt(
-            applyJinjaTemplate(forkConversation(chatNamePrompt)),
-            [this](auto &&...) { return !m_stopGenerating; },
-            handleResponse,
-            promptContextFromSettings(m_modelInfo)
+        promptModelWithTools(
+            m_llModelInfo.model.get(),
+            /*promptCallback*/ [this](auto &&...) { return !m_stopGenerating; },
+            respHandler, promptContextFromSettings(m_modelInfo),
+            applyJinjaTemplate(forkConversation(chatNamePrompt)).c_str(),
+            { ToolCallConstants::ThinkTagName }
         );
     } catch (const std::exception &e) {
         qWarning() << "ChatLLM failed to generate name:" << e.what();
@@ -1125,13 +1268,43 @@ void ChatLLM::handleChatIdChanged(const QString &id)
     m_llmThread.setObjectName(id);
 }
 
-void ChatLLM::generateQuestions(qint64 elapsed)
-{
+class QuestionResponseHandler : public SimpleResponseHandler {
+public:
+    using SimpleResponseHandler::SimpleResponseHandler;
+
+protected:
+    bool onSimpleResponse(const QString &response) override
+    {
+        auto responseUtf8Bytes = response.toUtf8().slice(m_offset);
+        auto responseUtf8 = std::string(responseUtf8Bytes.begin(), responseUtf8Bytes.end());
+        // extract all questions from response
+        ptrdiff_t lastMatchEnd = -1;
+        auto it = std::sregex_iterator(responseUtf8.begin(), responseUtf8.end(), s_reQuestion);
+        auto end = std::sregex_iterator();
+        for (; it != end; ++it) {
+            auto pos = it->position();
+            auto len = it->length();
+            lastMatchEnd = pos + len;
+            emit m_cllm->generatedQuestionFinished(QString::fromUtf8(&responseUtf8[pos], len));
+        }
+
+        // remove processed input from buffer
+        if (lastMatchEnd != -1)
+            m_offset += lastMatchEnd;
+        return true;
+    }
+
+private:
     // FIXME: This only works with response by the model in english which is not ideal for a multi-language
     // model.
     // match whole question sentences
-    static const std::regex reQuestion(R"(\b(?:What|Where|How|Why|When|Who|Which|Whose|Whom)\b[^?]*\?)");
+    static inline const std::regex s_reQuestion { R"(\b(?:What|Where|How|Why|When|Who|Which|Whose|Whom)\b[^?]*\?)" };
 
+    qsizetype m_offset = 0;
+};
+
+void ChatLLM::generateQuestions(qint64 elapsed)
+{
     Q_ASSERT(isModelLoaded());
     if (!isModelLoaded()) {
         emit responseStopped(elapsed);
@@ -1149,39 +1322,17 @@ void ChatLLM::generateQuestions(qint64 elapsed)
 
     emit generatingQuestions();
 
-    std::string response; // raw UTF-8
-
-    auto handleResponse = [this, &response](LLModel::Token token, std::string_view piece) -> bool {
-        Q_UNUSED(token)
-
-        // add token to buffer
-        response.append(piece);
-
-        // extract all questions from response
-        ptrdiff_t lastMatchEnd = -1;
-        auto it = std::sregex_iterator(response.begin(), response.end(), reQuestion);
-        auto end = std::sregex_iterator();
-        for (; it != end; ++it) {
-            auto pos = it->position();
-            auto len = it->length();
-            lastMatchEnd = pos + len;
-            emit generatedQuestionFinished(QString::fromUtf8(&response[pos], len));
-        }
-
-        // remove processed input from buffer
-        if (lastMatchEnd != -1)
-            response.erase(0, lastMatchEnd);
-        return true;
-    };
+    QuestionResponseHandler respHandler(this);
 
     QElapsedTimer totalTime;
     totalTime.start();
     try {
-        m_llModelInfo.model->prompt(
-            applyJinjaTemplate(forkConversation(suggestedFollowUpPrompt)),
-            [this](auto &&...) { return !m_stopGenerating; },
-            handleResponse,
-            promptContextFromSettings(m_modelInfo)
+        promptModelWithTools(
+            m_llModelInfo.model.get(),
+            /*promptCallback*/ [this](auto &&...) { return !m_stopGenerating; },
+            respHandler, promptContextFromSettings(m_modelInfo),
+            applyJinjaTemplate(forkConversation(suggestedFollowUpPrompt)).c_str(),
+            { ToolCallConstants::ThinkTagName }
         );
     } catch (const std::exception &e) {
         qWarning() << "ChatLLM failed to generate follow-up questions:" << e.what();

gpt4all-chat/src/chatllm.h

@@ -2,7 +2,7 @@
 #define CHATLLM_H
 
 #include "chatmodel.h"
-#include "database.h" // IWYU pragma: keep
+#include "database.h"
 #include "modellist.h"
 
 #include <gpt4all-backend/llmodel.h>
@@ -10,28 +10,30 @@
 #include <QByteArray>
 #include <QElapsedTimer>
 #include <QFileInfo>
-#include <QList>      // IWYU pragma: keep
+#include <QList>
 #include <QObject>
 #include <QPointer>
 #include <QString>
 #include <QStringList> // IWYU pragma: keep
-#include <QStringView>
 #include <QThread>
 #include <QVariantMap> // IWYU pragma: keep
-#include <QtGlobal>
+#include <QtNumeric>
 
 #include <atomic>
-#include <cstdint>
 #include <memory>
 #include <optional>
 #include <span>
 #include <string>
+#include <string_view>
 #include <variant>
+#include <vector>
 
 using namespace Qt::Literals::StringLiterals;
 
+class ChatLLM;
 class QDataStream;
 
+
 // NOTE: values serialized to disk, do not change or reuse
 enum class LLModelTypeV0 { // chat versions 2-5
     MPT       = 0,
@@ -88,9 +90,6 @@ inline LLModelTypeV1 parseLLModelTypeV0(int v0)
     }
 }
 
-class ChatLLM;
-class ChatModel;
-
 struct LLModelInfo {
     std::unique_ptr<LLModel> model;
     QFileInfo fileInfo;
@@ -285,6 +284,8 @@ private:
     bool m_isServer;
     bool m_forceMetal;
     bool m_reloadingToChangeVariant;
+    friend class ChatViewResponseHandler;
+    friend class SimpleResponseHandler;
 };
 
 #endif // CHATLLM_H

gpt4all-chat/src/chatmodel.cpp

@@ -2,9 +2,11 @@
 
 #include <QDebug>
 #include <QMap>
-#include <QtGlobal>
+#include <QTextStream>
 #include <QtLogging>
 
+#include <exception>
+
 
 QList<ResultInfo> ChatItem::consolidateSources(const QList<ResultInfo> &sources)
 {
@@ -41,6 +43,12 @@ void ChatItem::serializeText(QDataStream &stream, int version)
     stream << value;
 }
 
+void ChatItem::serializeThink(QDataStream &stream, int version)
+{
+    stream << value;
+    stream << thinkingTime;
+}
+
 void ChatItem::serializeSubItems(QDataStream &stream, int version)
 {
     stream << name;
@@ -50,6 +58,7 @@ void ChatItem::serializeSubItems(QDataStream &stream, int version)
         case ToolCall:      { serializeToolCall(stream, version);       break; }
         case ToolResponse:  { serializeToolResponse(stream, version);   break; }
         case Text:          { serializeText(stream, version);           break; }
+        case Think:         { serializeThink(stream, version);          break; }
         case System:
         case Prompt:
             throw std::invalid_argument(fmt::format("cannot serialize subitem type {}", int(typ)));
@@ -162,6 +171,13 @@ bool ChatItem::deserializeResponse(QDataStream &stream, int version)
     return true;
 }
 
+bool ChatItem::deserializeThink(QDataStream &stream, int version)
+{
+    stream >> value;
+    stream >> thinkingTime;
+    return true;
+}
+
 bool ChatItem::deserializeSubItems(QDataStream &stream, int version)
 {
     stream >> name;
@@ -177,6 +193,7 @@ bool ChatItem::deserializeSubItems(QDataStream &stream, int version)
         case ToolCall:      { deserializeToolCall(stream, version); break; }
         case ToolResponse:  { deserializeToolResponse(stream, version); break; }
         case Text:          { deserializeText(stream, version); break; }
+        case Think:         { deserializeThink(stream, version); break; }
         case System:
         case Prompt:
             throw std::invalid_argument(fmt::format("cannot serialize subitem type {}", int(typ)));

gpt4all-chat/src/chatmodel.h

@@ -4,32 +4,41 @@
 #include "database.h"
 #include "tool.h"
 #include "toolcallparser.h"
-#include "utils.h"
+#include "utils.h" // IWYU pragma: keep
 #include "xlsxtomd.h"
 
 #include <fmt/format.h>
 
-#include <QApplication>
 #include <QAbstractListModel>
 #include <QBuffer>
 #include <QByteArray>
 #include <QClipboard>
 #include <QDataStream>
-#include <QJsonDocument>
+#include <QFileInfo>
+#include <QGuiApplication>
+#include <QIODevice>
 #include <QHash>
 #include <QList>
+#include <QMutex>
+#include <QMutexLocker> // IWYU pragma: keep
 #include <QObject>
-#include <QPair>
+#include <QPair> // IWYU pragma: keep
 #include <QString>
+#include <QStringList> // IWYU pragma: keep
+#include <QUrl>
 #include <QVariant>
-#include <QVector>
 #include <Qt>
-#include <QtGlobal>
+#include <QtAssert>
+#include <QtPreprocessorSupport>
+#include <QtTypes>
 
 #include <algorithm>
 #include <iterator>
+#include <list>
+#include <optional>
 #include <ranges>
 #include <span>
+#include <stdexcept>
 #include <utility>
 #include <vector>
 
@@ -94,11 +103,28 @@ class MessageItem
 public:
     enum class Type { System, Prompt, Response, ToolResponse };
 
-    MessageItem(Type type, QString content)
-        : m_type(type), m_content(std::move(content)) {}
+    struct system_tag_t { explicit system_tag_t() = default; };
+    static inline constexpr system_tag_t system_tag = system_tag_t{};
 
-    MessageItem(Type type, QString content, const QList<ResultInfo> &sources, const QList<PromptAttachment> &promptAttachments)
-        : m_type(type), m_content(std::move(content)), m_sources(sources), m_promptAttachments(promptAttachments) {}
+    MessageItem(qsizetype index, Type type, QString content)
+        : m_index(index), m_type(type), m_content(std::move(content))
+    {
+        Q_ASSERT(type != Type::System); // use system_tag constructor
+    }
+
+    // Construct a system message with no index, since they are never stored in the chat
+    MessageItem(system_tag_t, QString content)
+        : m_type(Type::System), m_content(std::move(content)) {}
+
+    MessageItem(qsizetype index, Type type, QString content, const QList<ResultInfo> &sources, const QList<PromptAttachment> &promptAttachments)
+        : m_index(index)
+        , m_type(type)
+        , m_content(std::move(content))
+        , m_sources(sources)
+        , m_promptAttachments(promptAttachments) {}
+
+    // index of the parent ChatItem (system, prompt, response) in its container
+    std::optional<qsizetype> index()   const { return m_index; }
 
     Type           type()    const { return m_type;    }
     const QString &content() const { return m_content; }
@@ -126,10 +152,11 @@ public:
     }
 
 private:
-    Type                    m_type;
-    QString                 m_content;
-    QList<ResultInfo>       m_sources;
-    QList<PromptAttachment> m_promptAttachments;
+    std::optional<qsizetype> m_index;
+    Type                     m_type;
+    QString                  m_content;
+    QList<ResultInfo>        m_sources;
+    QList<PromptAttachment>  m_promptAttachments;
 };
 Q_DECLARE_METATYPE(MessageItem)
 
@@ -159,8 +186,11 @@ class ChatItem : public QObject
     Q_PROPERTY(bool    thumbsUpState   MEMBER thumbsUpState  )
     Q_PROPERTY(bool    thumbsDownState MEMBER thumbsDownState)
 
+    // thinking
+    Q_PROPERTY(int     thinkingTime    MEMBER thinkingTime NOTIFY thinkingTimeChanged)
+
 public:
-    enum class Type { System, Prompt, Response, Text, ToolCall, ToolResponse };
+    enum class Type { System, Prompt, Response, Text, ToolCall, ToolResponse, Think };
 
     // tags for constructing ChatItems
     struct prompt_tag_t        { explicit prompt_tag_t       () = default; };
@@ -169,19 +199,22 @@ public:
     struct text_tag_t          { explicit text_tag_t         () = default; };
     struct tool_call_tag_t     { explicit tool_call_tag_t    () = default; };
     struct tool_response_tag_t { explicit tool_response_tag_t() = default; };
+    struct think_tag_t         { explicit think_tag_t        () = default; };
     static inline constexpr prompt_tag_t        prompt_tag        = prompt_tag_t        {};
     static inline constexpr response_tag_t      response_tag      = response_tag_t      {};
     static inline constexpr system_tag_t        system_tag        = system_tag_t        {};
     static inline constexpr text_tag_t          text_tag          = text_tag_t          {};
     static inline constexpr tool_call_tag_t     tool_call_tag     = tool_call_tag_t     {};
     static inline constexpr tool_response_tag_t tool_response_tag = tool_response_tag_t {};
+    static inline constexpr think_tag_t         think_tag         = think_tag_t         {};
 
 public:
     ChatItem(QObject *parent)
         : QObject(nullptr)
     {
         moveToThread(parent->thread());
-        setParent(parent);
+        // setParent must be called from the thread the object lives in
+        QMetaObject::invokeMethod(this, [this, parent]() { this->setParent(parent); });
     }
 
     // NOTE: System messages are currently never serialized and only *stored* by the local server.
@@ -220,6 +253,10 @@ public:
         : ChatItem(parent)
     { this->name = u"ToolResponse: "_s; this->value = value; }
 
+    ChatItem(QObject *parent, think_tag_t, const QString &value)
+        : ChatItem(parent)
+    { this->name = u"Think: "_s; this->value = value; }
+
     Type type() const
     {
         if (name == u"System: "_s)
@@ -234,6 +271,8 @@ public:
             return Type::ToolCall;
         if (name == u"ToolResponse: "_s)
             return Type::ToolResponse;
+        if (name == u"Think: "_s)
+            return Type::Think;
         throw std::invalid_argument(fmt::format("Chat item has unknown label: {:?}", name));
     }
 
@@ -254,7 +293,7 @@ public:
         if (type() == Type::Response) {
             // We parse if this contains any part of a partial toolcall
             ToolCallParser parser;
-            parser.update(value);
+            parser.update(value.toUtf8());
 
             // If no tool call is detected, return the original value
             if (parser.startIndex() < 0)
@@ -265,9 +304,11 @@ public:
             return beforeToolCall;
         }
 
-        // For tool calls we only return content if it is the code interpreter
+        if (type() == Type::Think)
+            return thinkContent(value);
+
         if (type() == Type::ToolCall)
-            return codeInterpreterContent(value);
+            return toolCallContent(value);
 
         // We don't show any of content from the tool response in the GUI
         if (type() == Type::ToolResponse)
@@ -276,10 +317,21 @@ public:
         return value;
     }
 
-    QString codeInterpreterContent(const QString &value) const
+    QString thinkContent(const QString &value) const
     {
         ToolCallParser parser;
-        parser.update(value);
+        parser.update(value.toUtf8());
+
+        // Extract the content
+        QString content = parser.toolCall();
+        content = content.trimmed();
+        return content;
+    }
+
+    QString toolCallContent(const QString &value) const
+    {
+        ToolCallParser parser;
+        parser.update(value.toUtf8());
 
         // Extract the code
         QString code = parser.toolCall();
@@ -357,6 +409,12 @@ public:
         return toolCallInfo.error != ToolEnums::Error::NoError;
     }
 
+    void setThinkingTime(int t)
+    {
+        thinkingTime = t;
+        emit thinkingTimeChanged();
+    }
+
     // NB: Assumes response is not current.
     static ChatItem *fromMessageInput(QObject *parent, const MessageInput &message)
     {
@@ -369,7 +427,7 @@ public:
         Q_UNREACHABLE();
     }
 
-    MessageItem asMessageItem() const
+    MessageItem asMessageItem(qsizetype index) const
     {
         MessageItem::Type msgType;
         switch (auto typ = type()) {
@@ -380,9 +438,10 @@ public:
             case ToolResponse: msgType = MessageItem::Type::ToolResponse; break;
             case Text:
             case ToolCall:
+            case Think:
                 throw std::invalid_argument(fmt::format("cannot convert ChatItem type {} to message item", int(typ)));
         }
-        return { msgType, flattenedContent(), sources, promptAttachments };
+        return { index, msgType, flattenedContent(), sources, promptAttachments };
     }
 
     static QList<ResultInfo> consolidateSources(const QList<ResultInfo> &sources);
@@ -391,6 +450,7 @@ public:
     void serializeToolCall(QDataStream &stream, int version);
     void serializeToolResponse(QDataStream &stream, int version);
     void serializeText(QDataStream &stream, int version);
+    void serializeThink(QDataStream &stream, int version);
     void serializeSubItems(QDataStream &stream, int version); // recursive
     void serialize(QDataStream &stream, int version);
 
@@ -399,6 +459,7 @@ public:
     bool deserializeToolCall(QDataStream &stream, int version);
     bool deserializeToolResponse(QDataStream &stream, int version);
     bool deserializeText(QDataStream &stream, int version);
+    bool deserializeThink(QDataStream &stream, int version);
     bool deserializeSubItems(QDataStream &stream, int version); // recursive
     bool deserialize(QDataStream &stream, int version);
 
@@ -406,6 +467,7 @@ Q_SIGNALS:
     void contentChanged();
     void isTooCallErrorChanged();
     void isCurrentResponseChanged();
+    void thinkingTimeChanged();
 
 public:
 
@@ -429,6 +491,9 @@ public:
     bool    stopped         = false;
     bool    thumbsUpState   = false;
     bool    thumbsDownState = false;
+
+    // thinking time in ms
+    int     thinkingTime  = 0;
 };
 
 class ChatModel : public QAbstractListModel
@@ -500,6 +565,7 @@ private:
         return std::nullopt;
     }
 
+    // FIXME(jared): this should really be done at the parent level, not the sub-item level
     static std::optional<qsizetype> getPeerInternal(const MessageItem *arr, qsizetype size, qsizetype index)
     {
         qsizetype peer;
@@ -879,6 +945,70 @@ public:
         if (changed) emit dataChanged(createIndex(index, 0), createIndex(index, 0), {NewResponseRole});
     }
 
+    Q_INVOKABLE void splitThinking(const QPair<QString, QString> &split)
+    {
+        qsizetype index;
+        {
+            QMutexLocker locker(&m_mutex);
+            if (m_chatItems.isEmpty() || m_chatItems.cend()[-1]->type() != ChatItem::Type::Response)
+                throw std::logic_error("can only set thinking on a chat that ends with a response");
+
+            index = m_chatItems.count() - 1;
+            ChatItem *currentResponse = m_chatItems.back();
+            Q_ASSERT(currentResponse->isCurrentResponse);
+
+            // Create a new response container for any text and the thinking
+            ChatItem *newResponse = new ChatItem(this, ChatItem::response_tag);
+
+            // Add preceding text if any
+            if (!split.first.isEmpty()) {
+                ChatItem *textItem = new ChatItem(this, ChatItem::text_tag, split.first);
+                newResponse->subItems.push_back(textItem);
+            }
+
+            // Add the thinking item
+            Q_ASSERT(!split.second.isEmpty());
+            ChatItem *thinkingItem = new ChatItem(this, ChatItem::think_tag, split.second);
+            thinkingItem->isCurrentResponse = true;
+            newResponse->subItems.push_back(thinkingItem);
+
+            // Add new response and reset our value
+            currentResponse->subItems.push_back(newResponse);
+            currentResponse->value = QString();
+        }
+
+        emit dataChanged(createIndex(index, 0), createIndex(index, 0), {ChildItemsRole, ContentRole});
+    }
+
+    Q_INVOKABLE void endThinking(const QPair<QString, QString> &split, int thinkingTime)
+    {
+        qsizetype index;
+        {
+            QMutexLocker locker(&m_mutex);
+            if (m_chatItems.isEmpty() || m_chatItems.cend()[-1]->type() != ChatItem::Type::Response)
+                throw std::logic_error("can only end thinking on a chat that ends with a response");
+
+            index = m_chatItems.count() - 1;
+            ChatItem *currentResponse = m_chatItems.back();
+            Q_ASSERT(currentResponse->isCurrentResponse);
+
+            ChatItem *subResponse = currentResponse->subItems.back();
+            Q_ASSERT(subResponse->type() == ChatItem::Type::Response);
+            Q_ASSERT(subResponse->isCurrentResponse);
+            subResponse->setCurrentResponse(false);
+
+            ChatItem *thinkingItem = subResponse->subItems.back();
+            Q_ASSERT(thinkingItem->type() == ChatItem::Type::Think);
+            thinkingItem->setCurrentResponse(false);
+            thinkingItem->setValue(split.first);
+            thinkingItem->setThinkingTime(thinkingTime);
+
+            currentResponse->setValue(split.second);
+        }
+
+        emit dataChanged(createIndex(index, 0), createIndex(index, 0), {ChildItemsRole, ContentRole});
+    }
+
     Q_INVOKABLE void splitToolCall(const QPair<QString, QString> &split)
     {
         qsizetype index;
@@ -1013,10 +1143,12 @@ public:
         // A flattened version of the chat item tree used by the backend and jinja
         QMutexLocker locker(&m_mutex);
         std::vector<MessageItem> chatItems;
-        for (const ChatItem *item : m_chatItems) {
-            chatItems.reserve(chatItems.size() + item->subItems.size() + 1);
-            ranges::copy(item->subItems | views::transform(&ChatItem::asMessageItem), std::back_inserter(chatItems));
-            chatItems.push_back(item->asMessageItem());
+        for (qsizetype i : views::iota(0, m_chatItems.size())) {
+            auto *parent = m_chatItems.at(i);
+            chatItems.reserve(chatItems.size() + parent->subItems.size() + 1);
+            ranges::copy(parent->subItems | views::transform([&](auto *s) { return s->asMessageItem(i); }),
+                         std::back_inserter(chatItems));
+            chatItems.push_back(parent->asMessageItem(i));
         }
         return chatItems;
     }

gpt4all-chat/src/chatviewtextprocessor.cpp

@@ -1,29 +1,32 @@
 #include "chatviewtextprocessor.h"
 
+#include <QAbstractTextDocumentLayout>
 #include <QBrush>
 #include <QChar>
 #include <QClipboard>
+#include <QDebug>
+#include <QFlag>
 #include <QFont>
-#include <QFontMetricsF>
 #include <QGuiApplication>
-#include <QList>
-#include <QPainter>
+#include <QList> // IWYU pragma: keep
+#include <QPair>
 #include <QQuickTextDocument>
 #include <QRegularExpression>
-#include <QStringList>
-#include <QTextBlock>
-#include <QTextCharFormat>
+#include <QStringList> // IWYU pragma: keep
+#include <QTextBlock> // IWYU pragma: keep
+#include <QTextCharFormat> // IWYU pragma: keep
 #include <QTextCursor>
 #include <QTextDocument>
 #include <QTextDocumentFragment>
-#include <QTextFrame>
-#include <QTextFrameFormat>
+#include <QTextFrame> // IWYU pragma: keep
+#include <QTextFrameFormat> // IWYU pragma: keep
 #include <QTextTableCell>
-#include <QVariant>
-#include <Qt>
-#include <QtGlobal>
+#include <QtAssert>
+#include <QtLogging>
 
 #include <algorithm>
+#include <utility>
+
 
 enum Language {
     None,

gpt4all-chat/src/chatviewtextprocessor.h

@@ -3,18 +3,15 @@
 
 #include <QColor>
 #include <QObject>
-#include <QQmlEngine>
-#include <QQuickTextDocument> // IWYU pragma: keep
-#include <QRectF>
-#include <QSizeF>
+#include <QQmlEngine> // IWYU pragma: keep
+#include <QQuickTextDocument>
 #include <QString>
 #include <QSyntaxHighlighter>
-#include <QTextObjectInterface>
-#include <QVector>
+#include <QVector> // IWYU pragma: keep
+#include <QtTypes>
+
+// IWYU pragma: no_forward_declare QQuickTextDocument
 
-class QPainter;
-class QTextDocument;
-class QTextFormat;
 
 struct CodeColors {
     Q_GADGET

gpt4all-chat/src/codeinterpreter.cpp

@@ -1,14 +1,25 @@
 #include "codeinterpreter.h"
 
+#include <QJSEngine>
 #include <QJSValue>
-#include <QStringList>
+#include <QList>
+#include <QStringList> // IWYU pragma: keep
 #include <QThread>
 #include <QVariant>
+#include <Qt>
 
 using namespace Qt::Literals::StringLiterals;
 
 
-QString CodeInterpreter::run(const QList<ToolParam> &params, qint64 timeout)
+CodeInterpreter::CodeInterpreter()
+    : Tool()
+    , m_error(ToolEnums::Error::NoError)
+{
+    m_worker = new CodeInterpreterWorker;
+    connect(this, &CodeInterpreter::request, m_worker, &CodeInterpreterWorker::request, Qt::QueuedConnection);
+}
+
+void CodeInterpreter::run(const QList<ToolParam> &params)
 {
     m_error = ToolEnums::Error::NoError;
     m_errorString = QString();
@@ -18,27 +29,24 @@ QString CodeInterpreter::run(const QList<ToolParam> &params, qint64 timeout)
           && params.first().type == ToolEnums::ParamType::String);
 
     const QString code = params.first().value.toString();
-
-    QThread workerThread;
-    CodeInterpreterWorker worker;
-    worker.moveToThread(&workerThread);
-    connect(&worker, &CodeInterpreterWorker::finished, &workerThread, &QThread::quit, Qt::DirectConnection);
-    connect(&workerThread, &QThread::started, [&worker, code]() {
-        worker.request(code);
+    connect(m_worker, &CodeInterpreterWorker::finished, [this, params] {
+        m_error = m_worker->error();
+        m_errorString = m_worker->errorString();
+        emit runComplete({
+        ToolCallConstants::CodeInterpreterFunction,
+            params,
+            m_worker->response(),
+            m_error,
+            m_errorString
+        });
     });
-    workerThread.start();
-    bool timedOut = !workerThread.wait(timeout);
-    if (timedOut) {
-        worker.interrupt(timeout); // thread safe
-        m_error = ToolEnums::Error::TimeoutError;
-    }
-    workerThread.quit();
-    workerThread.wait();
-    if (!timedOut) {
-        m_error = worker.error();
-        m_errorString = worker.errorString();
-    }
-    return worker.response();
+
+    emit request(code);
+}
+
+bool CodeInterpreter::interrupt()
+{
+    return m_worker->interrupt();
 }
 
 QList<ToolParamInfo> CodeInterpreter::parameters() const
@@ -89,20 +97,20 @@ QString CodeInterpreter::exampleReply() const
 
 CodeInterpreterWorker::CodeInterpreterWorker()
     : QObject(nullptr)
+    , m_engine(new QJSEngine(this))
 {
-}
+    moveToThread(&m_thread);
 
-void CodeInterpreterWorker::request(const QString &code)
-{
-    JavaScriptConsoleCapture consoleCapture;
-    QJSValue consoleInternalObject = m_engine.newQObject(&consoleCapture);
-    m_engine.globalObject().setProperty("console_internal", consoleInternalObject);
+    QJSValue consoleInternalObject = m_engine->newQObject(&m_consoleCapture);
+    m_engine->globalObject().setProperty("console_internal", consoleInternalObject);
 
     // preprocess console.log args in JS since Q_INVOKE doesn't support varargs
-    auto consoleObject = m_engine.evaluate(uR"(
+    auto consoleObject = m_engine->evaluate(uR"(
         class Console {
             log(...args) {
-                if (args.length && typeof args[0] === 'string')
+                if (args.length == 0)
+                    return;
+                if (args.length >= 2 && typeof args[0] === 'string')
                     throw new Error('console.log string formatting not supported');
                 let cat = '';
                 for (const arg of args) {
@@ -114,15 +122,28 @@ void CodeInterpreterWorker::request(const QString &code)
 
         new Console();
     )"_s);
-    m_engine.globalObject().setProperty("console", consoleObject);
+    m_engine->globalObject().setProperty("console", consoleObject);
+    m_thread.start();
+}
 
-    const QJSValue result = m_engine.evaluate(code);
+void CodeInterpreterWorker::reset()
+{
+    m_response.clear();
+    m_error = ToolEnums::Error::NoError;
+    m_errorString.clear();
+    m_consoleCapture.output.clear();
+    m_engine->setInterrupted(false);
+}
 
+void CodeInterpreterWorker::request(const QString &code)
+{
+    reset();
+    const QJSValue result = m_engine->evaluate(code);
     QString resultString;
 
-    if (m_engine.isInterrupted()) {
-        resultString = QString("Error: code execution was timed out as it exceeded %1 ms. Code must be written to ensure execution does not timeout.").arg(m_timeout);
-    } else if (result.isError()) {
+    if (m_engine->isInterrupted()) {
+        resultString = QString("Error: code execution was interrupted or timed out.");
+   } else if (result.isError()) {
         // NOTE: We purposely do not set the m_error or m_errorString for the code interpreter since
         // we *want* the model to see the response has an error so it can hopefully correct itself. The
         // error member variables are intended for tools that have error conditions that cannot be corrected.
@@ -143,9 +164,16 @@ void CodeInterpreterWorker::request(const QString &code)
     }
 
     if (resultString.isEmpty())
-        resultString = consoleCapture.output;
-    else if (!consoleCapture.output.isEmpty())
-        resultString += "\n" + consoleCapture.output;
+        resultString = m_consoleCapture.output;
+    else if (!m_consoleCapture.output.isEmpty())
+        resultString += "\n" + m_consoleCapture.output;
     m_response = resultString;
     emit finished();
 }
+
+bool CodeInterpreterWorker::interrupt()
+{
+    m_error = ToolEnums::Error::TimeoutError;
+    m_engine->setInterrupted(true);
+    return true;
+}

gpt4all-chat/src/codeinterpreter.h

@@ -4,10 +4,13 @@
 #include "tool.h"
 #include "toolcallparser.h"
 
-#include <QJSEngine>
 #include <QObject>
 #include <QString>
-#include <QtGlobal>
+#include <QThread>
+#include <QtAssert>
+
+class QJSEngine;
+
 
 class JavaScriptConsoleCapture : public QObject
 {
@@ -39,32 +42,37 @@ public:
     CodeInterpreterWorker();
     virtual ~CodeInterpreterWorker() {}
 
+    void reset();
     QString response() const { return m_response; }
-
-    void request(const QString &code);
-    void interrupt(qint64 timeout) { m_timeout = timeout; m_engine.setInterrupted(true); }
     ToolEnums::Error error() const { return m_error; }
     QString errorString() const { return m_errorString; }
+    bool interrupt();
+
+public Q_SLOTS:
+    void request(const QString &code);
 
 Q_SIGNALS:
     void finished();
 
 private:
-    qint64 m_timeout = 0;
-    QJSEngine m_engine;
     QString m_response;
     ToolEnums::Error m_error = ToolEnums::Error::NoError;
     QString m_errorString;
+    QThread m_thread;
+    JavaScriptConsoleCapture m_consoleCapture;
+    QJSEngine *m_engine = nullptr;
 };
 
 class CodeInterpreter : public Tool
 {
     Q_OBJECT
 public:
-    explicit CodeInterpreter() : Tool(), m_error(ToolEnums::Error::NoError) {}
+    explicit CodeInterpreter();
     virtual ~CodeInterpreter() {}
 
-    QString run(const QList<ToolParam> &params, qint64 timeout = 2000) override;
+    void run(const QList<ToolParam> &params) override;
+    bool interrupt() override;
+
     ToolEnums::Error error() const override { return m_error; }
     QString errorString() const override { return m_errorString; }
 
@@ -77,9 +85,13 @@ public:
     QString exampleCall() const override;
     QString exampleReply() const override;
 
+Q_SIGNALS:
+    void request(const QString &code);
+
 private:
     ToolEnums::Error m_error = ToolEnums::Error::NoError;
     QString m_errorString;
+    CodeInterpreterWorker *m_worker;
 };
 
 #endif // CODEINTERPRETER_H

gpt4all-chat/src/config.h.in

@@ -0,0 +1,7 @@
+#pragma once
+
+#define APP_VERSION "@APP_VERSION@"
+
+#define G4A_CONFIG(name) (1/G4A_CONFIG_##name == 1)
+
+#define G4A_CONFIG_force_d3d12 @GPT4ALL_CONFIG_FORCE_D3D12@

gpt4all-chat/src/database.cpp

@@ -1,19 +1,21 @@
 #include "database.h"
 
 #include "mysettings.h"
-#include "utils.h"
+#include "utils.h" // IWYU pragma: keep
 
 #include <duckx/duckx.hpp>
 #include <fmt/format.h>
+#include <usearch/index.hpp>
 #include <usearch/index_plugins.hpp>
 
-#include <QByteArrayView>
 #include <QDebug>
 #include <QDir>
 #include <QDirIterator>
 #include <QFile>
 #include <QFileSystemWatcher>
+#include <QFlags>
 #include <QIODevice>
+#include <QKeyValueIterator>
 #include <QRegularExpression>
 #include <QSqlError>
 #include <QSqlQuery>
@@ -22,8 +24,9 @@
 #include <QMap>
 #include <QUtf8StringView>
 #include <QVariant>
-#include <Qt>
 #include <QtLogging>
+#include <QtMinMax>
+#include <QtTypes>
 
 #include <algorithm>
 #include <cmath>
@@ -46,6 +49,7 @@ namespace us = unum::usearch;
 //#define DEBUG
 //#define DEBUG_EXAMPLE
 
+
 namespace {
 
 /* QFile that checks input for binary data. If seen, it fails the read and returns true
@@ -1111,9 +1115,9 @@ class DocumentReader {
 public:
     struct Metadata { QString title, author, subject, keywords; };
 
-    static std::unique_ptr<DocumentReader> fromDocument(const DocumentInfo &info);
+    static std::unique_ptr<DocumentReader> fromDocument(DocumentInfo info);
 
-    const DocumentInfo           &doc     () const { return *m_info; }
+    const DocumentInfo           &doc     () const { return m_info; }
     const Metadata               &metadata() const { return m_metadata; }
     const std::optional<QString> &word    () const { return m_word; }
     const std::optional<QString> &nextWord()       { m_word = advance(); return m_word; }
@@ -1123,8 +1127,8 @@ public:
     virtual ~DocumentReader() = default;
 
 protected:
-    explicit DocumentReader(const DocumentInfo &info)
-        : m_info(&info) {}
+    explicit DocumentReader(DocumentInfo info)
+        : m_info(std::move(info)) {}
 
     void postInit(Metadata &&metadata = {})
     {
@@ -1134,9 +1138,9 @@ protected:
 
     virtual std::optional<QString> advance() = 0;
 
-    const DocumentInfo     *m_info;
-    Metadata                m_metadata;
-    std::optional<QString>  m_word;
+    DocumentInfo           m_info;
+    Metadata               m_metadata;
+    std::optional<QString> m_word;
 };
 
 namespace {
@@ -1144,8 +1148,8 @@ namespace {
 #ifdef GPT4ALL_USE_QTPDF
 class PdfDocumentReader final : public DocumentReader {
 public:
-    explicit PdfDocumentReader(const DocumentInfo &info)
-        : DocumentReader(info)
+    explicit PdfDocumentReader(DocumentInfo info)
+        : DocumentReader(std::move(info))
     {
         QString path = info.file.canonicalFilePath();
         if (m_doc.load(path) != QPdfDocument::Error::None)
@@ -1185,8 +1189,8 @@ private:
 #else
 class PdfDocumentReader final : public DocumentReader {
 public:
-    explicit PdfDocumentReader(const DocumentInfo &info)
-        : DocumentReader(info)
+    explicit PdfDocumentReader(DocumentInfo info)
+        : DocumentReader(std::move(info))
     {
         QString path = info.file.canonicalFilePath();
         m_doc = FPDF_LoadDocument(path.toUtf8().constData(), nullptr);
@@ -1209,7 +1213,6 @@ public:
             FPDF_ClosePage(m_page);
         if (m_doc)
             FPDF_CloseDocument(m_doc);
-        FPDF_DestroyLibrary();
     }
 
     int page() const override { return m_currentPage; }
@@ -1224,7 +1227,7 @@ private:
                     return std::nullopt;
 
                 if (m_page)
-                    FPDF_ClosePage(m_page);
+                    FPDF_ClosePage(std::exchange(m_page, nullptr));
                 m_page = FPDF_LoadPage(m_doc, m_currentPage++);
                 if (!m_page)
                     throw std::runtime_error("Failed to load page.");
@@ -1278,8 +1281,8 @@ private:
 
 class WordDocumentReader final : public DocumentReader {
 public:
-    explicit WordDocumentReader(const DocumentInfo &info)
-        : DocumentReader(info)
+    explicit WordDocumentReader(DocumentInfo info)
+        : DocumentReader(std::move(info))
         , m_doc(info.file.canonicalFilePath().toStdString())
     {
         m_doc.open();
@@ -1371,8 +1374,8 @@ protected:
 
 class TxtDocumentReader final : public DocumentReader {
 public:
-    explicit TxtDocumentReader(const DocumentInfo &info)
-        : DocumentReader(info)
+    explicit TxtDocumentReader(DocumentInfo info)
+        : DocumentReader(std::move(info))
         , m_file(info.file.canonicalFilePath())
     {
         if (!m_file.open(QIODevice::ReadOnly))
@@ -1413,13 +1416,13 @@ protected:
 
 } // namespace
 
-std::unique_ptr<DocumentReader> DocumentReader::fromDocument(const DocumentInfo &doc)
+std::unique_ptr<DocumentReader> DocumentReader::fromDocument(DocumentInfo doc)
 {
     if (doc.isPdf())
-        return std::make_unique<PdfDocumentReader>(doc);
+        return std::make_unique<PdfDocumentReader>(std::move(doc));
     if (doc.isDocx())
-        return std::make_unique<WordDocumentReader>(doc);
-    return std::make_unique<TxtDocumentReader>(doc);
+        return std::make_unique<WordDocumentReader>(std::move(doc));
+    return std::make_unique<TxtDocumentReader>(std::move(doc));
 }
 
 ChunkStreamer::ChunkStreamer(Database *database)
@@ -1427,12 +1430,12 @@ ChunkStreamer::ChunkStreamer(Database *database)
 
 ChunkStreamer::~ChunkStreamer() = default;
 
-void ChunkStreamer::setDocument(const DocumentInfo &doc, int documentId, const QString &embeddingModel)
+void ChunkStreamer::setDocument(DocumentInfo doc, int documentId, const QString &embeddingModel)
 {
     auto docKey = doc.key();
     if (!m_docKey || *m_docKey != docKey) {
         m_docKey         = docKey;
-        m_reader         = DocumentReader::fromDocument(doc);
+        m_reader         = DocumentReader::fromDocument(std::move(doc));
         m_documentId     = documentId;
         m_embeddingModel = embeddingModel;
         m_chunk.clear();
@@ -1442,7 +1445,8 @@ void ChunkStreamer::setDocument(const DocumentInfo &doc, int documentId, const Q
         if (m_database->m_documentIdCache.contains(documentId)) {
             QSqlQuery q(m_database->m_db);
             if (!m_database->removeChunksByDocumentId(q, documentId))
-                handleDocumentError("ERROR: Cannot remove chunks of document", documentId, doc.file.canonicalPath(), q.lastError());
+                handleDocumentError("ERROR: Cannot remove chunks of document",
+                                    documentId, m_reader->doc().file.canonicalPath(), q.lastError());
         }
     }
 }

gpt4all-chat/src/database.h

@@ -1,7 +1,7 @@
 #ifndef DATABASE_H
 #define DATABASE_H
 
-#include "embllm.h" // IWYU pragma: keep
+#include "embllm.h"
 
 #include <QByteArray>
 #include <QChar>
@@ -15,11 +15,11 @@
 #include <QSet>
 #include <QSqlDatabase>
 #include <QString>
-#include <QStringList>
+#include <QStringList> // IWYU pragma: keep
 #include <QThread>
 #include <QUrl>
-#include <QVector>
-#include <QtGlobal>
+#include <QVector> // IWYU pragma: keep
+#include <QtAssert>
 
 #include <atomic>
 #include <cstddef>
@@ -28,7 +28,7 @@
 #include <memory>
 #include <optional>
 #include <utility>
-#include <vector>
+#include <vector> // IWYU pragma: keep
 
 using namespace Qt::Literals::StringLiterals;
 
@@ -39,6 +39,7 @@ class QSqlQuery;
 class QTextStream;
 class QTimer;
 
+
 /* Version 0: GPT4All v2.4.3, full-text search
  * Version 1: GPT4All v2.5.3, embeddings in hsnwlib
  * Version 2: GPT4All v3.0.0, embeddings in sqlite
@@ -171,7 +172,7 @@ public:
     explicit ChunkStreamer(Database *database);
     ~ChunkStreamer();
 
-    void setDocument(const DocumentInfo &doc, int documentId, const QString &embeddingModel);
+    void setDocument(DocumentInfo doc, int documentId, const QString &embeddingModel);
     std::optional<DocumentInfo::key_type> currentDocKey() const;
     void reset();
 

gpt4all-chat/src/download.cpp

@@ -10,32 +10,37 @@
 #include <QDebug>
 #include <QGlobalStatic>
 #include <QGuiApplication>
-#include <QIODevice>
+#include <QIODevice> // IWYU pragma: keep
 #include <QJsonArray>
 #include <QJsonDocument>
 #include <QJsonObject>
 #include <QJsonValue>
+#include <QKeyValueIterator>
 #include <QLocale>
 #include <QNetworkRequest>
-#include <QPair>
+#include <QPair> // IWYU pragma: keep
+#include <QRegularExpression>
+#include <QRegularExpressionMatch>
 #include <QSettings>
 #include <QSslConfiguration>
 #include <QSslSocket>
-#include <QStringList>
+#include <QStringList> // IWYU pragma: keep
 #include <QTextStream>
 #include <QUrl>
 #include <QVariant>
-#include <QVector>
+#include <QVector> // IWYU pragma: keep
 #include <Qt>
+#include <QtAssert>
 #include <QtLogging>
+#include <QtMinMax>
 
-#include <algorithm>
 #include <compare>
 #include <cstddef>
 #include <utility>
 
 using namespace Qt::Literals::StringLiterals;
 
+
 class MyDownload: public Download { };
 Q_GLOBAL_STATIC(MyDownload, downloadInstance)
 Download *Download::globalInstance()

gpt4all-chat/src/download.h

@@ -13,10 +13,14 @@
 #include <QSslError>
 #include <QString>
 #include <QThread>
-#include <QtGlobal>
+#include <QtTypes>
 
+// IWYU pragma: no_forward_declare QFile
+// IWYU pragma: no_forward_declare QList
+// IWYU pragma: no_forward_declare QSslError
 class QByteArray;
 
+
 struct ReleaseInfo {
     Q_GADGET
     Q_PROPERTY(QString version MEMBER version)

gpt4all-chat/src/embllm.cpp

@@ -1,35 +1,35 @@
 #include "embllm.h"
 
-#include "modellist.h"
 #include "mysettings.h"
 
 #include <gpt4all-backend/llmodel.h>
 
 #include <QCoreApplication>
 #include <QDebug>
-#include <QFile>
 #include <QFileInfo>
 #include <QGuiApplication>
-#include <QIODevice>
 #include <QJsonArray>
 #include <QJsonDocument>
 #include <QJsonObject>
+#include <QJsonValue>
 #include <QList>
-#include <QMutexLocker>
+#include <QMutexLocker> // IWYU pragma: keep
 #include <QNetworkAccessManager>
 #include <QNetworkReply>
 #include <QNetworkRequest>
 #include <QUrl>
 #include <Qt>
-#include <QtGlobal>
+#include <QtAssert>
 #include <QtLogging>
 
 #include <exception>
+#include <string>
 #include <utility>
 #include <vector>
 
 using namespace Qt::Literals::StringLiterals;
 
+
 static const QString EMBEDDING_MODEL_NAME = u"nomic-embed-text-v1.5"_s;
 static const QString LOCAL_EMBEDDING_MODEL = u"nomic-embed-text-v1.5.f16.gguf"_s;
 
@@ -359,8 +359,11 @@ void EmbeddingLLMWorker::handleFinished()
     if (retrievedData.isValid() && retrievedData.canConvert<QVector<EmbeddingChunk>>())
         chunks = retrievedData.value<QVector<EmbeddingChunk>>();
 
-    QVariant response = reply->attribute(QNetworkRequest::HttpStatusCodeAttribute);
-    Q_ASSERT(response.isValid());
+    QVariant response;
+    if (reply->error() != QNetworkReply::NoError) {
+        response = reply->attribute(QNetworkRequest::HttpStatusCodeAttribute);
+        Q_ASSERT(response.isValid());
+    }
     bool ok;
     int code = response.toInt(&ok);
     if (!ok || code != 200) {

gpt4all-chat/src/embllm.h

@@ -5,10 +5,10 @@
 #include <QMutex>
 #include <QObject>
 #include <QString>
-#include <QStringList>
+#include <QStringList> // IWYU pragma: keep
 #include <QThread>
 #include <QVariant>
-#include <QVector>
+#include <QVector> // IWYU pragma: keep
 
 #include <atomic>
 #include <vector>
@@ -16,6 +16,7 @@
 class LLModel;
 class QNetworkAccessManager;
 
+
 struct EmbeddingChunk {
     QString model; // TODO(jared): use to select model
     int folder_id;

gpt4all-chat/src/jinja_helpers.cpp

@@ -1,117 +1,76 @@
 #include "jinja_helpers.h"
 
-#include "utils.h"
-
-#include <fmt/format.h>
-
 #include <QString>
 #include <QUrl>
 
-#include <memory>
-#include <vector>
+#include <ranges>
+#include <string>
+#include <utility>
 
-using namespace std::literals::string_view_literals;
+namespace views  = std::views;
+using json = nlohmann::ordered_json;
 
 
-JinjaResultInfo::~JinjaResultInfo() = default;
-
-const JinjaFieldMap<ResultInfo> JinjaResultInfo::s_fields = {
-    { "collection", [](auto &s) { return s.collection.toStdString(); } },
-    { "path",       [](auto &s) { return s.path      .toStdString(); } },
-    { "file",       [](auto &s) { return s.file      .toStdString(); } },
-    { "title",      [](auto &s) { return s.title     .toStdString(); } },
-    { "author",     [](auto &s) { return s.author    .toStdString(); } },
-    { "date",       [](auto &s) { return s.date      .toStdString(); } },
-    { "text",       [](auto &s) { return s.text      .toStdString(); } },
-    { "page",       [](auto &s) { return s.page;                     } },
-    { "file_uri",   [](auto &s) { return s.fileUri() .toStdString(); } },
-};
-
-JinjaPromptAttachment::~JinjaPromptAttachment() = default;
-
-const JinjaFieldMap<PromptAttachment> JinjaPromptAttachment::s_fields = {
-    { "url",               [](auto &s) { return s.url.toString()    .toStdString(); } },
-    { "file",              [](auto &s) { return s.file()            .toStdString(); } },
-    { "processed_content", [](auto &s) { return s.processedContent().toStdString(); } },
-};
-
-std::vector<std::string> JinjaMessage::GetKeys() const
+json::object_t JinjaResultInfo::AsJson() const
 {
-    std::vector<std::string> result;
-    auto &keys = this->keys();
-    result.reserve(keys.size());
-    result.assign(keys.begin(), keys.end());
-    return result;
+    return {
+        { "collection", m_source->collection.toStdString() },
+        { "path",       m_source->path      .toStdString() },
+        { "file",       m_source->file      .toStdString() },
+        { "title",      m_source->title     .toStdString() },
+        { "author",     m_source->author    .toStdString() },
+        { "date",       m_source->date      .toStdString() },
+        { "text",       m_source->text      .toStdString() },
+        { "page",       m_source->page                     },
+        { "file_uri",   m_source->fileUri() .toStdString() },
+    };
 }
 
-auto JinjaMessage::keys() const -> const std::unordered_set<std::string_view> &
+json::object_t JinjaPromptAttachment::AsJson() const
 {
-    static const std::unordered_set<std::string_view> baseKeys
-        { "role", "content" };
-    static const std::unordered_set<std::string_view> userKeys
-        { "role", "content", "sources", "prompt_attachments" };
-    switch (m_item->type()) {
-        using enum MessageItem::Type;
-    case System:
-    case Response:
-    case ToolResponse:
-        return baseKeys;
-    case Prompt:
-        return userKeys;
-        break;
-    }
-    Q_UNREACHABLE();
+    return {
+        { "url",               m_attachment->url.toString()    .toStdString() },
+        { "file",              m_attachment->file()            .toStdString() },
+        { "processed_content", m_attachment->processedContent().toStdString() },
+    };
 }
 
-bool operator==(const JinjaMessage &a, const JinjaMessage &b)
+json::object_t JinjaMessage::AsJson() const
 {
-    if (a.m_item == b.m_item)
-        return true;
-    const auto &[ia, ib] = std::tie(*a.m_item, *b.m_item);
-    auto type = ia.type();
-    if (type != ib.type() || ia.content() != ib.content())
-        return false;
-
-    switch (type) {
-        using enum MessageItem::Type;
-    case System:
-    case Response:
-    case ToolResponse:
-        return true;
-    case Prompt:
-        return ia.sources() == ib.sources() && ia.promptAttachments() == ib.promptAttachments();
-        break;
-    }
-    Q_UNREACHABLE();
-}
-
-const JinjaFieldMap<JinjaMessage> JinjaMessage::s_fields = {
-    { "role", [](auto &m) {
-        switch (m.item().type()) {
+    json::object_t obj;
+    {
+        json::string_t role;
+        switch (m_item->type()) {
             using enum MessageItem::Type;
-            case System:   return "system"sv;
-            case Prompt:   return "user"sv;
-            case Response: return "assistant"sv;
-            case ToolResponse: return "tool"sv;
-                break;
+            case System:       role = "system";    break;
+            case Prompt:       role = "user";      break;
+            case Response:     role = "assistant"; break;
+            case ToolResponse: role = "tool";      break;
         }
-        Q_UNREACHABLE();
-    } },
-    { "content", [](auto &m) {
-        if (m.version() == 0 && m.item().type() == MessageItem::Type::Prompt)
-            return m.item().bakedPrompt().toStdString();
-        return m.item().content().toStdString();
-    } },
-    { "sources", [](auto &m) {
-        auto sources = m.item().sources() | views::transform([](auto &r) {
-            return jinja2::GenericMap([map = std::make_shared<JinjaResultInfo>(r)] { return map.get(); });
-        });
-        return jinja2::ValuesList(sources.begin(), sources.end());
-    } },
-    { "prompt_attachments", [](auto &m) {
-        auto attachments = m.item().promptAttachments() | views::transform([](auto &pa) {
-            return jinja2::GenericMap([map = std::make_shared<JinjaPromptAttachment>(pa)] { return map.get(); });
-        });
-        return jinja2::ValuesList(attachments.begin(), attachments.end());
-    } },
-};
+        obj.emplace_back("role", std::move(role));
+    }
+    {
+        QString content;
+        if (m_version == 0 && m_item->type() == MessageItem::Type::Prompt) {
+            content = m_item->bakedPrompt();
+        } else {
+            content = m_item->content();
+        }
+        obj.emplace_back("content", content.toStdString());
+    }
+    if (m_item->type() == MessageItem::Type::Prompt) {
+        {
+            auto sources = m_item->sources() | views::transform([](auto &r) {
+                return JinjaResultInfo(r).AsJson();
+            });
+            obj.emplace("sources", json::array_t(sources.begin(), sources.end()));
+        }
+        {
+            auto attachments = m_item->promptAttachments() | views::transform([](auto &pa) {
+                return JinjaPromptAttachment(pa).AsJson();
+            });
+            obj.emplace("prompt_attachments", json::array_t(attachments.begin(), attachments.end()));
+        }
+    }
+    return obj;
+}

gpt4all-chat/src/jinja_helpers.h

@@ -3,47 +3,21 @@
 #include "chatmodel.h"
 #include "database.h"
 
-#include <jinja2cpp/value.h>
+#include <nlohmann/json.hpp>
 
-#include <functional>
-#include <ranges>
-#include <string>
-#include <string_view>
-#include <unordered_map>
-#include <unordered_set>
+#include <QtTypes> // IWYU pragma: keep
 
-#include <QtGlobal>
+// IWYU pragma: no_forward_declare MessageItem
+// IWYU pragma: no_forward_declare PromptAttachment
+// IWYU pragma: no_forward_declare ResultInfo
 
-namespace views = std::views;
+using json = nlohmann::ordered_json;
 
 
-template <typename T>
-using JinjaFieldMap = std::unordered_map<std::string_view, std::function<jinja2::Value (const T &)>>;
-
 template <typename Derived>
-class JinjaComparable : public jinja2::IMapItemAccessor {
+class JinjaHelper {
 public:
-    JinjaComparable() = default;
-
-    bool IsEqual(const jinja2::IComparable &other) const override;
-
-private:
-    Q_DISABLE_COPY_MOVE(JinjaComparable)
-};
-
-template <typename Derived>
-class JinjaHelper : public JinjaComparable<Derived> {
-public:
-    size_t GetSize() const override
-    { return Derived::s_fields.size(); }
-
-    bool HasValue(const std::string &name) const override
-    { return Derived::s_fields.contains(name); }
-
-    jinja2::Value GetValueByName(const std::string &name) const override;
-
-    std::vector<std::string> GetKeys() const override
-    { auto keys = views::elements<0>(Derived::s_fields); return { keys.begin(), keys.end() }; }
+    json::object_t AsJson() const { return static_cast<const Derived *>(this)->AsJson(); }
 };
 
 class JinjaResultInfo : public JinjaHelper<JinjaResultInfo> {
@@ -51,18 +25,10 @@ public:
     explicit JinjaResultInfo(const ResultInfo &source) noexcept
         : m_source(&source) {}
 
-    ~JinjaResultInfo() override;
-
-    const ResultInfo &value() const { return *m_source; }
-
-    friend bool operator==(const JinjaResultInfo &a, const JinjaResultInfo &b)
-    { return a.m_source == b.m_source || *a.m_source == *b.m_source; }
+    json::object_t AsJson() const;
 
 private:
-    static const JinjaFieldMap<ResultInfo> s_fields;
     const ResultInfo *m_source;
-
-    friend class JinjaHelper<JinjaResultInfo>;
 };
 
 class JinjaPromptAttachment : public JinjaHelper<JinjaPromptAttachment> {
@@ -70,18 +36,10 @@ public:
     explicit JinjaPromptAttachment(const PromptAttachment &attachment) noexcept
         : m_attachment(&attachment) {}
 
-    ~JinjaPromptAttachment() override;
-
-    const PromptAttachment &value() const { return *m_attachment; }
-
-    friend bool operator==(const JinjaPromptAttachment &a, const JinjaPromptAttachment &b)
-    { return a.m_attachment == b.m_attachment || *a.m_attachment == *b.m_attachment; }
+    json::object_t AsJson() const;
 
 private:
-    static const JinjaFieldMap<PromptAttachment> s_fields;
     const PromptAttachment *m_attachment;
-
-    friend class JinjaHelper<JinjaPromptAttachment>;
 };
 
 class JinjaMessage : public JinjaHelper<JinjaMessage> {
@@ -89,28 +47,9 @@ public:
     explicit JinjaMessage(uint version, const MessageItem &item) noexcept
         : m_version(version), m_item(&item) {}
 
-    const JinjaMessage &value  () const { return *this;     }
-    uint                version() const { return m_version; }
-    const MessageItem      &item  () const { return *m_item;   }
-
-    size_t GetSize() const override { return keys().size(); }
-    bool HasValue(const std::string &name) const override { return keys().contains(name); }
-
-    jinja2::Value GetValueByName(const std::string &name) const override
-    { return HasValue(name) ? JinjaHelper::GetValueByName(name) : jinja2::EmptyValue(); }
-
-    std::vector<std::string> GetKeys() const override;
+    json::object_t AsJson() const;
 
 private:
-    auto keys() const -> const std::unordered_set<std::string_view> &;
-
-private:
-    static const JinjaFieldMap<JinjaMessage> s_fields;
-    uint            m_version;
+    uint               m_version;
     const MessageItem *m_item;
-
-    friend class JinjaHelper<JinjaMessage>;
-    friend bool operator==(const JinjaMessage &a, const JinjaMessage &b);
 };
-
-#include "jinja_helpers.inl"

gpt4all-chat/src/jinja_helpers.inl

@@ -1,17 +0,0 @@
-template <typename D>
-bool JinjaComparable<D>::IsEqual(const jinja2::IComparable &other) const
-{
-    if (auto *omsg = dynamic_cast<const D *>(&other))
-        return *static_cast<const D *>(this) == *omsg;
-    return false;
-}
-
-template <typename D>
-jinja2::Value JinjaHelper<D>::GetValueByName(const std::string &name) const
-{
-    if (auto it = D::s_fields.find(name); it != D::s_fields.end()) {
-        auto [_, func] = *it;
-        return func(static_cast<const D *>(this)->value());
-    }
-    return jinja2::EmptyValue();
-}

gpt4all-chat/src/jinja_replacements.cpp

@@ -2,29 +2,15 @@
 
 #include "jinja_replacements.h"
 
+#include <utility>
+
+
 // This is a list of prompt templates known to GPT4All and their associated replacements which are automatically used
 // instead when loading the chat template from GGUF. These exist for two primary reasons:
 // - HuggingFace model authors make ugly chat templates because they do not expect the end user to see them;
-// - and our Jinja2Cpp-based template parsing is not fully compatible with HuggingFace transformers and jinja2.
-
-// Below is a list of known incompatibilities with the official HF jinja2 implementation. These are not all necessarily
-// reflected in the below substitution list, and this cannot be an exhaustive list because there are a plethora of edge
-// cases in template parsing in which jinja2 and Jinja2Cpp differ. These are differences that could be reasonably
-// expected to affect chat templates that could be seen in the wild, or that cause a crash:
-// - Jinja2Cpp crashes (in debug builds) if given the template `a[""(`
-// - Jinja2Cpp does not support these jinja2 constructs:
-//   - `is not none`
-//   - list slicing, e.g. `messages[1:]`
-//   - the jinja2.ext.loopcontrols extension, which HF enables by default
-//   - a missing space after a quote in substitution (e.g. `{{ 'foo'}}`), which *has* been seen in the wild
-// - GPT4All does not currently support these HuggingFace template features:
-//   - customized "tojson" filter (we provide the built-in Jinja2Cpp one)
-//   - the AssistantTracker extension
-
+// - and chat templates occasionally use features we do not support. This is less true now that we use minja.
 
 // The substitution list.
-// For templates that apply to models listed in models3.json, these should be copied there as well for best
-// compatibility with older versions of GPT4All.
 
 const std::unordered_map<std::string_view, std::string_view> CHAT_TEMPLATE_SUBSTITUTIONS {
     // calme-2.1-phi3.5-4b.Q6_K.gguf (reported by ThilotE on Discord), Phi-3.5-mini-instruct-Q4_0.gguf (nomic-ai/gpt4all#3345)
@@ -52,6 +38,30 @@ const std::unordered_map<std::string_view, std::string_view> CHAT_TEMPLATE_SUBST
     {{- '<|assistant|>\n' }}
 {%- else %}
     {{- eos_token }}
+{%- endif %})TEMPLATE",
+    },
+    // DeepSeek-R1-Distill-Qwen-7B-Q4_0.gguf
+    {
+        // original
+        R"TEMPLATE({% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>'}}{% endif %})TEMPLATE",
+        // replacement
+        R"TEMPLATE({%- if not add_generation_prompt is defined %}
+    {%- set add_generation_prompt = false %}
+{%- endif %}
+{%- if messages[0]['role'] == 'system' %}
+    {{- messages[0]['content'] }}
+{%- endif %}
+{%- for message in messages %}
+    {%- if message['role'] == 'user' %}
+        {{- '<｜User｜>' + message['content'] }}
+    {%- endif %}
+    {%- if message['role'] == 'assistant' %}
+        {%- set content = message['content'] | regex_replace('^[\\s\\S]*</think>', '') %}
+        {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' }}
+    {%- endif %}
+{%- endfor -%}
+{%- if add_generation_prompt %}
+    {{- '<｜Assistant｜>' }}
 {%- endif %})TEMPLATE",
     },
     // gemma-2-9b-it-Q4_0.gguf (nomic-ai/gpt4all#3282)
@@ -106,11 +116,31 @@ const std::unordered_map<std::string_view, std::string_view> CHAT_TEMPLATE_SUBST
     {%- elif message['role'] == 'system' %}
         {{- '<|system|>\n' + message['content'] + eos_token }}
     {%- elif message['role'] == 'assistant' %}
-        {{- '<|assistant|>\n'  + message['content'] + eos_token }}
+        {{- '<|assistant|>\n' + message['content'] + eos_token }}
     {%- endif %}
     {%- if loop.last and add_generation_prompt %}
         {{- '<|assistant|>' }}
     {%- endif %}
+{%- endfor %})TEMPLATE",
+    },
+    // granite-3.1-3b-a800m-instruct-Q4_0.gguf, granite-3.1-8b-instruct-Q4_0.gguf (nomic-ai/gpt4all#3471)
+    {
+        // original
+        R"TEMPLATE({%- if messages[0]['role'] == 'system' %}{%- set system_message = messages[0]['content'] %}{%- set loop_messages = messages[1:] %}{%- else %}{%- set system_message = "Knowledge Cutoff Date: April 2024. You are Granite, developed by IBM." %}{%- if tools and documents %}{%- set system_message = system_message + " You are a helpful AI assistant with access to the following tools. When a tool is required to answer the user's query, respond with <|tool_call|> followed by a JSON list of tools used. If a tool does not exist in the provided list of tools, notify the user that you do not have the ability to fulfill the request. Write the response to the user's input by strictly aligning with the facts in the provided documents. If the information needed to answer the question is not available in the documents, inform the user that the question cannot be answered based on the available data." %}{%- elif tools %}{%- set system_message = system_message + " You are a helpful AI assistant with access to the following tools. When a tool is required to answer the user's query, respond with <|tool_call|> followed by a JSON list of tools used. If a tool does not exist in the provided list of tools, notify the user that you do not have the ability to fulfill the request." %}{%- elif documents %}{%- set system_message = system_message + " Write the response to the user's input by strictly aligning with the facts in the provided documents. If the information needed to answer the question is not available in the documents, inform the user that the question cannot be answered based on the available data." %}{%- else %}{%- set system_message = system_message + " You are a helpful AI assistant." %}{%- endif %}{%- if controls and 'citations' in controls and documents %}{%- set system_message = system_message + ' In your response, use the symbols <co> and </co> to indicate when a fact comes from a document in the search result, e.g <co>0</co> for a fact from document 0. Afterwards, list all the citations with their corresponding documents in an ordered list.' %}{%- endif %}{%- if controls and 'hallucinations' in controls and documents %}{%- set system_message = system_message + ' Finally, after the response is written, include a numbered list of sentences from the response that are potentially hallucinated and not based in the documents.' %}{%- endif %}{%- set loop_messages = messages %}{%- endif %}{{- '<|start_of_role|>system<|end_of_role|>' + system_message + '<|end_of_text|> ' }}{%- if tools %}{{- '<|start_of_role|>tools<|end_of_role|>' }}{{- tools | tojson(indent=4) }}{{- '<|end_of_text|> ' }}{%- endif %}{%- if documents %}{{- '<|start_of_role|>documents<|end_of_role|>' }}{%- for document in documents %}{{- 'Document ' + loop.index0 | string + ' ' }}{{- document['text'] }}{%- if not loop.last %}{{- ' '}}{%- endif%}{%- endfor %}{{- '<|end_of_text|> ' }}{%- endif %}{%- for message in loop_messages %}{{- '<|start_of_role|>' + message['role'] + '<|end_of_role|>' + message['content'] + '<|end_of_text|> ' }}{%- if loop.last and add_generation_prompt %}{{- '<|start_of_role|>assistant' }}{%- if controls %}{{- ' ' + controls | tojson()}}{%- endif %}{{- '<|end_of_role|>' }}{%- endif %}{%- endfor %})TEMPLATE",
+        // replacement
+        R"TEMPLATE({%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content'] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "Knowledge Cutoff Date: April 2024. You are Granite, developed by IBM. You are a helpful AI assistant." %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{{- '<|start_of_role|>system<|end_of_role|>' + system_message + '<|end_of_text|> ' }}
+{%- for message in loop_messages %}
+    {{- '<|start_of_role|>' + message['role'] + '<|end_of_role|>' + message['content'] + '<|end_of_text|> ' }}
+    {%- if loop.last and add_generation_prompt %}
+        {{- '<|start_of_role|>assistant<|end_of_role|>' }}
+    {%- endif %}
 {%- endfor %})TEMPLATE",
     },
     // Hermes-3-Llama-3.2-3B.Q4_0.gguf, mistral-7b-openorca.gguf2.Q4_0.gguf
@@ -611,6 +641,70 @@ const std::unordered_map<std::string_view, std::string_view> CHAT_TEMPLATE_SUBST
 {%- if add_generation_prompt %}
     {{- '<|im_start|>assistant\n' }}
 {%- endif %})TEMPLATE",
+    },
+    // OLMoE-1B-7B-0125-Instruct-Q4_0.gguf (nomic-ai/gpt4all#3471)
+    {
+        // original
+        R"TEMPLATE({{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'user' %}{{ '<|user|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
+'  + message['content'] + eos_token + '
+' }}{% else %}{{ '<|assistant|>
+'  + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
+' }}{% endif %}{% endfor %})TEMPLATE",
+        // replacement
+        R"TEMPLATE({{- bos_token }}
+{%- for message in messages %}
+    {%- if message['role'] == 'system' %}
+        {{- '<|system|>\n' + message['content'] + '\n' }}
+    {%- elif message['role'] == 'user' %}
+        {{- '<|user|>\n' + message['content'] + '\n' }}
+    {%- elif message['role'] == 'assistant' %}
+        {%- if not loop.last %}
+            {{- '<|assistant|>\n' + message['content'] + eos_token + '\n' }}
+        {%- else %}
+            {{- '<|assistant|>\n' + message['content'] + eos_token }}
+        {%- endif %}
+    {%- endif %}
+    {%- if loop.last and add_generation_prompt %}
+        {{- '<|assistant|>\n' }}
+    {%- endif %}
+{%- endfor %})TEMPLATE",
+    },
+    // OLMoE-1B-7B-0924-Instruct-Q4_0.gguf (nomic-ai/gpt4all#3471)
+    {
+        // original
+        R"TEMPLATE({{ bos_token }}{% for message in messages %}
+{% if message['role'] == 'system' %}
+{{ '<|system|>
+' + message['content'] }}
+{% elif message['role'] == 'user' %}
+{{ '<|user|>
+' + message['content'] }}
+{% elif message['role'] == 'assistant' %}
+{{ '<|assistant|>
+'  + message['content'] + eos_token }}
+{% endif %}
+{% if loop.last and add_generation_prompt %}
+{{ '<|assistant|>' }}
+{% endif %}
+{% endfor %})TEMPLATE",
+        // replacement
+        R"TEMPLATE({{- bos_token }}
+{%- for message in messages %}
+    {%- if message['role'] == 'system' %}
+        {{- '<|system|>\n' + message['content'] }}
+    {%- elif message['role'] == 'user' %}
+        {{- '<|user|>\n' + message['content'] }}
+    {%- elif message['role'] == 'assistant' %}
+        {{- '<|assistant|>\n' + message['content'] + eos_token }}
+    {%- endif %}
+    {%- if loop.last and add_generation_prompt %}
+        {{- '<|assistant|>' }}
+    {%- endif %}
+{%- endfor %})TEMPLATE",
     },
     // Phi-3.1-mini-128k-instruct-Q4_0.gguf (nomic-ai/gpt4all#3346)
     {
@@ -626,7 +720,7 @@ const std::unordered_map<std::string_view, std::string_view> CHAT_TEMPLATE_SUBST
         // replacement
         R"TEMPLATE({%- for message in messages %}
     {%- if message['role'] == 'system' %}
-        {{-'<|system|>\n' + message['content'] + '<|end|>\n'}}
+        {{- '<|system|>\n' + message['content'] + '<|end|>\n' }}
     {%- elif message['role'] == 'user' %}
         {{- '<|user|>\n' + message['content'] + '<|end|>\n' }}
     {%- elif message['role'] == 'assistant' %}

gpt4all-chat/src/llm.cpp

@@ -12,6 +12,9 @@
 #include <QSettings>
 #include <QUrl>
 #include <QtLogging>
+#include <QtSystemDetection>
+
+#include <string>
 
 #ifdef GPT4ALL_OFFLINE_INSTALLER
 #   include <QDesktopServices>
@@ -25,6 +28,7 @@
 
 using namespace Qt::Literals::StringLiterals;
 
+
 class MyLLM: public LLM { };
 Q_GLOBAL_STATIC(MyLLM, llmInstance)
 LLM *LLM::globalInstance()

gpt4all-chat/src/llm.h

@@ -3,7 +3,8 @@
 
 #include <QObject>
 #include <QString>
-#include <QtGlobal>
+#include <QtTypes>
+
 
 class LLM : public QObject
 {

gpt4all-chat/src/localdocs.cpp

@@ -5,10 +5,14 @@
 #include "mysettings.h"
 
 #include <QCoreApplication>
+#include <QDebug>
 #include <QGlobalStatic>
 #include <QGuiApplication>
+#include <QList>
 #include <QUrl>
 #include <Qt>
+#include <QtLogging>
+
 
 class MyLocalDocs: public LocalDocs { };
 Q_GLOBAL_STATIC(MyLocalDocs, localDocsInstance)

gpt4all-chat/src/localdocs.h

@@ -2,11 +2,14 @@
 #define LOCALDOCS_H
 
 #include "database.h"
-#include "localdocsmodel.h" // IWYU pragma: keep
+#include "localdocsmodel.h"
 
 #include <QObject>
 #include <QString>
-#include <QStringList>
+#include <QStringList> // IWYU pragma: keep
+
+// IWYU pragma: no_forward_declare LocalDocsModel
+
 
 class LocalDocs : public QObject
 {

gpt4all-chat/src/localdocsmodel.cpp

@@ -5,11 +5,11 @@
 
 #include <QDateTime>
 #include <QMap>
-#include <QVector>
-#include <QtGlobal>
+#include <QVector> // IWYU pragma: keep
 
 #include <utility>
 
+
 LocalDocsCollectionsModel::LocalDocsCollectionsModel(QObject *parent)
     : QSortFilterProxyModel(parent)
 {

gpt4all-chat/src/localdocsmodel.h

@@ -4,17 +4,19 @@
 #include "database.h"
 
 #include <QAbstractListModel>
-#include <QByteArray>
-#include <QHash>
 #include <QList>
-#include <QObject>
+#include <QObject> // IWYU pragma: keep
 #include <QSortFilterProxyModel>
 #include <QString>
-#include <QVariant>
 #include <Qt>
 
 #include <functional>
 
+class QByteArray;
+class QVariant;
+template <typename Key, typename T> class QHash;
+
+
 class LocalDocsCollectionsModel : public QSortFilterProxyModel
 {
     Q_OBJECT

gpt4all-chat/src/logger.cpp

@@ -2,8 +2,10 @@
 
 #include <QDateTime>
 #include <QDebug>
+#include <QFlags>
 #include <QGlobalStatic>
 #include <QIODevice>
+#include <QMutexLocker> // IWYU pragma: keep
 #include <QStandardPaths>
 
 #include <cstdio>
@@ -12,6 +14,7 @@
 
 using namespace Qt::Literals::StringLiterals;
 
+
 class MyLogger: public Logger { };
 Q_GLOBAL_STATIC(MyLogger, loggerInstance)
 Logger *Logger::globalInstance()
@@ -62,8 +65,11 @@ void Logger::messageHandler(QtMsgType type, const QMessageLogContext &, const QS
     }
     // Get time and date
     auto timestamp = QDateTime::currentDateTime().toString();
-    // Write message
+
     const std::string out = u"[%1] (%2): %3\n"_s.arg(typeString, timestamp, msg).toStdString();
+
+    // Write message
+    QMutexLocker locker(&logger->m_mutex);
     logger->m_file.write(out.c_str());
     logger->m_file.flush();
     std::cerr << out;

gpt4all-chat/src/logger.h

@@ -2,19 +2,24 @@
 #define LOGGER_H
 
 #include <QFile>
+#include <QMutex>
 #include <QString>
 #include <QtLogging>
 
-class Logger
-{
-    QFile m_file;
-
-    static void messageHandler(QtMsgType type, const QMessageLogContext &context, const QString &msg);
 
+class Logger {
 public:
+    explicit Logger();
+
     static Logger *globalInstance();
 
-    explicit Logger();
+private:
+    static void messageHandler(QtMsgType type, const QMessageLogContext &context, const QString &msg);
+
+private:
+    QFile  m_file;
+    QMutex m_mutex;
+
     friend class MyLogger;
 };
 

gpt4all-chat/src/macosdock.mm

@@ -2,6 +2,7 @@
 
 #include <Cocoa/Cocoa.h>
 
+
 void MacOSDock::showIcon()
 {
     [[NSApplication sharedApplication] setActivationPolicy:NSApplicationActivationPolicyRegular];

gpt4all-chat/src/main.cpp

@@ -12,18 +12,28 @@
 #include <gpt4all-backend/llmodel.h>
 #include <singleapplication.h>
 
+#include <QByteArray>
 #include <QCoreApplication>
 #include <QFont>
 #include <QFontDatabase>
+#include <QList>
 #include <QObject>
 #include <QQmlApplicationEngine>
 #include <QQmlContext>
 #include <QQuickWindow>
 #include <QSettings>
 #include <QString>
+#include <QStringList>
 #include <QUrl>
 #include <QVariant>
+#include <QWindow>
 #include <Qt>
+#include <QtAssert>
+#include <QtSystemDetection>
+
+#if G4A_CONFIG(force_d3d12)
+#   include <QSGRendererInterface>
+#endif
 
 #ifndef GPT4ALL_USE_QTPDF
 #   include <fpdfview.h>
@@ -83,24 +93,27 @@ int main(int argc, char *argv[])
         return 0;
     }
 
+#if G4A_CONFIG(force_d3d12)
+    QQuickWindow::setGraphicsApi(QSGRendererInterface::Direct3D12);
+#endif
+
 #ifdef Q_OS_LINUX
     app.setWindowIcon(QIcon(":/gpt4all/icons/gpt4all.svg"));
 #endif
 
     // set search path before constructing the MySettings instance, which relies on this
-    QString llmodelSearchPaths = QCoreApplication::applicationDirPath();
-    const QString libDir = QCoreApplication::applicationDirPath() + "/../lib/";
-    if (LLM::directoryExists(libDir))
-        llmodelSearchPaths += ";" + libDir;
-#if defined(Q_OS_MAC)
-    const QString binDir = QCoreApplication::applicationDirPath() + "/../../../";
-    if (LLM::directoryExists(binDir))
-        llmodelSearchPaths += ";" + binDir;
-    const QString frameworksDir = QCoreApplication::applicationDirPath() + "/../Frameworks/";
-    if (LLM::directoryExists(frameworksDir))
-        llmodelSearchPaths += ";" + frameworksDir;
+    {
+        auto appDirPath = QCoreApplication::applicationDirPath();
+        QStringList searchPaths {
+#ifdef Q_OS_DARWIN
+            u"%1/../Frameworks"_s.arg(appDirPath),
+#else
+            appDirPath,
+            u"%1/../lib"_s.arg(appDirPath),
 #endif
-    LLModel::Implementation::setImplementationsSearchPath(llmodelSearchPaths.toStdString());
+        };
+        LLModel::Implementation::setImplementationsSearchPath(searchPaths.join(u';').toStdString());
+    }
 
     // Set the local and language translation before the qml engine has even been started. This will
     // use the default system locale unless the user has explicitly set it to use a different one.
@@ -174,5 +187,9 @@ int main(int argc, char *argv[])
     // Otherwise, we can get a heap-use-after-free inside of llama.cpp.
     ChatListModel::globalInstance()->destroyChats();
 
+#ifndef GPT4ALL_USE_QTPDF
+    FPDF_DestroyLibrary();
+#endif
+
     return res;
 }

gpt4all-chat/src/modellist.cpp

@@ -9,9 +9,11 @@
 
 #include <QChar>
 #include <QCoreApplication>
+#include <QCryptographicHash>
 #include <QDebug>
 #include <QDir>
 #include <QDirIterator>
+#include <QEvent>
 #include <QEventLoop>
 #include <QFile>
 #include <QFileInfo>
@@ -29,14 +31,15 @@
 #include <QSslConfiguration>
 #include <QSslSocket>
 #include <QStandardPaths>
-#include <QStringList>
+#include <QStringList> // IWYU pragma: keep
 #include <QTextStream>
 #include <QTimer>
 #include <QUrl>
+#include <QtAssert>
 #include <QtLogging>
+#include <QtPreprocessorSupport>
 
 #include <algorithm>
-#include <cstddef>
 #include <iterator>
 #include <optional>
 #include <string>
@@ -499,10 +502,11 @@ bool GPT4AllDownloadableModels::filterAcceptsRow(int sourceRow,
     bool hasDescription = !description.isEmpty();
     bool isClone = sourceModel()->data(index, ModelList::IsCloneRole).toBool();
     bool isDiscovered = sourceModel()->data(index, ModelList::IsDiscoveredRole).toBool();
+    bool isOnline = sourceModel()->data(index, ModelList::OnlineRole).toBool();
     bool satisfiesKeyword = m_keywords.isEmpty();
     for (const QString &k : m_keywords)
         satisfiesKeyword = description.contains(k) ? true : satisfiesKeyword;
-    return !isDiscovered && hasDescription && !isClone && satisfiesKeyword;
+    return !isOnline && !isDiscovered && hasDescription && !isClone && satisfiesKeyword;
 }
 
 int GPT4AllDownloadableModels::count() const
@@ -1621,7 +1625,6 @@ void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
         QString requiresVersion = obj["requires"].toString();
         QString versionRemoved = obj["removedIn"].toString();
         QString url = obj["url"].toString();
-        QByteArray modelHash = obj["md5sum"].toString().toLatin1();
         bool isDefault = obj.contains("isDefault") && obj["isDefault"] == u"true"_s;
         bool disableGUI = obj.contains("disableGUI") && obj["disableGUI"] == u"true"_s;
         QString description = obj["description"].toString();
@@ -1632,6 +1635,16 @@ void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
         QString type = obj["type"].toString();
         bool isEmbeddingModel = obj["embeddingModel"].toBool();
 
+        QByteArray modelHash;
+        ModelInfo::HashAlgorithm hashAlgorithm;
+        if (auto it = obj.find("sha256sum"_L1); it != obj.end()) {
+            modelHash = it->toString().toLatin1();
+            hashAlgorithm = ModelInfo::Sha256;
+        } else {
+            modelHash = obj["md5sum"].toString().toLatin1();
+            hashAlgorithm = ModelInfo::Md5;
+        }
+
         // Some models aren't supported in the GUI at all
         if (disableGUI)
             continue;
@@ -1660,7 +1673,7 @@ void ModelList::parseModelsJsonFile(const QByteArray &jsonData, bool save)
             { ModelList::FilenameRole, modelFilename },
             { ModelList::FilesizeRole, modelFilesize },
             { ModelList::HashRole, modelHash },
-            { ModelList::HashAlgorithmRole, ModelInfo::Md5 },
+            { ModelList::HashAlgorithmRole, hashAlgorithm },
             { ModelList::DefaultRole, isDefault },
             { ModelList::DescriptionRole, description },
             { ModelList::RequiresVersionRole, requiresVersion },
@@ -2344,3 +2357,56 @@ void ModelList::handleDiscoveryItemErrorOccurred(QNetworkReply::NetworkError cod
     qWarning() << u"ERROR: Discovery item failed with error code \"%1-%2\""_s
                       .arg(code).arg(reply->errorString()).toStdString();
 }
+
+QStringList ModelList::remoteModelList(const QString &apiKey, const QUrl &baseUrl)
+{
+    QStringList modelList;
+
+    // Create the request
+    QNetworkRequest request;
+    request.setUrl(baseUrl.resolved(QUrl("models")));
+    request.setHeader(QNetworkRequest::ContentTypeHeader, "application/json");
+
+    // Add the Authorization header
+    const QString bearerToken = QString("Bearer %1").arg(apiKey);
+    request.setRawHeader("Authorization", bearerToken.toUtf8());
+
+    // Make the GET request
+    QNetworkReply *reply = m_networkManager.get(request);
+
+    // We use a local event loop to wait for the request to complete
+    QEventLoop loop;
+    connect(reply, &QNetworkReply::finished, &loop, &QEventLoop::quit);
+    loop.exec();
+
+    // Check for errors
+    if (reply->error() == QNetworkReply::NoError) {
+        // Parse the JSON response
+        const QByteArray responseData = reply->readAll();
+        const QJsonDocument jsonDoc = QJsonDocument::fromJson(responseData);
+
+        if (!jsonDoc.isNull() && jsonDoc.isObject()) {
+            QJsonObject rootObj = jsonDoc.object();
+            QJsonValue dataValue = rootObj.value("data");
+
+            if (dataValue.isArray()) {
+                QJsonArray dataArray = dataValue.toArray();
+                for (const QJsonValue &val : dataArray) {
+                    if (val.isObject()) {
+                        QJsonObject obj = val.toObject();
+                        const QString modelId = obj.value("id").toString();
+                        modelList.append(modelId);
+                    }
+                }
+            }
+        }
+    } else {
+        // Handle network error (e.g. print it to qDebug)
+        qWarning() << "Error retrieving models:" << reply->errorString();
+    }
+
+    // Clean up
+    reply->deleteLater();
+
+    return modelList;
+}

gpt4all-chat/src/modellist.h

@@ -5,25 +5,29 @@
 #include <QByteArray>
 #include <QDateTime>
 #include <QHash>
-#include <QLatin1StringView>
+#include <QLatin1StringView> // IWYU pragma: keep
 #include <QList>
 #include <QMutex>
 #include <QNetworkAccessManager>
 #include <QNetworkReply>
 #include <QObject>
-#include <QPair>
-#include <QQmlEngine>
+#include <QPair> // IWYU pragma: keep
+#include <QQmlEngine> // IWYU pragma: keep
 #include <QSortFilterProxyModel>
 #include <QSslError>
 #include <QString>
 #include <QVariant>
-#include <QVector>
+#include <QVector> // IWYU pragma: keep
 #include <Qt>
-#include <QtGlobal>
+#include <QtTypes>
 
 #include <optional>
 #include <utility>
 
+// IWYU pragma: no_forward_declare QObject
+// IWYU pragma: no_forward_declare QSslError
+class QUrl;
+
 using namespace Qt::Literals::StringLiterals;
 
 
@@ -269,7 +273,7 @@ private:
             std::optional<QString> m_chatTemplate;
     mutable std::optional<QString> m_modelChatTemplate;
     QString m_systemMessage;
-    QString m_chatNamePrompt          = "Describe the above conversation in seven words or less.";
+    QString m_chatNamePrompt          = "Describe the above conversation. Your entire response must be three words or less.";
     QString m_suggestedFollowUpPrompt = "Suggest three very short factual follow-up questions that have not been answered yet or cannot be found inspired by the previous conversation and excerpts.";
     friend class MySettings;
     friend class ModelList;
@@ -530,6 +534,8 @@ public:
 
     Q_INVOKABLE void discoverSearch(const QString &discover);
 
+    Q_INVOKABLE QStringList remoteModelList(const QString &apiKey, const QUrl &baseUrl);
+
 Q_SIGNALS:
     void countChanged();
     void installedModelsChanged();

gpt4all-chat/src/mysettings.cpp

@@ -11,22 +11,27 @@
 #include <QFileInfo>
 #include <QGlobalStatic>
 #include <QGuiApplication>
-#include <QIODevice>
+#include <QIODevice> // IWYU pragma: keep
 #include <QMap>
 #include <QMetaObject>
 #include <QStandardPaths>
 #include <QThread>
 #include <QUrl>
-#include <QVariant>
 #include <QtLogging>
+#include <QtAssert>
 
 #include <algorithm>
 #include <string>
 #include <thread>
 #include <vector>
 
+#if !(defined(Q_OS_MAC) && defined(__aarch64__))
+    #include <cstring>
+#endif
+
 using namespace Qt::Literals::StringLiterals;
 
+
 // used only for settings serialization, do not translate
 static const QStringList suggestionModeNames { "LocalDocsOnly", "On", "Off" };
 static const QStringList chatThemeNames      { "Light", "Dark", "LegacyDark" };

gpt4all-chat/src/mysettings.h

@@ -4,20 +4,24 @@
 #include "modellist.h" // IWYU pragma: keep
 
 #include <QDateTime>
-#include <QLatin1StringView>
+#include <QLatin1StringView> // IWYU pragma: keep
 #include <QList>
 #include <QModelIndex>
 #include <QObject>
 #include <QSettings>
 #include <QString>
-#include <QStringList>
+#include <QStringList> // IWYU pragma: keep
 #include <QTranslator>
-#include <QVector>
+#include <QVariant>
 
 #include <cstdint>
 #include <memory>
 #include <optional>
 
+// IWYU pragma: no_forward_declare QModelIndex
+class QLocale;
+
+
 namespace MySettingsEnums {
     Q_NAMESPACE
 

gpt4all-chat/src/network.cpp

@@ -8,7 +8,6 @@
 #include "localdocsmodel.h"
 #include "modellist.h"
 #include "mysettings.h"
-#include "utils.h"
 
 #include <gpt4all-backend/llmodel.h>
 
@@ -29,7 +28,6 @@
 #include <QSslSocket>
 #include <QSysInfo>
 #include <Qt>
-#include <QtGlobal>
 #include <QtLogging>
 #include <QUrl>
 #include <QUuid>
@@ -49,6 +47,7 @@ using namespace Qt::Literals::StringLiterals;
 #define STR_(x) #x
 #define STR(x) STR_(x)
 
+
 static const char MIXPANEL_TOKEN[] = "ce362e568ddaee16ed243eaffb5860a2";
 
 #ifdef __clang__
@@ -242,6 +241,12 @@ void Network::handleJsonUploadFinished()
 
     m_activeUploads.removeAll(jsonReply);
 
+    if (jsonReply->error() != QNetworkReply::NoError) {
+        qWarning() << "Request to" << jsonReply->url().toString() << "failed:" << jsonReply->errorString();
+        jsonReply->deleteLater();
+        return;
+    }
+
     QVariant response = jsonReply->attribute(QNetworkRequest::HttpStatusCodeAttribute);
     Q_ASSERT(response.isValid());
     bool ok;
@@ -449,6 +454,11 @@ void Network::handleIpifyFinished()
     QNetworkReply *reply = qobject_cast<QNetworkReply *>(sender());
     if (!reply)
         return;
+    if (reply->error() != QNetworkReply::NoError) {
+        qWarning() << "Request to" << reply->url().toString() << "failed:" << reply->errorString();
+        reply->deleteLater();
+        return;
+    }
 
     QVariant response = reply->attribute(QNetworkRequest::HttpStatusCodeAttribute);
     Q_ASSERT(response.isValid());
@@ -473,6 +483,11 @@ void Network::handleMixpanelFinished()
     QNetworkReply *reply = qobject_cast<QNetworkReply *>(sender());
     if (!reply)
         return;
+    if (reply->error() != QNetworkReply::NoError) {
+        qWarning() << "Request to" << reply->url().toString() << "failed:" << reply->errorString();
+        reply->deleteLater();
+        return;
+    }
 
     QVariant response = reply->attribute(QNetworkRequest::HttpStatusCodeAttribute);
     Q_ASSERT(response.isValid());
@@ -511,6 +526,11 @@ void Network::handleHealthFinished()
     QNetworkReply *healthReply = qobject_cast<QNetworkReply *>(sender());
     if (!healthReply)
         return;
+    if (healthReply->error() != QNetworkReply::NoError) {
+        qWarning() << "Request to" << healthReply->url().toString() << "failed:" << healthReply->errorString();
+        healthReply->deleteLater();
+        return;
+    }
 
     QVariant response = healthReply->attribute(QNetworkRequest::HttpStatusCodeAttribute);
     Q_ASSERT(response.isValid());

gpt4all-chat/src/network.h

@@ -11,7 +11,14 @@
 #include <QSslError>
 #include <QString>
 #include <QVariant>
-#include <QVector>
+#include <QVariantMap> // IWYU pragma: keep
+#include <QVector> // IWYU pragma: keep
+
+// IWYU pragma: no_forward_declare QByteArray
+// IWYU pragma: no_forward_declare QNetworkReply
+// IWYU pragma: no_forward_declare QSslError
+class QUrl;
+
 
 struct KeyValue {
     QString key;

gpt4all-chat/src/server.cpp

@@ -4,9 +4,10 @@
 #include "chatmodel.h"
 #include "modellist.h"
 #include "mysettings.h"
-#include "utils.h"
+#include "utils.h" // IWYU pragma: keep
 
 #include <fmt/format.h>
+#include <gpt4all-backend/llmodel.h>
 
 #include <QByteArray>
 #include <QCborArray>
@@ -15,32 +16,38 @@
 #include <QDateTime>
 #include <QDebug>
 #include <QHostAddress>
+#include <QHttpHeaders>
 #include <QHttpServer>
+#include <QHttpServerRequest>
 #include <QHttpServerResponder>
 #include <QJsonArray>
 #include <QJsonDocument>
 #include <QJsonObject>
 #include <QJsonValue>
 #include <QLatin1StringView>
-#include <QPair>
+#include <QPair> // IWYU pragma: keep
+#include <QTcpServer>
 #include <QVariant>
 #include <Qt>
+#include <QtAssert>
 #include <QtCborCommon>
-#include <QtGlobal>
 #include <QtLogging>
+#include <QtMinMax>
+#include <QtPreprocessorSupport>
+#include <QtTypes>
 
 #include <cstdint>
+#include <exception>
 #include <iostream>
 #include <optional>
+#include <span>
 #include <stdexcept>
 #include <string>
-#include <type_traits>
+#include <string_view>
 #include <unordered_map>
 #include <utility>
-
-#if QT_VERSION >= QT_VERSION_CHECK(6, 8, 0)
-#   include <QTcpServer>
-#endif
+#include <variant>
+#include <vector>
 
 using namespace std::string_literals;
 using namespace Qt::Literals::StringLiterals;
@@ -451,23 +458,17 @@ static QJsonObject requestFromJson(const QByteArray &request)
 void Server::start()
 {
     m_server = std::make_unique<QHttpServer>(this);
-#if QT_VERSION >= QT_VERSION_CHECK(6, 8, 0)
     auto *tcpServer = new QTcpServer(m_server.get());
-#else
-    auto *tcpServer = m_server.get();
-#endif
 
     auto port = MySettings::globalInstance()->networkPort();
     if (!tcpServer->listen(QHostAddress::LocalHost, port)) {
         qWarning() << "Server ERROR: Failed to listen on port" << port;
         return;
     }
-#if QT_VERSION >= QT_VERSION_CHECK(6, 8, 0)
     if (!m_server->bind(tcpServer)) {
         qWarning() << "Server ERROR: Failed to HTTP server to socket" << port;
         return;
     }
-#endif
 
     m_server->route("/v1/models", QHttpServerRequest::Method::Get,
         [](const QHttpServerRequest &) {
@@ -607,19 +608,12 @@ void Server::start()
         }
     );
 
-#if QT_VERSION >= QT_VERSION_CHECK(6, 8, 0)
     m_server->addAfterRequestHandler(this, [](const QHttpServerRequest &req, QHttpServerResponse &resp) {
         Q_UNUSED(req);
         auto headers = resp.headers();
         headers.append("Access-Control-Allow-Origin"_L1, "*"_L1);
         resp.setHeaders(std::move(headers));
     });
-#else
-    m_server->afterRequest([](QHttpServerResponse &&resp) {
-        resp.addHeader("Access-Control-Allow-Origin", "*");
-        return std::move(resp);
-    });
-#endif
 
     connect(this, &Server::requestResetResponseState, m_chat, &Chat::resetResponseState, Qt::BlockingQueuedConnection);
 }

gpt4all-chat/src/server.h

@@ -8,7 +8,7 @@
 #include <QHttpServerResponse>
 #include <QJsonObject>
 #include <QList>
-#include <QObject>
+#include <QObject> // IWYU pragma: keep
 #include <QString>
 
 #include <memory>

gpt4all-chat/src/tool.cpp

@@ -1,12 +1,16 @@
 #include "tool.h"
 
-#include <jinja2cpp/value.h>
+#include <QDataStream>
+#include <QtTypes>
 
 #include <string>
 
-jinja2::Value Tool::jinjaValue() const
+using json = nlohmann::ordered_json;
+
+
+json::object_t Tool::jinjaValue() const
 {
-    jinja2::ValuesList paramList;
+    json::array_t paramList;
     const QList<ToolParamInfo> p = parameters();
     for (auto &info : p) {
         std::string typeStr;
@@ -20,26 +24,24 @@ jinja2::Value Tool::jinjaValue() const
         case Boolean:  typeStr = "boolean"; break;
         case Null:     typeStr = "null"; break;
         }
-        jinja2::ValuesMap infoMap {
-            { "name", info.name.toStdString() },
-            { "type", typeStr},
+        paramList.emplace_back(json::initializer_list_t {
+            { "name",        info.name.toStdString()        },
+            { "type",        typeStr                        },
             { "description", info.description.toStdString() },
-            { "required", info.required }
-        };
-        paramList.push_back(infoMap);
+            { "required",    info.required                  },
+        });
     }
 
-    jinja2::ValuesMap params {
-        { "name", name().toStdString() },
-        { "description", description().toStdString() },
-        { "function", function().toStdString() },
-        { "parameters", paramList },
+    return {
+        { "name",           name().toStdString()           },
+        { "description",    description().toStdString()    },
+        { "function",       function().toStdString()       },
+        { "parameters",     paramList                      },
         { "symbolicFormat", symbolicFormat().toStdString() },
-        { "examplePrompt", examplePrompt().toStdString() },
-        { "exampleCall", exampleCall().toStdString() },
-        { "exampleReply", exampleReply().toStdString() }
+        { "examplePrompt",  examplePrompt().toStdString()  },
+        { "exampleCall",    exampleCall().toStdString()    },
+        { "exampleReply",   exampleReply().toStdString()   },
     };
-    return params;
 }
 
 void ToolCallInfo::serialize(QDataStream &stream, int version)

gpt4all-chat/src/tool.h

@@ -1,13 +1,18 @@
 #ifndef TOOL_H
 #define TOOL_H
 
+#include <nlohmann/json.hpp>
+
 #include <QList>
 #include <QObject>
 #include <QString>
 #include <QVariant>
 #include <QtGlobal>
 
-#include <jinja2cpp/value.h>
+class QDataStream;
+
+using json = nlohmann::ordered_json;
+
 
 namespace ToolEnums
 {
@@ -25,6 +30,7 @@ namespace ToolEnums
 
     enum class ParseState {
         None,
+        InTagChoice,
         InStart,
         Partial,
         Complete,
@@ -87,7 +93,8 @@ public:
     Tool() : QObject(nullptr) {}
     virtual ~Tool() {}
 
-    virtual QString run(const QList<ToolParam> &params, qint64 timeout = 2000) = 0;
+    virtual void run(const QList<ToolParam> &params) = 0;
+    virtual bool interrupt() = 0;
 
     // Tools should set these if they encounter errors. For instance, a tool depending upon the network
     // might set these error variables if the network is not available.
@@ -121,7 +128,10 @@ public:
 
     bool operator==(const Tool &other) const { return function() == other.function(); }
 
-    jinja2::Value jinjaValue() const;
+    json::object_t jinjaValue() const;
+
+Q_SIGNALS:
+    void runComplete(const ToolCallInfo &info);
 };
 
 #endif // TOOL_H

gpt4all-chat/src/toolcallparser.cpp

@@ -1,16 +1,31 @@
 #include "toolcallparser.h"
 
-#include <QDebug>
-#include <QtGlobal>
-#include <QtLogging>
+#include "tool.h"
 
-#include <cstddef>
+#include <QChar>
+#include <QSet>
+#include <QtAssert>
+#include <QtTypes>
+
+#include <stdexcept>
 
-static const QString ToolCallStart = ToolCallConstants::CodeInterpreterTag;
-static const QString ToolCallEnd = ToolCallConstants::CodeInterpreterEndTag;
 
 ToolCallParser::ToolCallParser()
+    : ToolCallParser(ToolCallConstants::AllTagNames)
+    {}
+
+ToolCallParser::ToolCallParser(const QStringList &tagNames)
 {
+    QSet<QChar> firstChars;
+    for (auto &name : tagNames) {
+        if (name.isEmpty())
+            throw std::invalid_argument("ToolCallParser(): tag names must not be empty");
+        if (firstChars.contains(name.at(0)))
+            throw std::invalid_argument("ToolCallParser(): tag names must not share any prefix");
+        firstChars << name.at(0);
+        m_possibleStartTags << makeStartTag(name).toUtf8();
+        m_possibleEndTags   << makeEndTag  (name).toUtf8();
+    }
     reset();
 }
 
@@ -20,36 +35,68 @@ void ToolCallParser::reset()
     resetSearchState();
 
     // These are global states maintained between update calls
-    m_buffer.clear();
-    m_hasSplit = false;
+    m_buffers.clear();
+    m_buffers << QByteArray();
 }
 
 void ToolCallParser::resetSearchState()
 {
-    m_expected = ToolCallStart.at(0);
+    m_expected = {'<'};
     m_expectedIndex = 0;
     m_state = ToolEnums::ParseState::None;
+
     m_toolCall.clear();
+    m_startTagBuffer.clear();
     m_endTagBuffer.clear();
+
+    m_currentTagIndex = -1;
     m_startIndex = -1;
+    m_endIndex = -1;
+}
+
+bool ToolCallParser::isExpected(char c) const
+{
+    return m_expected.isEmpty() || m_expected.contains(c);
+}
+
+void ToolCallParser::setExpected(const QList<QByteArray> &tags)
+{
+    m_expected.clear();
+    for (const auto &tag : tags) {
+        Q_ASSERT(tag.size() > m_expectedIndex);
+        m_expected << tag.at(m_expectedIndex);
+    }
+}
+
+QByteArray ToolCallParser::startTag() const
+{
+    if (m_currentTagIndex < 0)
+        return {};
+    return m_possibleStartTags.at(m_currentTagIndex);
+}
+
+QByteArray ToolCallParser::endTag() const
+{
+    if (m_currentTagIndex < 0)
+        return {};
+    return m_possibleEndTags.at(m_currentTagIndex);
+}
+
+QByteArray &ToolCallParser::currentBuffer()
+{
+    return m_buffers.last();
 }
 
 // This method is called with an arbitrary string and a current state. This method should take the
 // current state into account and then parse through the update character by character to arrive at
 // the new state.
-void ToolCallParser::update(const QString &update)
+void ToolCallParser::update(const QByteArray &update)
 {
-    Q_ASSERT(m_state != ToolEnums::ParseState::Complete);
-    if (m_state == ToolEnums::ParseState::Complete) {
-        qWarning() << "ERROR: ToolCallParser::update already found a complete toolcall!";
-        return;
-    }
+    currentBuffer().append(update);
 
-    m_buffer.append(update);
-
-    for (size_t i = m_buffer.size() - update.size(); i < m_buffer.size(); ++i) {
-        const QChar c = m_buffer[i];
-        const bool foundMatch = m_expected.isNull() || c == m_expected;
+    for (qsizetype i = currentBuffer().size() - update.size(); i < currentBuffer().size(); ++i) {
+        const char c = currentBuffer()[i];
+        const bool foundMatch = isExpected(c);
         if (!foundMatch) {
             resetSearchState();
             continue;
@@ -59,34 +106,58 @@ void ToolCallParser::update(const QString &update)
         case ToolEnums::ParseState::None:
             {
                 m_expectedIndex = 1;
-                m_expected = ToolCallStart.at(1);
-                m_state = ToolEnums::ParseState::InStart;
+                setExpected(m_possibleStartTags);
+                m_state = ToolEnums::ParseState::InTagChoice;
                 m_startIndex = i;
                 break;
             }
+        case ToolEnums::ParseState::InTagChoice:
+            {
+                for (int i = 0; i < m_possibleStartTags.size(); ++i) {
+                    const auto &tag = m_possibleStartTags.at(i);
+                    if (c == tag.at(1)) m_currentTagIndex = i;
+                }
+                if (m_currentTagIndex >= 0) {
+                    m_expectedIndex = 2;
+                    setExpected({m_possibleStartTags.at(m_currentTagIndex)});
+                    m_state = ToolEnums::ParseState::InStart;
+                } else
+                    resetSearchState();
+                break;
+            }
         case ToolEnums::ParseState::InStart:
             {
-                if (m_expectedIndex == ToolCallStart.size() - 1) {
+                m_startTagBuffer.append(c);
+
+                const auto startTag = this->startTag();
+                Q_ASSERT(!startTag.isEmpty());
+                if (m_expectedIndex == startTag.size() - 1) {
                     m_expectedIndex = 0;
-                    m_expected = QChar();
+                    setExpected({});
                     m_state = ToolEnums::ParseState::Partial;
                 } else {
                     ++m_expectedIndex;
-                    m_expected = ToolCallStart.at(m_expectedIndex);
+                    Q_ASSERT(m_currentTagIndex >= 0);
+                    setExpected({startTag});
                 }
                 break;
             }
         case ToolEnums::ParseState::Partial:
             {
+                Q_ASSERT(m_currentTagIndex >= 0);
+                const auto endTag = this->endTag();
+                Q_ASSERT(!endTag.isEmpty());
                 m_toolCall.append(c);
                 m_endTagBuffer.append(c);
-                if (m_endTagBuffer.size() > ToolCallEnd.size())
+                if (m_endTagBuffer.size() > endTag.size())
                     m_endTagBuffer.remove(0, 1);
-                if (m_endTagBuffer == ToolCallEnd) {
-                    m_toolCall.chop(ToolCallEnd.size());
+                if (m_endTagBuffer == endTag) {
+                    m_endIndex = i + 1;
+                    m_toolCall.chop(endTag.size());
                     m_state = ToolEnums::ParseState::Complete;
                     m_endTagBuffer.clear();
                 }
+                break;
             }
         case ToolEnums::ParseState::Complete:
             {
@@ -97,15 +168,35 @@ void ToolCallParser::update(const QString &update)
     }
 }
 
-QPair<QString, QString> ToolCallParser::split()
+bool ToolCallParser::splitIfPossible()
 {
-    Q_ASSERT(m_state == ToolEnums::ParseState::Partial
-        || m_state == ToolEnums::ParseState::Complete);
+    // The first split happens when we're in a partial state
+    if (m_buffers.size() < 2 && m_state == ToolEnums::ParseState::Partial) {
+        Q_ASSERT(m_startIndex >= 0);
+        const auto beforeToolCall = currentBuffer().left(m_startIndex);
+        const auto toolCall       = currentBuffer().mid (m_startIndex);
+        m_buffers = { beforeToolCall, toolCall };
+        return true;
+    }
 
-    Q_ASSERT(m_startIndex >= 0);
-    m_hasSplit = true;
-    const QString beforeToolCall = m_buffer.left(m_startIndex);
-    m_buffer = m_buffer.mid(m_startIndex);
-    m_startIndex = 0;
-    return { beforeToolCall, m_buffer };
+    // The second split happens when we're in the complete state
+    if (m_buffers.size() < 3 && m_state == ToolEnums::ParseState::Complete) {
+        Q_ASSERT(m_endIndex >= 0);
+        const auto &beforeToolCall = m_buffers.first();
+        const auto toolCall        = currentBuffer().left(m_endIndex);
+        const auto afterToolCall   = currentBuffer().mid (m_endIndex);
+        m_buffers = { beforeToolCall, toolCall, afterToolCall };
+        return true;
+    }
+
+    return false;
+}
+
+QStringList ToolCallParser::buffers() const
+{
+    QStringList result;
+    result.reserve(m_buffers.size());
+    for (const auto &buffer : m_buffers)
+        result << QString::fromUtf8(buffer);
+    return result;
 }

gpt4all-chat/src/toolcallparser.h

@@ -1,47 +1,73 @@
 #ifndef TOOLCALLPARSER_H
 #define TOOLCALLPARSER_H
 
-#include "tool.h"
-
-#include <QChar>
+#include <QByteArray>
+#include <QList>
 #include <QString>
-#include <QPair>
+#include <QStringList> // IWYU pragma: keep
+
+namespace ToolEnums { enum class ParseState; }
+
+using namespace Qt::Literals::StringLiterals;
 
-namespace ToolCallConstants
-{
-    const QString CodeInterpreterFunction = R"(javascript_interpret)";
-    const QString CodeInterpreterTag = R"(<)" + CodeInterpreterFunction + R"(>)";
-    const QString CodeInterpreterEndTag = R"(</)" + CodeInterpreterFunction + R"(>)";
-    const QString CodeInterpreterPrefix = CodeInterpreterTag + "\n```javascript\n";
-    const QString CodeInterpreterSuffix = "```\n" + CodeInterpreterEndTag;
-}
 
 class ToolCallParser
 {
 public:
     ToolCallParser();
+    ToolCallParser(const QStringList &tagNames);
+
     void reset();
-    void update(const QString &update);
-    QString buffer() const { return m_buffer; }
-    QString toolCall() const { return m_toolCall; }
+    void update(const QByteArray &update);
+    QString toolCall() const { return QString::fromUtf8(m_toolCall); }
     int startIndex() const { return m_startIndex; }
     ToolEnums::ParseState state() const { return m_state; }
+    QByteArray startTag() const;
+    QByteArray endTag() const;
 
-    // Splits
-    QPair<QString, QString> split();
-    bool hasSplit() const { return m_hasSplit; }
+    bool splitIfPossible();
+    QStringList buffers() const;
+    int numberOfBuffers() const { return m_buffers.size(); }
+
+    static QString makeStartTag(const QString &name) { return u"<%1>"_s .arg(name); }
+    static QString makeEndTag  (const QString &name) { return u"</%1>"_s.arg(name); }
 
 private:
+    QByteArray &currentBuffer();
     void resetSearchState();
+    bool isExpected(char c) const;
+    void setExpected(const QList<QByteArray> &tags);
 
-    QChar m_expected;
+    QList<QByteArray> m_possibleStartTags;
+    QList<QByteArray> m_possibleEndTags;
+    QByteArray m_startTagBuffer;
+    QByteArray m_endTagBuffer;
+    int m_currentTagIndex;
+
+    QList<char> m_expected;
     int m_expectedIndex;
     ToolEnums::ParseState m_state;
-    QString m_buffer;
-    QString m_toolCall;
-    QString m_endTagBuffer;
+    QList<QByteArray> m_buffers;
+    QByteArray m_toolCall;
     int m_startIndex;
-    bool m_hasSplit;
+    int m_endIndex;
 };
 
+namespace ToolCallConstants
+{
+    // NB: the parsing code assumes the first char of the various tags differ
+
+    inline const QString CodeInterpreterFunction = u"javascript_interpret"_s;
+    inline const QString CodeInterpreterStartTag = ToolCallParser::makeStartTag(CodeInterpreterFunction);
+    inline const QString CodeInterpreterEndTag   = ToolCallParser::makeEndTag  (CodeInterpreterFunction);
+    inline const QString CodeInterpreterPrefix   = u"%1\n```javascript\n"_s.arg(CodeInterpreterStartTag);
+    inline const QString CodeInterpreterSuffix   = u"```\n%1"_s            .arg(CodeInterpreterEndTag  );
+
+    inline const QString ThinkTagName  = u"think"_s;
+    inline const QString ThinkStartTag = ToolCallParser::makeStartTag(ThinkTagName);
+    inline const QString ThinkEndTag   = ToolCallParser::makeEndTag  (ThinkTagName);
+
+    inline const QStringList AllTagNames { CodeInterpreterFunction, ThinkTagName };
+}
+
 #endif // TOOLCALLPARSER_H

gpt4all-chat/src/toolmodel.cpp

@@ -6,6 +6,7 @@
 #include <QEvent>
 #include <QGlobalStatic>
 
+
 class MyToolModel: public ToolModel { };
 Q_GLOBAL_STATIC(MyToolModel, toolModelInstance)
 ToolModel *ToolModel::globalInstance()

gpt4all-chat/src/toolmodel.h

@@ -9,7 +9,8 @@
 #include <QList>
 #include <QString>
 #include <QVariant>
-#include <QtGlobal>
+#include <QtPreprocessorSupport>
+
 
 class ToolModel : public QAbstractListModel
 {

gpt4all-chat/src/utils.h

@@ -5,7 +5,7 @@
 
 #include <QByteArray>
 #include <QJsonValue>
-#include <QLatin1StringView>
+#include <QLatin1StringView> // IWYU pragma: keep
 #include <QString>
 #include <QStringView>
 #include <QUtf8StringView>
@@ -13,8 +13,9 @@
 
 #include <initializer_list>
 #include <string_view>
-#include <utility>
+#include <utility> // IWYU pragma: keep
 
+// IWYU pragma: no_forward_declare QJsonValue
 class QJsonObject;
 
 
@@ -40,4 +41,4 @@ MAKE_FORMATTER(QVariant,        value.toString().toUtf8());
 // alternative to QJsonObject's initializer_list constructor that accepts Latin-1 strings
 QJsonObject makeJsonObject(std::initializer_list<std::pair<QLatin1StringView, QJsonValue>> args);
 
-#include "utils.inl"
+#include "utils.inl" // IWYU pragma: export

gpt4all-chat/src/utils.inl

@@ -1,5 +1,6 @@
 #include <QJsonObject>
 
+
 inline QJsonObject makeJsonObject(std::initializer_list<std::pair<QLatin1StringView, QJsonValue>> args)
 {
     QJsonObject obj;

gpt4all-chat/src/xlsxtomd.cpp

@@ -7,15 +7,16 @@
 #include <xlsxformat.h>
 #include <xlsxworksheet.h>
 
+#include <QChar>
 #include <QDateTime>
 #include <QDebug>
+#include <QLatin1StringView>
 #include <QList>
 #include <QRegularExpression>
 #include <QString>
-#include <QStringList>
+#include <QStringList> // IWYU pragma: keep
 #include <QStringView>
 #include <QVariant>
-#include <QtGlobal>
 #include <QtLogging>
 
 #include <memory>

gpt4all-chat/src/xlsxtomd.h

@@ -4,6 +4,7 @@
 class QIODevice;
 class QString;
 
+
 class XLSXToMD
 {
 public: