Version 0.2.2

- Improve chat history
- Fix system prompt
- Fix ollama model livetime
- Fix providers and prompt default value

DocumentContextReader.cpp

@@ -250,9 +250,6 @@ QString DocumentContextReader::getInstructions() const
 {
     QString instructions;
 
-    if (Settings::contextSettings().useSpecificInstructions())
-        instructions += getSpecificInstructions();
-
     if (Settings::contextSettings().useFilePathInContext())
         instructions += getLanguageAndFileInfo();
 

LLMClientInterface.cpp

@@ -30,6 +30,7 @@
 #include "PromptTemplateManager.hpp"
 #include "QodeAssistUtils.hpp"
 #include "core/LLMRequestConfig.hpp"
+#include "settings/ContextSettings.hpp"
 #include "settings/GeneralSettings.hpp"
 
 namespace QodeAssist {
@@ -159,6 +160,9 @@ void LLMClientInterface::handleCompletion(const QJsonObject &request)
                               {"stop",
                                QJsonArray::fromStringList(config.promptTemplate->stopWords())}};
 
+    if (Settings::contextSettings().useSpecificInstructions())
+        config.providerRequest["system"] = Settings::contextSettings().specificInstractions();
+
     config.promptTemplate->prepareRequest(config.providerRequest, updatedContext);
     config.provider->prepareRequest(config.providerRequest);
 

QodeAssist.json.in

@@ -1,6 +1,6 @@
 {
     "Name" : "QodeAssist",
-    "Version" : "0.2.1",
+    "Version" : "0.2.2",
     "CompatVersion" : "${IDE_VERSION_COMPAT}",
     "Vendor" : "Petr Mironychev",
     "Copyright" : "(C) ${IDE_COPYRIGHT_YEAR} Petr Mironychev, (C) ${IDE_COPYRIGHT_YEAR} The Qt Company Ltd",

QodeAssistConstants.hpp

@@ -53,10 +53,10 @@ const char AUTO_COMPLETION_CHAR_THRESHOLD[] = "QodeAssist.autoCompletionCharThre
 const char AUTO_COMPLETION_TYPING_INTERVAL[] = "QodeAssist.autoCompletionTypingInterval";
 const char MAX_FILE_THRESHOLD[] = "QodeAssist.maxFileThreshold";
 const char OLLAMA_LIVETIME[] = "QodeAssist.ollamaLivetime";
-const char SPECIFIC_INSTRUCTIONS[] = "QodeAssist.specificInstractions";
+const char SYSTEM_PROMPT[] = "QodeAssist.systemPrompt";
 const char MULTILINE_COMPLETION[] = "QodeAssist.multilineCompletion";
 const char API_KEY[] = "QodeAssist.apiKey";
-const char USE_SPECIFIC_INSTRUCTIONS[] = "QodeAssist.useSpecificInstructions";
+const char USE_SYSTEM_PROMPT[] = "QodeAssist.useSystemPrompt";
 const char USE_FILE_PATH_IN_CONTEXT[] = "QodeAssist.useFilePathInContext";
 const char CUSTOM_JSON_TEMPLATE[] = "QodeAssist.customJsonTemplate";
 const char USE_PROJECT_CHANGES_CACHE[] = "QodeAssist.useProjectChangesCache";

README.md

@@ -65,9 +65,15 @@ If you've successfully used a model that's not listed here, please let us know b
 
 1. Install QtCreator 14.0
 2. Install [Ollama](https://ollama.com). Make sure to review the system requirements before installation.
-3. Install a language model in Ollama. For example, you can run:
+3. Install a language models in Ollama. For example, you can run:
+
+For suggestions:
 ```
-ollama run starcoder2:7b
+ollama run codellama:7b-code
+```
+For chat:
+```
+ollama run codellama:7b-instruct
 ```
 4. Download the QodeAssist plugin.
 5. Launch Qt Creator and install the plugin:

chat/ChatClientInterface.cpp

@@ -31,6 +31,51 @@
 
 namespace QodeAssist::Chat {
 
+int ChatHistory::estimateTokenCount(const QString &text) const
+{
+    return text.length() / 4;
+}
+
+void ChatHistory::addMessage(ChatMessage::Role role, const QString &content)
+{
+    int tokenCount = estimateTokenCount(content);
+    m_messages.append({role, content, tokenCount});
+    m_totalTokens += tokenCount;
+    trim();
+}
+void ChatHistory::clear()
+{
+    m_messages.clear();
+    m_totalTokens = 0;
+}
+
+QVector<ChatMessage> ChatHistory::getMessages() const
+{
+    return m_messages;
+}
+
+QString ChatHistory::getSystemPrompt() const
+{
+    return m_systemPrompt;
+}
+
+void ChatHistory::setSystemPrompt(const QString &prompt)
+{
+    m_systemPrompt = prompt;
+}
+
+void ChatHistory::trim()
+{
+    while (m_messages.size() > MAX_HISTORY_SIZE || m_totalTokens > MAX_TOKENS) {
+        if (!m_messages.isEmpty()) {
+            m_totalTokens -= m_messages.first().tokenCount;
+            m_messages.removeFirst();
+        } else {
+            break;
+        }
+    }
+}
+
 ChatClientInterface::ChatClientInterface(QObject *parent)
     : QObject(parent)
     , m_requestHandler(new LLMRequestHandler(this))
@@ -51,15 +96,10 @@ ChatClientInterface::ChatClientInterface(QObject *parent)
                 }
             });
 
-    // QJsonObject systemMessage;
+    m_chatHistory.setSystemPrompt("You are a helpful C++ and QML programming assistant.");
-    // systemMessage["role"] = "system";
-    // systemMessage["content"] = "You are a helpful C++ and QML programming assistant.";
-    // m_chatHistory.append(systemMessage);
 }
 
-ChatClientInterface::~ChatClientInterface()
+ChatClientInterface::~ChatClientInterface() = default;
-{
-}
 
 void ChatClientInterface::sendMessage(const QString &message)
 {
@@ -79,14 +119,11 @@ void ChatClientInterface::sendMessage(const QString &message)
     QJsonObject providerRequest;
     providerRequest["model"] = Settings::generalSettings().chatModelName();
     providerRequest["stream"] = true;
-
+    providerRequest["messages"] = prepareMessagesForRequest();
-    providerRequest["messages"] = m_chatHistory;
 
     chatTemplate->prepareRequest(providerRequest, context);
     chatProvider->prepareRequest(providerRequest);
 
-    m_chatHistory = providerRequest["messages"].toArray();
-
     LLMConfig config;
     config.requestType = RequestType::Chat;
     config.provider = chatProvider;
@@ -99,18 +136,22 @@ void ChatClientInterface::sendMessage(const QString &message)
     request["id"] = QUuid::createUuid().toString();
 
     m_accumulatedResponse.clear();
-    m_pendingMessage = message;
+    m_chatHistory.addMessage(ChatMessage::Role::User, message);
     m_requestHandler->sendLLMRequest(config, request);
 }
 
 void ChatClientInterface::clearMessages()
 {
-    m_chatHistory = {};
+    m_chatHistory.clear();
     m_accumulatedResponse.clear();
-    m_pendingMessage.clear();
     logMessage("Chat history cleared");
 }
 
+QVector<ChatMessage> ChatClientInterface::getChatHistory() const
+{
+    return m_chatHistory.getMessages();
+}
+
 void ChatClientInterface::handleLLMResponse(const QString &response, bool isComplete)
 {
     m_accumulatedResponse += response;
@@ -119,42 +160,33 @@ void ChatClientInterface::handleLLMResponse(const QString &response, bool isComp
         logMessage("Message completed. Final response: " + m_accumulatedResponse);
         emit messageReceived(m_accumulatedResponse.trimmed());
 
-        QJsonObject assistantMessage;
+        m_chatHistory.addMessage(ChatMessage::Role::Assistant, m_accumulatedResponse.trimmed());
-        assistantMessage["role"] = "assistant";
-        assistantMessage["content"] = m_accumulatedResponse.trimmed();
-        m_chatHistory.append(assistantMessage);
-
-        m_pendingMessage.clear();
         m_accumulatedResponse.clear();
-
-        trimChatHistory();
     }
 }
 
-void ChatClientInterface::trimChatHistory()
+QJsonArray ChatClientInterface::prepareMessagesForRequest() const
 {
-    int maxTokens = 4000;
+    QJsonArray messages;
-    int totalTokens = 0;
-    QJsonArray newHistory;
 
-    if (!m_chatHistory.isEmpty()
+    messages.append(QJsonObject{{"role", "system"}, {"content", m_chatHistory.getSystemPrompt()}});
-        && m_chatHistory.first().toObject()["role"].toString() == "system") {
-        newHistory.append(m_chatHistory.first());
-    }
 
-    for (int i = m_chatHistory.size() - 1; i >= 0; --i) {
+    for (const auto &message : m_chatHistory.getMessages()) {
-        QJsonObject message = m_chatHistory[i].toObject();
+        QString role;
-        int messageTokens = message["content"].toString().length() / 4;
+        switch (message.role) {
-
+        case ChatMessage::Role::User:
-        if (totalTokens + messageTokens > maxTokens) {
+            role = "user";
             break;
+        case ChatMessage::Role::Assistant:
+            role = "assistant";
+            break;
+        default:
+            continue;
         }
-
+        messages.append(QJsonObject{{"role", role}, {"content", message.content}});
-        newHistory.prepend(message);
-        totalTokens += messageTokens;
     }
 
-    m_chatHistory = newHistory;
+    return messages;
 }
 
 } // namespace QodeAssist::Chat

chat/ChatClientInterface.hpp

@@ -20,12 +20,41 @@
 #pragma once
 
 #include <QObject>
-#include <QtCore/qjsonarray.h>
+#include <QString>
+#include <QVector>
 #include "QodeAssistData.hpp"
 #include "core/LLMRequestHandler.hpp"
 
 namespace QodeAssist::Chat {
 
+struct ChatMessage
+{
+    enum class Role { System, User, Assistant };
+    Role role;
+    QString content;
+    int tokenCount;
+};
+
+class ChatHistory
+{
+public:
+    void addMessage(ChatMessage::Role role, const QString &content);
+    void clear();
+    QVector<ChatMessage> getMessages() const;
+    QString getSystemPrompt() const;
+    void setSystemPrompt(const QString &prompt);
+    void trim();
+
+private:
+    QVector<ChatMessage> m_messages;
+    QString m_systemPrompt;
+    int m_totalTokens = 0;
+    static const int MAX_HISTORY_SIZE = 50;
+    static const int MAX_TOKENS = 4000;
+
+    int estimateTokenCount(const QString &text) const;
+};
+
 class ChatClientInterface : public QObject
 {
     Q_OBJECT
@@ -36,6 +65,7 @@ public:
 
     void sendMessage(const QString &message);
     void clearMessages();
+    QVector<ChatMessage> getChatHistory() const;
 
 signals:
     void messageReceived(const QString &message);
@@ -43,12 +73,11 @@ signals:
 
 private:
     void handleLLMResponse(const QString &response, bool isComplete);
-    void trimChatHistory();
+    QJsonArray prepareMessagesForRequest() const;
 
     LLMRequestHandler *m_requestHandler;
     QString m_accumulatedResponse;
-    QString m_pendingMessage;
+    ChatHistory m_chatHistory;
-    QJsonArray m_chatHistory;
 };
 
 } // namespace QodeAssist::Chat

providers/LMStudioProvider.cpp

@@ -55,9 +55,20 @@ QString LMStudioProvider::chatEndpoint() const
 void LMStudioProvider::prepareRequest(QJsonObject &request)
 {
     auto &settings = Settings::presetPromptsSettings();
+    QJsonArray messages;
+
+    if (request.contains("system")) {
+        QJsonObject systemMessage{{"role", "system"},
+                                  {"content", request.take("system").toString()}};
+        messages.append(systemMessage);
+    }
+
     if (request.contains("prompt")) {
-        QJsonArray messages{
+        QJsonObject userMessage{{"role", "user"}, {"content", request.take("prompt").toString()}};
-            {QJsonObject{{"role", "user"}, {"content", request.take("prompt").toString()}}}};
+        messages.append(userMessage);
+    }
+
+    if (!messages.isEmpty()) {
         request["messages"] = std::move(messages);
     }
 

providers/OllamaProvider.cpp

@@ -59,7 +59,6 @@ void OllamaProvider::prepareRequest(QJsonObject &request)
 
     QJsonObject options;
     options["num_predict"] = settings.maxTokens();
-    options["keep_alive"] = settings.ollamaLivetime();
     options["temperature"] = settings.temperature();
     if (settings.useTopP())
         options["top_p"] = settings.topP();
@@ -70,6 +69,7 @@ void OllamaProvider::prepareRequest(QJsonObject &request)
     if (settings.usePresencePenalty())
         options["presence_penalty"] = settings.presencePenalty();
     request["options"] = options;
+    request["keep_alive"] = settings.ollamaLivetime();
 }
 
 bool OllamaProvider::handleResponse(QNetworkReply *reply, QString &accumulatedResponse)

providers/OpenAICompatProvider.cpp

@@ -24,7 +24,6 @@
 #include <QJsonObject>
 #include <QNetworkReply>
 
-#include "PromptTemplateManager.hpp"
 #include "settings/PresetPromptsSettings.hpp"
 
 namespace QodeAssist::Providers {
@@ -54,9 +53,20 @@ QString OpenAICompatProvider::chatEndpoint() const
 void OpenAICompatProvider::prepareRequest(QJsonObject &request)
 {
     auto &settings = Settings::presetPromptsSettings();
+    QJsonArray messages;
+
+    if (request.contains("system")) {
+        QJsonObject systemMessage{{"role", "system"},
+                                  {"content", request.take("system").toString()}};
+        messages.append(systemMessage);
+    }
+
     if (request.contains("prompt")) {
-        QJsonArray messages{
+        QJsonObject userMessage{{"role", "user"}, {"content", request.take("prompt").toString()}};
-            {QJsonObject{{"role", "user"}, {"content", request.take("prompt").toString()}}}};
+        messages.append(userMessage);
+    }
+
+    if (!messages.isEmpty()) {
         request["messages"] = std::move(messages);
     }
 

settings/ContextSettings.cpp

@@ -17,8 +17,6 @@
  * along with QodeAssist. If not, see <https://www.gnu.org/licenses/>.
  */
 
-#pragma once
-
 #include "ContextSettings.hpp"
 
 #include <QMessageBox>
@@ -60,11 +58,11 @@ ContextSettings::ContextSettings()
     useFilePathInContext.setDefaultValue(false);
     useFilePathInContext.setLabelText(Tr::tr("Use File Path in Context"));
 
-    useSpecificInstructions.setSettingsKey(Constants::USE_SPECIFIC_INSTRUCTIONS);
+    useSpecificInstructions.setSettingsKey(Constants::USE_SYSTEM_PROMPT);
     useSpecificInstructions.setDefaultValue(true);
-    useSpecificInstructions.setLabelText(Tr::tr("Use Specific Instructions"));
+    useSpecificInstructions.setLabelText(Tr::tr("Use System Prompt"));
 
-    specificInstractions.setSettingsKey(Constants::SPECIFIC_INSTRUCTIONS);
+    specificInstractions.setSettingsKey(Constants::SYSTEM_PROMPT);
     specificInstractions.setDisplayStyle(Utils::StringAspect::TextEditDisplay);
     specificInstractions.setLabelText(
         Tr::tr("Instructions: Please keep %1 for languge name, warning, it shouldn't too big"));

settings/GeneralSettings.cpp

@@ -131,6 +131,11 @@ GeneralSettings::GeneralSettings()
     loadProviders();
     loadPrompts();
 
+    llmProviders.setDefaultValue(llmProviders.indexForDisplay("Ollama"));
+    chatLlmProviders.setDefaultValue(chatLlmProviders.indexForDisplay("Ollama"));
+    fimPrompts.setDefaultValue(fimPrompts.indexForDisplay("CodeLLama FIM"));
+    chatPrompts.setDefaultValue(chatPrompts.indexForDisplay("CodeLLama Chat"));
+
     readSettings();
 
     auto fimProviderName = llmProviders.displayForIndex(llmProviders.value());
@@ -273,12 +278,12 @@ void GeneralSettings::resetPageToDefaults()
         resetAspect(startSuggestionTimer);
         resetAspect(autoCompletionTypingInterval);
         resetAspect(autoCompletionCharThreshold);
+        resetAspect(llmProviders);
+        resetAspect(chatLlmProviders);
+        resetAspect(fimPrompts);
+        resetAspect(chatPrompts);
     }
 
-    int fimIndex = llmProviders.indexForDisplay("Ollama");
-    llmProviders.setVolatileValue(fimIndex);
-    int chatIndex = chatLlmProviders.indexForDisplay("Ollama");
-    chatLlmProviders.setVolatileValue(chatIndex);
     modelName.setVolatileValue("");
     chatModelName.setVolatileValue("");
 

templates/CodeLlamaFimTemplate.hpp

@@ -27,7 +27,7 @@ class CodeLlamaFimTemplate : public PromptTemplate
 {
 public:
     TemplateType type() const override { return TemplateType::Fim; }
-    QString name() const override { return "CodeLlama FIM"; }
+    QString name() const override { return "CodeLLama FIM"; }
     QString promptTemplate() const override { return "%1<PRE> %2 <SUF>%3 <MID>"; }
     QStringList stopWords() const override
     {

templates/CodeLlamaInstruct.hpp

@@ -28,7 +28,7 @@ class CodeLlamaInstructTemplate : public PromptTemplate
 {
 public:
     TemplateType type() const override { return TemplateType::Chat; }
-    QString name() const override { return "CodeLlama Chat"; }
+    QString name() const override { return "CodeLLama Chat"; }
     QString promptTemplate() const override { return "[INST] %1 [/INST]"; }
     QStringList stopWords() const override { return QStringList() << "[INST]" << "[/INST]"; }