🐛 fix: Set content type for requests

Add Alpaca, Llama3, LLama2, ChatML templates

CMakeLists.txt

@@ -40,16 +40,22 @@ add_qtc_plugin(QodeAssist
     QodeAssistConstants.hpp
     QodeAssisttr.h
     LLMClientInterface.hpp LLMClientInterface.cpp
+    templates/Templates.hpp
     templates/CodeLlamaFim.hpp
     templates/StarCoder2Fim.hpp
     templates/DeepSeekCoderFim.hpp
     templates/CustomFimTemplate.hpp
-    templates/DeepSeekCoderChat.hpp
-    templates/CodeLlamaChat.hpp
+
+
     templates/Qwen.hpp
-    templates/StarCoderChat.hpp
+
     templates/Ollama.hpp
     templates/BasicChat.hpp
+    templates/Llama3.hpp
+    templates/ChatML.hpp
+    templates/Alpaca.hpp
+    templates/Llama2.hpp
+    providers/Providers.hpp
     providers/OllamaProvider.hpp providers/OllamaProvider.cpp
     providers/LMStudioProvider.hpp providers/LMStudioProvider.cpp
     providers/OpenAICompatProvider.hpp providers/OpenAICompatProvider.cpp

ChatView/ClientInterface.cpp

@@ -68,6 +68,8 @@ void ClientInterface::sendMessage(const QString &message, bool includeCurrentFil
 {
     cancelRequest();
 
+    m_chatModel->addMessage(message, ChatModel::ChatRole::User, "");
+
     auto &chatAssistantSettings = Settings::chatAssistantSettings();
 
     auto providerName = Settings::generalSettings().caProvider();
@@ -124,11 +126,11 @@ void ClientInterface::sendMessage(const QString &message, bool includeCurrentFil
     config.url = QString("%1%2").arg(Settings::generalSettings().caUrl(), provider->chatEndpoint());
     config.providerRequest = providerRequest;
     config.multiLineCompletion = false;
+    config.apiKey = Settings::chatAssistantSettings().apiKey();
 
     QJsonObject request;
     request["id"] = QUuid::createUuid().toString();
 
-    m_chatModel->addMessage(message, ChatModel::ChatRole::User, "");
     m_requestHandler->sendLLMRequest(config, request);
 }
 
@@ -148,11 +150,16 @@ void ClientInterface::handleLLMResponse(const QString &response,
                                         const QJsonObject &request,
                                         bool isComplete)
 {
-    QString messageId = request["id"].toString();
-    m_chatModel->addMessage(response.trimmed(), ChatModel::ChatRole::Assistant, messageId);
+    const auto message = response.trimmed();
 
-    if (isComplete) {
-        LOG_MESSAGE("Message completed. Final response for message " + messageId + ": " + response);
+    if (!message.isEmpty()) {
+        QString messageId = request["id"].toString();
+        m_chatModel->addMessage(message, ChatModel::ChatRole::Assistant, messageId);
+
+        if (isComplete) {
+            LOG_MESSAGE(
+                "Message completed. Final response for message " + messageId + ": " + response);
+        }
     }
 }
 

DocumentContextReader.cpp

@@ -210,10 +210,11 @@ LLMCore::ContextData DocumentContextReader::prepareContext(int lineNumber, int c
 
     QString fileContext;
     if (Settings::codeCompletionSettings().useFilePathInContext())
-        fileContext += getLanguageAndFileInfo();
+        fileContext.append("\n ").append(getLanguageAndFileInfo());
 
     if (Settings::codeCompletionSettings().useProjectChangesCache())
-        fileContext += ChangesManager::instance().getRecentChangesContext(m_textDocument);
+        fileContext.append("\n ").append(
+            ChangesManager::instance().getRecentChangesContext(m_textDocument));
 
     return {contextBefore, contextAfter, fileContext};
 }

LLMClientInterface.cpp

@@ -172,6 +172,7 @@ void LLMClientInterface::handleCompletion(const QJsonObject &request)
     config.promptTemplate = promptTemplate;
     config.url = QUrl(
         QString("%1%2").arg(Settings::generalSettings().ccUrl(), provider->completionEndpoint()));
+    config.apiKey = Settings::codeCompletionSettings().apiKey();
 
     config.providerRequest = {{"model", Settings::generalSettings().ccModel()}, {"stream", true}};
 

QodeAssist.json.in

@@ -1,6 +1,6 @@
 {
     "Name" : "QodeAssist",
-    "Version" : "0.3.9",
+    "Version" : "0.3.10",
     "CompatVersion" : "${IDE_VERSION_COMPAT}",
     "Vendor" : "Petr Mironychev",
     "Copyright" : "(C) ${IDE_COPYRIGHT_YEAR} Petr Mironychev, (C) ${IDE_COPYRIGHT_YEAR} The Qt Company Ltd",

README.md

@@ -29,8 +29,8 @@ QodeAssist is an AI-powered coding assistant plugin for Qt Creator. It provides
   - Side and Bottom panels
 - Support for multiple LLM providers:
   - Ollama
-  - LM Studio (experimental)
-  - OpenAI-compatible providers (experimental)
+  - LM Studio
+  - OpenAI-compatible providers(eg. https://openrouter.ai)
 - Extensive library of model-specific templates
 - Custom template support
 - Easy configuration and model selection
@@ -120,6 +120,28 @@ ollama run qwen2.5-coder:7b-instruct
 ollama run deepseek-coder-v2
 ```
 
+### Template-Model Compatibility
+
+| Template | Compatible Models | Purpose |
+|----------|------------------|----------|
+| CodeLlama FIM | `codellama:code` | Code completion |
+| DeepSeekCoder FIM | `deepseek-coder-v2`, `deepseek-v2.5` | Code completion |
+| Ollama Auto FIM | `Any Ollama base model` | Code completion |
+| Qwen FIM | `Qwen 2.5 models` | Code completion |
+| StarCoder2 FIM | `starcoder2 base model` | Code completion |
+| Alpaca | `starcoder2:instruct` | Chat assistance |
+| Basic Chat| `Messages without tokens` | Chat assistance |
+| ChatML | `Qwen 2.5 models` | Chat assistance |
+| Llama2 | `llama2 model family`, `codellama:instruct` | Chat assistance |
+| Llama3 | `llama3 model family` | Chat assistance |
+| Ollama Auto Chat | `Any Ollama chat model` | Chat assistance |
+
+> Note: 
+> - FIM (Fill-in-Middle) templates are optimized for code completion
+> - Chat templates are designed for interactive dialogue
+> - The Ollama Auto templates automatically adapt to most Ollama models
+> - Custom Template allows you to define your own prompt format
+
 ## QtCreator Version Compatibility
 
 - QtCreator 14.0.2 - 0.2.3 - 0.3.x 

llmcore/PromptTemplate.hpp

@@ -38,5 +38,6 @@ public:
     virtual QString promptTemplate() const = 0;
     virtual QStringList stopWords() const = 0;
     virtual void prepareRequest(QJsonObject &request, const ContextData &context) const = 0;
+    virtual QString description() const = 0;
 };
 } // namespace QodeAssist::LLMCore

llmcore/RequestConfig.hpp

@@ -35,6 +35,7 @@ struct LLMConfig
     QJsonObject providerRequest;
     RequestType requestType;
     bool multiLineCompletion;
+    QString apiKey;
 };
 
 } // namespace QodeAssist::LLMCore

llmcore/RequestHandler.cpp

@@ -38,7 +38,7 @@ void RequestHandler::sendLLMRequest(const LLMConfig &config, const QJsonObject &
                             QJsonDocument(config.providerRequest).toJson(QJsonDocument::Indented))));
 
     QNetworkRequest networkRequest(config.url);
-    prepareNetworkRequest(networkRequest, config.providerRequest);
+    prepareNetworkRequest(networkRequest, config.apiKey);
 
     QNetworkReply *reply = m_manager->post(networkRequest,
                                            QJsonDocument(config.providerRequest).toJson());
@@ -84,6 +84,8 @@ void RequestHandler::handleLLMResponse(QNetworkReply *reply,
         if (isComplete) {
             auto cleanedCompletion = removeStopWords(accumulatedResponse,
                                                      config.promptTemplate->stopWords());
+            removeCodeBlockWrappers(cleanedCompletion);
+
             emit completionReceived(cleanedCompletion, request, true);
         }
     } else if (config.requestType == RequestType::Chat) {
@@ -107,33 +109,33 @@ bool RequestHandler::cancelRequest(const QString &id)
     return false;
 }
 
-void RequestHandler::prepareNetworkRequest(QNetworkRequest &networkRequest,
-                                           const QJsonObject &providerRequest)
+void RequestHandler::prepareNetworkRequest(
+    QNetworkRequest &networkRequest, const QString &apiKey) const
 {
     networkRequest.setHeader(QNetworkRequest::ContentTypeHeader, "application/json");
 
-    if (providerRequest.contains("api_key")) {
-        QString apiKey = providerRequest["api_key"].toString();
+    if (!apiKey.isEmpty()) {
         networkRequest.setRawHeader("Authorization", QString("Bearer %1").arg(apiKey).toUtf8());
     }
 }
 
-bool RequestHandler::processSingleLineCompletion(QNetworkReply *reply,
-                                                 const QJsonObject &request,
-                                                 const QString &accumulatedResponse,
-                                                 const LLMConfig &config)
+bool RequestHandler::processSingleLineCompletion(
+    QNetworkReply *reply,
+    const QJsonObject &request,
+    const QString &accumulatedResponse,
+    const LLMConfig &config)
 {
-    int newlinePos = accumulatedResponse.indexOf('\n');
+    QString cleanedResponse = accumulatedResponse;
+    removeCodeBlockWrappers(cleanedResponse);
 
+    int newlinePos = cleanedResponse.indexOf('\n');
     if (newlinePos != -1) {
-        QString singleLineCompletion = accumulatedResponse.left(newlinePos).trimmed();
-        singleLineCompletion = removeStopWords(singleLineCompletion,
-                                               config.promptTemplate->stopWords());
-
+        QString singleLineCompletion = cleanedResponse.left(newlinePos).trimmed();
+        singleLineCompletion
+            = removeStopWords(singleLineCompletion, config.promptTemplate->stopWords());
         emit completionReceived(singleLineCompletion, request, true);
         m_accumulatedResponses.remove(reply);
         reply->abort();
-
         return true;
     }
     return false;
@@ -150,4 +152,36 @@ QString RequestHandler::removeStopWords(const QStringView &completion, const QSt
     return filteredCompletion;
 }
 
+void RequestHandler::removeCodeBlockWrappers(QString &response)
+{
+    static const QRegularExpression
+        fullCodeBlockRegex(R"(```[\w\s]*\n([\s\S]*?)```)", QRegularExpression::MultilineOption);
+    static const QRegularExpression
+        partialStartBlockRegex(R"(```[\w\s]*\n([\s\S]*?)$)", QRegularExpression::MultilineOption);
+    static const QRegularExpression
+        partialEndBlockRegex(R"(^([\s\S]*?)```)", QRegularExpression::MultilineOption);
+
+    QRegularExpressionMatchIterator matchIterator = fullCodeBlockRegex.globalMatch(response);
+    while (matchIterator.hasNext()) {
+        QRegularExpressionMatch match = matchIterator.next();
+        QString codeBlock = match.captured(0);
+        QString codeContent = match.captured(1).trimmed();
+        response.replace(codeBlock, codeContent);
+    }
+
+    QRegularExpressionMatch startMatch = partialStartBlockRegex.match(response);
+    if (startMatch.hasMatch()) {
+        QString partialBlock = startMatch.captured(0);
+        QString codeContent = startMatch.captured(1).trimmed();
+        response.replace(partialBlock, codeContent);
+    }
+
+    QRegularExpressionMatch endMatch = partialEndBlockRegex.match(response);
+    if (endMatch.hasMatch()) {
+        QString partialBlock = endMatch.captured(0);
+        QString codeContent = endMatch.captured(1).trimmed();
+        response.replace(partialBlock, codeContent);
+    }
+}
+
 } // namespace QodeAssist::LLMCore

llmcore/RequestHandler.hpp

@@ -52,12 +52,13 @@ private:
     QMap<QString, QNetworkReply *> m_activeRequests;
     QMap<QNetworkReply *, QString> m_accumulatedResponses;
 
-    void prepareNetworkRequest(QNetworkRequest &networkRequest, const QJsonObject &providerRequest);
+    void prepareNetworkRequest(QNetworkRequest &networkRequest, const QString &apiKey) const;
     bool processSingleLineCompletion(QNetworkReply *reply,
                                      const QJsonObject &request,
                                      const QString &accumulatedResponse,
                                      const LLMConfig &config);
     QString removeStopWords(const QStringView &completion, const QStringList &stopWords);
+    void removeCodeBlockWrappers(QString &response);
 };
 
 } // namespace QodeAssist::LLMCore

providers/OpenAICompatProvider.cpp

@@ -26,13 +26,15 @@
 #include <QJsonObject>
 #include <QNetworkReply>
 
+#include "logger/Logger.hpp"
+
 namespace QodeAssist::Providers {
 
 OpenAICompatProvider::OpenAICompatProvider() {}
 
 QString OpenAICompatProvider::name() const
 {
-    return "OpenAI Compatible (experimental)";
+    return "OpenAI Compatible";
 }
 
 QString OpenAICompatProvider::url() const
@@ -82,10 +84,6 @@ void OpenAICompatProvider::prepareRequest(QJsonObject &request, LLMCore::Request
             request["frequency_penalty"] = settings.frequencyPenalty();
         if (settings.usePresencePenalty())
             request["presence_penalty"] = settings.presencePenalty();
-        const QString &apiKey = settings.apiKey();
-        if (!apiKey.isEmpty()) {
-            request["api_key"] = apiKey;
-        }
     };
 
     QJsonArray messages = prepareMessages(request);
@@ -103,24 +101,41 @@ void OpenAICompatProvider::prepareRequest(QJsonObject &request, LLMCore::Request
 bool OpenAICompatProvider::handleResponse(QNetworkReply *reply, QString &accumulatedResponse)
 {
     bool isComplete = false;
+    QString tempResponse = accumulatedResponse;
+
     while (reply->canReadLine()) {
         QByteArray line = reply->readLine().trimmed();
         if (line.isEmpty()) {
             continue;
         }
-        if (line == "data: [DONE]") {
+
+        if (!line.startsWith("data:")) {
+            continue;
+        }
+
+        line = line.mid(6);
+
+        if (line == "[DONE]") {
             isComplete = true;
             break;
         }
-        if (line.startsWith("data: ")) {
-            line = line.mid(6); // Remove "data: " prefix
-        }
+
         QJsonDocument jsonResponse = QJsonDocument::fromJson(line);
         if (jsonResponse.isNull()) {
-            qWarning() << "Invalid JSON response from LM Studio:" << line;
+            LOG_MESSAGE(
+                "Invalid JSON response from OpenAI compatible provider: " + QString::fromUtf8(line));
             continue;
         }
+
         QJsonObject responseObj = jsonResponse.object();
+
+        if (responseObj.contains("error")) {
+            LOG_MESSAGE(
+                "OpenAI compatible provider error: "
+                + QString::fromUtf8(QJsonDocument(responseObj).toJson(QJsonDocument::Indented)));
+            return false;
+        }
+
         if (responseObj.contains("choices")) {
             QJsonArray choices = responseObj["choices"].toArray();
             if (!choices.isEmpty()) {
@@ -128,16 +143,30 @@ bool OpenAICompatProvider::handleResponse(QNetworkReply *reply, QString &accumul
                 QJsonObject delta = choice["delta"].toObject();
                 if (delta.contains("content")) {
                     QString completion = delta["content"].toString();
-
-                    accumulatedResponse += completion;
+                    if (!completion.isEmpty()) {
+                        tempResponse += completion;
+                    }
                 }
-                if (choice["finish_reason"].toString() == "stop") {
+                QString finishReason = choice["finish_reason"].toString();
+                if (!finishReason.isNull() && finishReason == "stop") {
                     isComplete = true;
-                    break;
                 }
             }
         }
+
+        if (responseObj.contains("usage")) {
+            QJsonObject usage = responseObj["usage"].toObject();
+            LOG_MESSAGE(QString("Token usage - Prompt: %1, Completion: %2, Total: %3")
+                            .arg(usage["prompt_tokens"].toInt())
+                            .arg(usage["completion_tokens"].toInt())
+                            .arg(usage["total_tokens"].toInt()));
+        }
     }
+
+    if (!tempResponse.isEmpty()) {
+        accumulatedResponse = tempResponse;
+    }
+
     return isComplete;
 }
 

providers/Providers.hpp

@@ -0,0 +1,37 @@
+/* 
+ * Copyright (C) 2024 Petr Mironychev
+ *
+ * This file is part of QodeAssist.
+ *
+ * QodeAssist is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * QodeAssist is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with QodeAssist. If not, see <https://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+
+#include "llmcore/ProvidersManager.hpp"
+#include "providers/LMStudioProvider.hpp"
+#include "providers/OllamaProvider.hpp"
+#include "providers/OpenAICompatProvider.hpp"
+
+namespace QodeAssist::Providers {
+
+inline void registerProviders()
+{
+    auto &providerManager = LLMCore::ProvidersManager::instance();
+    providerManager.registerProvider<OllamaProvider>();
+    providerManager.registerProvider<LMStudioProvider>();
+    providerManager.registerProvider<OpenAICompatProvider>();
+}
+
+} // namespace QodeAssist::Providers

qodeassist.cpp

@@ -43,22 +43,9 @@
 #include "QodeAssistClient.hpp"
 #include "chat/ChatOutputPane.h"
 #include "chat/NavigationPanel.hpp"
-#include "llmcore/PromptTemplateManager.hpp"
-#include "llmcore/ProvidersManager.hpp"
-#include "providers/LMStudioProvider.hpp"
-#include "providers/OllamaProvider.hpp"
-#include "providers/OpenAICompatProvider.hpp"
 
-#include "templates/BasicChat.hpp"
-#include "templates/CodeLlamaChat.hpp"
-#include "templates/CodeLlamaFim.hpp"
-#include "templates/CustomFimTemplate.hpp"
-#include "templates/DeepSeekCoderChat.hpp"
-#include "templates/DeepSeekCoderFim.hpp"
-#include "templates/Ollama.hpp"
-#include "templates/Qwen.hpp"
-#include "templates/StarCoder2Fim.hpp"
-#include "templates/StarCoderChat.hpp"
+#include "providers/Providers.hpp"
+#include "templates/Templates.hpp"
 
 using namespace Utils;
 using namespace Core;
@@ -85,25 +72,8 @@ public:
 
     void initialize() final
     {
-        auto &providerManager = LLMCore::ProvidersManager::instance();
-        providerManager.registerProvider<Providers::OllamaProvider>();
-        providerManager.registerProvider<Providers::LMStudioProvider>();
-        providerManager.registerProvider<Providers::OpenAICompatProvider>();
-
-        auto &templateManager = LLMCore::PromptTemplateManager::instance();
-        templateManager.registerTemplate<Templates::CodeLlamaFim>();
-        templateManager.registerTemplate<Templates::StarCoder2Fim>();
-        templateManager.registerTemplate<Templates::DeepSeekCoderFim>();
-        templateManager.registerTemplate<Templates::CustomTemplate>();
-        templateManager.registerTemplate<Templates::DeepSeekCoderChat>();
-        templateManager.registerTemplate<Templates::CodeLlamaChat>();
-        templateManager.registerTemplate<Templates::LlamaChat>();
-        templateManager.registerTemplate<Templates::StarCoderChat>();
-        templateManager.registerTemplate<Templates::QwenChat>();
-        templateManager.registerTemplate<Templates::QwenFim>();
-        templateManager.registerTemplate<Templates::OllamaAutoFim>();
-        templateManager.registerTemplate<Templates::OllamaAutoChat>();
-        templateManager.registerTemplate<Templates::BasicChat>();
+        Providers::registerProviders();
+        Templates::registerTemplates();
 
         Utils::Icon QCODEASSIST_ICON(
             {{":/resources/images/qoderassist-icon.png", Utils::Theme::IconsBaseColor}});

settings/CodeCompletionSettings.cpp

@@ -143,9 +143,8 @@ CodeCompletionSettings::CodeCompletionSettings()
 
     systemPrompt.setSettingsKey(Constants::CC_SYSTEM_PROMPT);
     systemPrompt.setDisplayStyle(Utils::StringAspect::TextEditDisplay);
-    systemPrompt.setDefaultValue(
-        "You are an expert C++, Qt, and QML code completion AI. Your task is to provide accurate "
-        "and contextually appropriate code suggestions.");
+    systemPrompt.setDefaultValue("You are an expert C++, Qt, and QML code completion AI. ANSWER "
+                                 "should be SHORT and in CODE");
 
     useFilePathInContext.setSettingsKey(Constants::CC_USE_FILE_PATH_IN_CONTEXT);
     useFilePathInContext.setDefaultValue(true);

templates/Alpaca.hpp

@@ -0,0 +1,67 @@
+/* 
+ * Copyright (C) 2024 Petr Mironychev
+ *
+ * This file is part of QodeAssist.
+ *
+ * QodeAssist is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * QodeAssist is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with QodeAssist. If not, see <https://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+
+#include "llmcore/PromptTemplate.hpp"
+#include <QJsonArray>
+
+namespace QodeAssist::Templates {
+
+class Alpaca : public LLMCore::PromptTemplate
+{
+public:
+    QString name() const override { return "Alpaca"; }
+    LLMCore::TemplateType type() const override { return LLMCore::TemplateType::Chat; }
+    QString promptTemplate() const override { return {}; }
+    QStringList stopWords() const override
+    {
+        return QStringList() << "### Instruction:" << "### Response:";
+    }
+    void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
+    {
+        QJsonArray messages = request["messages"].toArray();
+
+        for (int i = 0; i < messages.size(); ++i) {
+            QJsonObject message = messages[i].toObject();
+            QString role = message["role"].toString();
+            QString content = message["content"].toString();
+
+            QString formattedContent;
+            if (role == "system") {
+                formattedContent = content + "\n\n";
+            } else if (role == "user") {
+                formattedContent = "### Instruction:\n" + content + "\n\n";
+            } else if (role == "assistant") {
+                formattedContent = "### Response:\n" + content + "\n\n";
+            }
+
+            message["content"] = formattedContent;
+            messages[i] = message;
+        }
+
+        request["messages"] = messages;
+    }
+    QString description() const override
+    {
+        return "The message will contain the following tokens: ### Instruction:\n### Response:\n";
+    }
+};
+
+} // namespace QodeAssist::Templates

templates/BasicChat.hpp

@@ -32,18 +32,9 @@ public:
     QString name() const override { return "Basic Chat"; }
     QString promptTemplate() const override { return {}; }
     QStringList stopWords() const override { return QStringList(); }
-
     void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
-    {
-        QJsonArray messages = request["messages"].toArray();
-
-        QJsonObject newMessage;
-        newMessage["role"] = "user";
-        newMessage["content"] = context.prefix;
-        messages.append(newMessage);
-
-        request["messages"] = messages;
-    }
+    {}
+    QString description() const override { return "chat without tokens"; }
 };
 
 } // namespace QodeAssist::Templates

templates/ChatML.hpp

@@ -20,35 +20,41 @@
 #pragma once
 
 #include <QJsonArray>
+
 #include "llmcore/PromptTemplate.hpp"
 
 namespace QodeAssist::Templates {
 
-class DeepSeekCoderChat : public LLMCore::PromptTemplate
+class ChatML : public LLMCore::PromptTemplate
 {
 public:
-    QString name() const override { return "DeepSeekCoder Chat"; }
+    QString name() const override { return "ChatML"; }
     LLMCore::TemplateType type() const override { return LLMCore::TemplateType::Chat; }
-
-    QString promptTemplate() const override { return "### Instruction:\n%1\n### Response:\n"; }
-
+    QString promptTemplate() const override { return {}; }
     QStringList stopWords() const override
     {
-        return QStringList() << "### Instruction:" << "### Response:" << "\n\n### " << "<|EOT|>";
+        return QStringList() << "<|im_start|>" << "<|im_end|>";
     }
-
     void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
     {
-        QString formattedPrompt = promptTemplate().arg(context.prefix);
         QJsonArray messages = request["messages"].toArray();
 
-        QJsonObject newMessage;
-        newMessage["role"] = "user";
-        newMessage["content"] = formattedPrompt;
-        messages.append(newMessage);
+        for (int i = 0; i < messages.size(); ++i) {
+            QJsonObject message = messages[i].toObject();
+            QString role = message["role"].toString();
+            QString content = message["content"].toString();
+
+            message["content"] = QString("<|im_start|>%1\n%2\n<|im_end|>").arg(role, content);
+
+            messages[i] = message;
+        }
 
         request["messages"] = messages;
     }
+    QString description() const override
+    {
+        return "The message will contain the following tokens: <|im_start|>%1\n%2\n<|im_end|>";
+    }
 };
 
 } // namespace QodeAssist::Templates

templates/CodeLlamaFim.hpp

@@ -33,12 +33,15 @@ public:
     {
         return QStringList() << "<EOT>" << "<PRE>" << "<SUF" << "<MID>";
     }
-
     void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
     {
         QString formattedPrompt = promptTemplate().arg(context.prefix, context.suffix);
         request["prompt"] = formattedPrompt;
     }
+    QString description() const override
+    {
+        return "The message will contain the following tokens: <PRE> %1 <SUF>%2 <MID>";
+    }
 };
 
 } // namespace QodeAssist::Templates

templates/CustomFimTemplate.hpp

@@ -39,7 +39,6 @@ public:
         return Settings::customPromptSettings().customJsonTemplate();
     }
     QStringList stopWords() const override { return QStringList(); }
-
     void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
     {
         QJsonDocument doc = QJsonDocument::fromJson(promptTemplate().toUtf8());
@@ -56,6 +55,7 @@ public:
             request[it.key()] = it.value();
         }
     }
+    QString description() const override { return promptTemplate(); }
 
 private:
     QJsonValue processJsonValue(const QJsonValue &value, const LLMCore::ContextData &context) const

templates/DeepSeekCoderFim.hpp

@@ -38,6 +38,11 @@ public:
         QString formattedPrompt = promptTemplate().arg(context.prefix, context.suffix);
         request["prompt"] = formattedPrompt;
     }
+    QString description() const override
+    {
+        return "The message will contain the following tokens: "
+               "<｜fim▁begin｜>%1<｜fim▁hole｜>%2<｜fim▁end｜>";
+    }
 };
 
 } // namespace QodeAssist::Templates

templates/Llama2.hpp

@@ -19,38 +19,46 @@
 
 #pragma once
 
-#include <QJsonArray>
-
 #include "llmcore/PromptTemplate.hpp"
+#include <QJsonArray>
 
 namespace QodeAssist::Templates {
 
-class CodeLlamaChat : public LLMCore::PromptTemplate
+class Llama2 : public LLMCore::PromptTemplate
 {
 public:
+    QString name() const override { return "Llama 2"; }
     LLMCore::TemplateType type() const override { return LLMCore::TemplateType::Chat; }
-    QString name() const override { return "CodeLlama Chat"; }
-    QString promptTemplate() const override { return "[INST] %1 [/INST]"; }
-    QStringList stopWords() const override { return QStringList() << "[INST]" << "[/INST]"; }
-
+    QString promptTemplate() const override { return {}; }
+    QStringList stopWords() const override { return QStringList() << "[INST]"; }
     void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
     {
-        QString formattedPrompt = promptTemplate().arg(context.prefix);
         QJsonArray messages = request["messages"].toArray();
 
-        QJsonObject newMessage;
-        newMessage["role"] = "user";
-        newMessage["content"] = formattedPrompt;
-        messages.append(newMessage);
+        for (int i = 0; i < messages.size(); ++i) {
+            QJsonObject message = messages[i].toObject();
+            QString role = message["role"].toString();
+            QString content = message["content"].toString();
+
+            QString formattedContent;
+            if (role == "system") {
+                formattedContent = QString("[INST]<<SYS>>\n%1\n<</SYS>>[/INST]\n").arg(content);
+            } else if (role == "user") {
+                formattedContent = QString("[INST]%1[/INST]\n").arg(content);
+            } else if (role == "assistant") {
+                formattedContent = content + "\n";
+            }
+
+            message["content"] = formattedContent;
+            messages[i] = message;
+        }
 
         request["messages"] = messages;
     }
-};
-
-class LlamaChat : public CodeLlamaChat
-{
-public:
-    QString name() const override { return "Llama Chat"; }
+    QString description() const override
+    {
+        return "The message will contain the following tokens: [INST]%1[/INST]\n";
+    }
 };
 
 } // namespace QodeAssist::Templates

templates/Llama3.hpp

@@ -20,32 +20,43 @@
 #pragma once
 
 #include <QJsonArray>
+
 #include "llmcore/PromptTemplate.hpp"
 
 namespace QodeAssist::Templates {
 
-class StarCoderChat : public LLMCore::PromptTemplate
+class Llama3 : public LLMCore::PromptTemplate
 {
 public:
-    QString name() const override { return "StarCoder Chat"; }
+    QString name() const override { return "Llama 3"; }
     LLMCore::TemplateType type() const override { return LLMCore::TemplateType::Chat; }
-    QString promptTemplate() const override { return "### Instruction:\n%1\n### Response:\n"; }
+    QString promptTemplate() const override { return ""; }
     QStringList stopWords() const override
     {
-        return QStringList() << "###"
-                             << "<|endoftext|>" << "<file_sep>";
+        return QStringList() << "<|start_header_id|>" << "<|end_header_id|>" << "<|eot_id|>";
     }
     void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
     {
-        QString formattedPrompt = promptTemplate().arg(context.prefix);
         QJsonArray messages = request["messages"].toArray();
 
-        QJsonObject newMessage;
-        newMessage["role"] = "user";
-        newMessage["content"] = formattedPrompt;
-        messages.append(newMessage);
+        for (int i = 0; i < messages.size(); ++i) {
+            QJsonObject message = messages[i].toObject();
+            QString role = message["role"].toString();
+            QString content = message["content"].toString();
+
+            message["content"]
+                = QString("<|start_header_id|>%1<|end_header_id|>%2<|eot_id|>").arg(role, content);
+
+            messages[i] = message;
+        }
 
         request["messages"] = messages;
     }
+    QString description() const override
+    {
+        return "The message will contain the following tokens: "
+               "<|start_header_id|>%1<|end_header_id|>%2<|eot_id|>";
+    }
 };
+
 } // namespace QodeAssist::Templates

templates/Ollama.hpp

@@ -32,12 +32,12 @@ public:
     QString name() const override { return "Ollama Auto FIM"; }
     QString promptTemplate() const override { return {}; }
     QStringList stopWords() const override { return QStringList(); }
-
     void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
     {
         request["prompt"] = context.prefix;
         request["suffix"] = context.suffix;
     }
+    QString description() const override { return "template will take from ollama modelfile"; }
 };
 
 class OllamaAutoChat : public LLMCore::PromptTemplate
@@ -59,6 +59,7 @@ public:
 
         request["messages"] = messages;
     }
+    QString description() const override { return "template will take from ollama modelfile"; }
 };
 
 } // namespace QodeAssist::Templates

templates/Qwen.hpp

@@ -24,33 +24,6 @@
 
 namespace QodeAssist::Templates {
 
-class QwenChat : public LLMCore::PromptTemplate
-{
-public:
-    QString name() const override { return "Qwen Chat"; }
-    LLMCore::TemplateType type() const override { return LLMCore::TemplateType::Chat; }
-
-    QString promptTemplate() const override { return "### Instruction:\n%1\n### Response:\n"; }
-
-    QStringList stopWords() const override
-    {
-        return QStringList() << "### Instruction:" << "### Response:" << "\n\n### " << "<|EOT|>";
-    }
-
-    void prepareRequest(QJsonObject &request, const LLMCore::ContextData &context) const override
-    {
-        QString formattedPrompt = promptTemplate().arg(context.prefix);
-        QJsonArray messages = request["messages"].toArray();
-
-        QJsonObject newMessage;
-        newMessage["role"] = "user";
-        newMessage["content"] = formattedPrompt;
-        messages.append(newMessage);
-
-        request["messages"] = messages;
-    }
-};
-
 class QwenFim : public LLMCore::PromptTemplate
 {
 public:
@@ -66,6 +39,11 @@ public:
         QString formattedPrompt = promptTemplate().arg(context.prefix, context.suffix);
         request["prompt"] = formattedPrompt;
     }
+    QString description() const override
+    {
+        return "The message will contain the following tokens: "
+               "<|fim_prefix|>%1<|fim_suffix|>%2<|fim_middle|>";
+    }
 };
 
 } // namespace QodeAssist::Templates

templates/StarCoder2Fim.hpp

@@ -39,6 +39,11 @@ public:
         QString formattedPrompt = promptTemplate().arg(context.prefix, context.suffix);
         request["prompt"] = formattedPrompt;
     }
+    QString description() const override
+    {
+        return "The message will contain the following tokens: "
+               "<fim_prefix>%1<fim_suffix>%2<fim_middle>";
+    }
 };
 
 } // namespace QodeAssist::Templates

templates/Templates.hpp

@@ -0,0 +1,54 @@
+/* 
+ * Copyright (C) 2024 Petr Mironychev
+ *
+ * This file is part of QodeAssist.
+ *
+ * QodeAssist is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * QodeAssist is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with QodeAssist. If not, see <https://www.gnu.org/licenses/>.
+ */
+
+#pragma once
+
+#include "llmcore/PromptTemplateManager.hpp"
+#include "templates/Alpaca.hpp"
+#include "templates/BasicChat.hpp"
+#include "templates/ChatML.hpp"
+#include "templates/CodeLlamaFim.hpp"
+#include "templates/CustomFimTemplate.hpp"
+#include "templates/DeepSeekCoderFim.hpp"
+#include "templates/Llama2.hpp"
+#include "templates/Llama3.hpp"
+#include "templates/Ollama.hpp"
+#include "templates/Qwen.hpp"
+#include "templates/StarCoder2Fim.hpp"
+
+namespace QodeAssist::Templates {
+
+inline void registerTemplates()
+{
+    auto &templateManager = LLMCore::PromptTemplateManager::instance();
+    templateManager.registerTemplate<CodeLlamaFim>();
+    templateManager.registerTemplate<StarCoder2Fim>();
+    templateManager.registerTemplate<DeepSeekCoderFim>();
+    templateManager.registerTemplate<CustomTemplate>();
+    templateManager.registerTemplate<QwenFim>();
+    templateManager.registerTemplate<OllamaAutoFim>();
+    templateManager.registerTemplate<OllamaAutoChat>();
+    templateManager.registerTemplate<BasicChat>();
+    templateManager.registerTemplate<Llama3>();
+    templateManager.registerTemplate<ChatML>();
+    templateManager.registerTemplate<Alpaca>();
+    templateManager.registerTemplate<Llama2>();
+}
+
+} // namespace QodeAssist::Templates