Don't block the gui thread for tool calls (#3435)

Signed-off-by: Adam Treat <treat.adam@gmail.com>

gpt4all-chat/CHANGELOG.md

@@ -15,6 +15,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - 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))
 
 ## [3.7.0] - 2025-01-21
 

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
                         }
                     }

gpt4all-chat/src/chat.cpp

@@ -181,6 +181,11 @@ 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 +247,71 @@ 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);
+    if (parser.state() == ToolEnums::ParseState::Complete)
+        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;
 }
 

gpt4all-chat/src/chat.h

@@ -55,7 +55,8 @@ public:
         LocalDocsProcessing,
         PromptProcessing,
         GeneratingQuestions,
-        ResponseGeneration
+        ResponseGeneration,
+        ToolCallGeneration
     };
     Q_ENUM(ResponseState)
 
@@ -166,6 +167,9 @@ 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/codeinterpreter.cpp

@@ -7,8 +7,15 @@
 
 using namespace Qt::Literals::StringLiterals;
 
+CodeInterpreter::CodeInterpreter()
+    : Tool()
+    , m_error(ToolEnums::Error::NoError)
+{
+    m_worker = new CodeInterpreterWorker;
+    connect(this, &CodeInterpreter::request, m_worker, &CodeInterpreterWorker::request, Qt::QueuedConnection);
+}
 
-QString CodeInterpreter::run(const QList<ToolParam> &params, qint64 timeout)
+void CodeInterpreter::run(const QList<ToolParam> &params)
 {
     m_error = ToolEnums::Error::NoError;
     m_errorString = QString();
@@ -18,27 +25,24 @@ 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
@@ -89,17 +93,15 @@ 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 == 0)
@@ -116,15 +118,28 @@ 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.
@@ -145,9 +160,16 @@ 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

@@ -8,6 +8,7 @@
 #include <QObject>
 #include <QString>
 #include <QtGlobal>
+#include <QThread>
 
 class JavaScriptConsoleCapture : public QObject
 {
@@ -39,32 +40,37 @@ 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 +83,13 @@ 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/tool.h

@@ -87,7 +87,8 @@ 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.
@@ -122,6 +123,9 @@ public:
     bool operator==(const Tool &other) const { return function() == other.function(); }
 
     jinja2::Value jinjaValue() const;
+
+Q_SIGNALS:
+    void runComplete(const ToolCallInfo &info);
 };
 
 #endif // TOOL_H