chat: fix emoji corruption (#3443)

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

gpt4all-chat/CHANGELOG.md

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

gpt4all-chat/src/chat.cpp

@@ -254,7 +254,7 @@ void Chat::responseStopped(qint64 promptResponseMs)
     });
 
     ToolCallParser parser;
-    parser.update(possibleToolcall);
+    parser.update(possibleToolcall.toUtf8());
     if (parser.state() == ToolEnums::ParseState::Complete && parser.startTag() != ToolCallConstants::ThinkTag)
         processToolCall(parser.toolCall());
     else

gpt4all-chat/src/chatllm.cpp

@@ -938,13 +938,11 @@ auto ChatLLM::promptInternal(
         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()));
+        toolCallParser.update(piece.data());
 
         // Split the response into two if needed and create chat items
         if (toolCallParser.numberOfBuffers() < 2 && toolCallParser.splitIfPossible()) {
-            const QVector<QString> &parseBuffers = toolCallParser.buffers();
+            const auto parseBuffers = toolCallParser.buffers();
             Q_ASSERT(parseBuffers.size() == 2);
             if (toolCallParser.startTag() == ToolCallConstants::ThinkTag)
                 m_chatModel->splitThinking({parseBuffers.at(0), parseBuffers.at(1)});
@@ -955,7 +953,7 @@ auto ChatLLM::promptInternal(
         // Split the response into three if needed and create chat items
         if (toolCallParser.numberOfBuffers() < 3 && toolCallParser.startTag() == ToolCallConstants::ThinkTag
             && toolCallParser.splitIfPossible()) {
-            const QVector<QString> &parseBuffers = toolCallParser.buffers();
+            const auto parseBuffers = toolCallParser.buffers();
             Q_ASSERT(parseBuffers.size() == 3);
             m_chatModel->endThinking({parseBuffers.at(1), parseBuffers.at(2)}, totalTime.elapsed());
         }
@@ -964,7 +962,7 @@ auto ChatLLM::promptInternal(
         auto respStr = QString::fromUtf8(result.response);
 
         try {
-            const QVector<QString> &parseBuffers = toolCallParser.buffers();
+            const auto parseBuffers = toolCallParser.buffers();
             if (parseBuffers.size() > 1)
                 m_chatModel->setResponseValue(parseBuffers.last());
             else
@@ -998,7 +996,7 @@ auto ChatLLM::promptInternal(
     m_timer->stop();
     qint64 elapsed = totalTime.elapsed();
 
-    const QVector<QString> &parseBuffers = toolCallParser.buffers();
+    const auto parseBuffers = toolCallParser.buffers();
     const bool shouldExecuteToolCall = toolCallParser.state() == ToolEnums::ParseState::Complete
         && toolCallParser.startTag() != ToolCallConstants::ThinkTag;
 

gpt4all-chat/src/chatmodel.h

@@ -265,7 +265,7 @@ 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)
@@ -292,7 +292,7 @@ public:
     QString thinkContent(const QString &value) const
     {
         ToolCallParser parser;
-        parser.update(value);
+        parser.update(value.toUtf8());
 
         // Extract the content
         QString content = parser.toolCall();
@@ -303,7 +303,7 @@ public:
     QString toolCallContent(const QString &value) const
     {
         ToolCallParser parser;
-        parser.update(value);
+        parser.update(value.toUtf8());
 
         // Extract the code
         QString code = parser.toolCall();

gpt4all-chat/src/toolcallparser.cpp

@@ -8,10 +8,10 @@
 
 ToolCallParser::ToolCallParser()
 {
-    m_possibleStartTags << ToolCallConstants::CodeInterpreterTag
-                        << ToolCallConstants::ThinkTag;
-    m_possibleEndTags   << ToolCallConstants::CodeInterpreterEndTag
-                        << ToolCallConstants::ThinkEndTag;
+    m_possibleStartTags << ToolCallConstants::CodeInterpreterTag.toUtf8()
+                        << ToolCallConstants::ThinkTag.toUtf8();
+    m_possibleEndTags   << ToolCallConstants::CodeInterpreterEndTag.toUtf8()
+                        << ToolCallConstants::ThinkEndTag.toUtf8();
     reset();
 }
 
@@ -22,7 +22,7 @@ void ToolCallParser::reset()
 
     // These are global states maintained between update calls
     m_buffers.clear();
-    m_buffers.append(QString());
+    m_buffers << QByteArray();
 }
 
 void ToolCallParser::resetSearchState()
@@ -40,35 +40,35 @@ void ToolCallParser::resetSearchState()
     m_endIndex = -1;
 }
 
-bool ToolCallParser::isExpected(QChar c) const
+bool ToolCallParser::isExpected(char c) const
 {
     return m_expected.isEmpty() || m_expected.contains(c);
 }
 
-void ToolCallParser::setExpected(const QStringList &tags)
+void ToolCallParser::setExpected(const QList<QByteArray> &tags)
 {
     m_expected.clear();
-    for (const QString &tag : tags) {
+    for (const auto &tag : tags) {
         Q_ASSERT(tag.size() > m_expectedIndex);
         m_expected << tag.at(m_expectedIndex);
     }
 }
 
-QString ToolCallParser::startTag() const
+QByteArray ToolCallParser::startTag() const
 {
     if (m_currentTagIndex < 0)
-        return QString();
+        return {};
     return m_possibleStartTags.at(m_currentTagIndex);
 }
 
-QString ToolCallParser::endTag() const
+QByteArray ToolCallParser::endTag() const
 {
     if (m_currentTagIndex < 0)
-        return QString();
+        return {};
     return m_possibleEndTags.at(m_currentTagIndex);
 }
 
-QString &ToolCallParser::currentBuffer()
+QByteArray &ToolCallParser::currentBuffer()
 {
     return m_buffers.last();
 }
@@ -76,12 +76,12 @@ QString &ToolCallParser::currentBuffer()
 // 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)
 {
     currentBuffer().append(update);
 
     for (size_t i = currentBuffer().size() - update.size(); i < currentBuffer().size(); ++i) {
-        const QChar c = currentBuffer()[i];
+        const char c = currentBuffer()[i];
         const bool foundMatch = isExpected(c);
         if (!foundMatch) {
             resetSearchState();
@@ -100,7 +100,7 @@ void ToolCallParser::update(const QString &update)
         case ToolEnums::ParseState::InTagChoice:
             {
                 for (int i = 0; i < m_possibleStartTags.size(); ++i) {
-                    const QString tag = m_possibleStartTags.at(i);
+                    const auto &tag = m_possibleStartTags.at(i);
                     if (c == tag.at(1)) m_currentTagIndex = i;
                 }
                 if (m_currentTagIndex >= 0) {
@@ -115,7 +115,7 @@ void ToolCallParser::update(const QString &update)
             {
                 m_startTagBuffer.append(c);
 
-                const QString startTag = this->startTag();
+                const auto startTag = this->startTag();
                 Q_ASSERT(!startTag.isEmpty());
                 if (m_expectedIndex == startTag.size() - 1) {
                     m_expectedIndex = 0;
@@ -131,7 +131,7 @@ void ToolCallParser::update(const QString &update)
         case ToolEnums::ParseState::Partial:
             {
                 Q_ASSERT(m_currentTagIndex >= 0);
-                const QString endTag = this->endTag();
+                const auto endTag = this->endTag();
                 Q_ASSERT(!endTag.isEmpty());
                 m_toolCall.append(c);
                 m_endTagBuffer.append(c);
@@ -159,8 +159,8 @@ bool ToolCallParser::splitIfPossible()
     // 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 QString beforeToolCall = currentBuffer().left(m_startIndex);
-        const QString toolCall = currentBuffer().mid(m_startIndex);
+        const auto beforeToolCall = currentBuffer().left(m_startIndex);
+        const auto toolCall       = currentBuffer().mid (m_startIndex);
         m_buffers = { beforeToolCall, toolCall };
         return true;
     }
@@ -168,9 +168,9 @@ bool ToolCallParser::splitIfPossible()
     // 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 QString beforeToolCall = m_buffers.first();
-        const QString toolCall = currentBuffer().left(m_endIndex);
-        const QString afterToolCall = currentBuffer().mid(m_endIndex);
+        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;
     }
@@ -178,7 +178,11 @@ bool ToolCallParser::splitIfPossible()
     return false;
 }
 
-const QVector<QString> &ToolCallParser::buffers() const
+QStringList ToolCallParser::buffers() const
 {
-    return m_buffers;
+    QStringList result;
+    result.reserve(m_buffers.size());
+    for (const auto &buffer : m_buffers)
+        result << QString::fromUtf8(buffer);
+    return result;
 }

gpt4all-chat/src/toolcallparser.h

@@ -3,9 +3,10 @@
 
 #include "tool.h"
 
-#include <QChar>
+#include <QByteArray>
+#include <QList>
 #include <QString>
-#include <QPair>
+#include <QStringList>
 
 namespace ToolCallConstants
 {
@@ -25,34 +26,34 @@ class ToolCallParser
 public:
     ToolCallParser();
     void reset();
-    void update(const QString &update);
-    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; }
-    QString startTag() const;
-    QString endTag() const;
+    QByteArray startTag() const;
+    QByteArray endTag() const;
 
     bool splitIfPossible();
-    const QVector<QString> &buffers() const;
+    QStringList buffers() const;
     int numberOfBuffers() const { return m_buffers.size(); }
 
 private:
-    QString &currentBuffer();
+    QByteArray &currentBuffer();
     void resetSearchState();
-    bool isExpected(QChar c) const;
-    void setExpected(const QStringList &tags);
+    bool isExpected(char c) const;
+    void setExpected(const QList<QByteArray> &tags);
 
-    QStringList m_possibleStartTags;
-    QStringList m_possibleEndTags;
-    QString m_startTagBuffer;
-    QString m_endTagBuffer;
+    QList<QByteArray> m_possibleStartTags;
+    QList<QByteArray> m_possibleEndTags;
+    QByteArray m_startTagBuffer;
+    QByteArray m_endTagBuffer;
     int m_currentTagIndex;
 
-    QVector<QChar> m_expected;
+    QList<char> m_expected;
     int m_expectedIndex;
     ToolEnums::ParseState m_state;
-    QVector<QString> m_buffers;
-    QString m_toolCall;
+    QList<QByteArray> m_buffers;
+    QByteArray m_toolCall;
     int m_startIndex;
     int m_endIndex;
 };