✨ feat: Add using instruct model in code completion

* ✨ feat: Add MessageBuilder for code completion
* ✨ feat: Add move text from request to comments
* ✨ feat: Add settings for process text of instruct model
* 🐛 fix: Add stop to ollama request validator
* 🐛 fix: Template double delete

CMakeLists.txt

@@ -67,4 +67,5 @@ add_qtc_plugin(QodeAssist
     chat/ChatOutputPane.h chat/ChatOutputPane.cpp
     chat/NavigationPanel.hpp chat/NavigationPanel.cpp
     ConfigurationManager.hpp ConfigurationManager.cpp
+    CodeHandler.hpp CodeHandler.cpp
 )

CodeHandler.cpp

@@ -0,0 +1,115 @@
+/* 
+ * 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/>.
+ */
+
+#include "CodeHandler.hpp"
+#include <QHash>
+
+namespace QodeAssist {
+
+QString CodeHandler::processText(QString text)
+{
+    QString result;
+    QStringList lines = text.split('\n');
+    bool inCodeBlock = false;
+    QString pendingComments;
+    QString currentLanguage;
+
+    for (const QString &line : lines) {
+        if (line.trimmed().startsWith("```")) {
+            if (!inCodeBlock) {
+                currentLanguage = detectLanguage(line);
+            }
+            inCodeBlock = !inCodeBlock;
+            continue;
+        }
+
+        if (inCodeBlock) {
+            if (!pendingComments.isEmpty()) {
+                QStringList commentLines = pendingComments.split('\n');
+                QString commentPrefix = getCommentPrefix(currentLanguage);
+
+                for (const QString &commentLine : commentLines) {
+                    if (!commentLine.trimmed().isEmpty()) {
+                        result += commentPrefix + " " + commentLine.trimmed() + "\n";
+                    } else {
+                        result += "\n";
+                    }
+                }
+                pendingComments.clear();
+            }
+            result += line + "\n";
+        } else {
+            QString trimmed = line.trimmed();
+            if (!trimmed.isEmpty()) {
+                pendingComments += trimmed + "\n";
+            } else {
+                pendingComments += "\n";
+            }
+        }
+    }
+
+    return result;
+}
+
+QString CodeHandler::getCommentPrefix(const QString &language)
+{
+    static const QHash<QString, QString> commentPrefixes
+        = {{"python", "#"},  {"py", "#"},          {"lua", "--"},   {"javascript", "//"},
+           {"js", "//"},     {"typescript", "//"}, {"ts", "//"},    {"cpp", "//"},
+           {"c++", "//"},    {"c", "//"},          {"java", "//"},  {"csharp", "//"},
+           {"cs", "//"},     {"php", "//"},        {"ruby", "#"},   {"rb", "#"},
+           {"rust", "//"},   {"rs", "//"},         {"go", "//"},    {"swift", "//"},
+           {"kotlin", "//"}, {"kt", "//"},         {"scala", "//"}, {"r", "#"},
+           {"shell", "#"},   {"bash", "#"},        {"sh", "#"},     {"perl", "#"},
+           {"pl", "#"},      {"haskell", "--"},    {"hs", "--"}};
+
+    return commentPrefixes.value(language.toLower(), "//");
+}
+
+QString CodeHandler::detectLanguage(const QString &line)
+{
+    QString trimmed = line.trimmed();
+    if (trimmed.length() <= 3) { // Если только ```
+        return QString();
+    }
+
+    return trimmed.mid(3).trimmed();
+}
+
+const QRegularExpression &CodeHandler::getFullCodeBlockRegex()
+{
+    static const QRegularExpression
+        regex(R"(```[\w\s]*\n([\s\S]*?)```)", QRegularExpression::MultilineOption);
+    return regex;
+}
+
+const QRegularExpression &CodeHandler::getPartialStartBlockRegex()
+{
+    static const QRegularExpression
+        regex(R"(```[\w\s]*\n([\s\S]*?)$)", QRegularExpression::MultilineOption);
+    return regex;
+}
+
+const QRegularExpression &CodeHandler::getPartialEndBlockRegex()
+{
+    static const QRegularExpression regex(R"(^([\s\S]*?)```)", QRegularExpression::MultilineOption);
+    return regex;
+}
+
+} // namespace QodeAssist

CodeHandler.hpp

@@ -0,0 +1,42 @@
+/* 
+ * 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 <QObject>
+#include <QRegularExpression>
+#include <QString>
+
+namespace QodeAssist {
+
+class CodeHandler
+{
+public:
+    static QString processText(QString text);
+
+private:
+    static QString getCommentPrefix(const QString &language);
+    static QString detectLanguage(const QString &line);
+
+    static const QRegularExpression &getFullCodeBlockRegex();
+    static const QRegularExpression &getPartialStartBlockRegex();
+    static const QRegularExpression &getPartialEndBlockRegex();
+};
+
+} // namespace QodeAssist

LLMClientInterface.cpp

@@ -26,7 +26,9 @@
 #include <llmcore/RequestConfig.hpp>
 #include <texteditor/textdocument.h>
 
+#include "CodeHandler.hpp"
 #include "DocumentContextReader.hpp"
+#include "llmcore/MessageBuilder.hpp"
 #include "llmcore/PromptTemplateManager.hpp"
 #include "llmcore/ProvidersManager.hpp"
 #include "logger/Logger.hpp"
@@ -167,11 +169,13 @@ void LLMClientInterface::handleCompletion(const QJsonObject &request)
     }
 
     LLMCore::LLMConfig config;
-    config.requestType = LLMCore::RequestType::Fim;
+    config.requestType = LLMCore::RequestType::CodeCompletion;
     config.provider = provider;
     config.promptTemplate = promptTemplate;
-    config.url = QUrl(
-        QString("%1%2").arg(Settings::generalSettings().ccUrl(), provider->completionEndpoint()));
+    config.url = QUrl(QString("%1%2").arg(
+        Settings::generalSettings().ccUrl(),
+        promptTemplate->type() == LLMCore::TemplateType::Fim ? provider->completionEndpoint()
+                                                             : provider->chatEndpoint()));
     config.apiKey = Settings::codeCompletionSettings().apiKey();
 
     config.providerRequest
@@ -180,20 +184,27 @@ void LLMClientInterface::handleCompletion(const QJsonObject &request)
 
     config.multiLineCompletion = completeSettings.multiLineCompletion();
 
+    const auto stopWords = QJsonArray::fromStringList(config.promptTemplate->stopWords());
+    if (!stopWords.isEmpty())
+        config.providerRequest["stop"] = stopWords;
+
     QString systemPrompt;
     if (completeSettings.useSystemPrompt())
         systemPrompt.append(completeSettings.systemPrompt());
     if (!updatedContext.fileContext.isEmpty())
         systemPrompt.append(updatedContext.fileContext);
-    if (!systemPrompt.isEmpty())
-        config.providerRequest["system"] = systemPrompt;
 
-    const auto stopWords = QJsonArray::fromStringList(config.promptTemplate->stopWords());
-    if (!stopWords.isEmpty())
-        config.providerRequest["stop"] = stopWords;
+    auto message = LLMCore::MessageBuilder()
+                       .addSystemMessage(systemPrompt)
+                       .addUserMessage(updatedContext.prefix)
+                       .addSuffix(updatedContext.suffix)
+                       .addtTokenizer(promptTemplate);
 
-    config.promptTemplate->prepareRequest(config.providerRequest, updatedContext);
-    config.provider->prepareRequest(config.providerRequest, LLMCore::RequestType::Fim);
+    message.saveTo(
+        config.providerRequest,
+        providerName == "Ollama" ? LLMCore::ProvidersApi::Ollama : LLMCore::ProvidersApi::OpenAI);
+
+    config.provider->prepareRequest(config.providerRequest, LLMCore::RequestType::CodeCompletion);
 
     auto errors = config.provider->validateRequest(config.providerRequest, promptTemplate->type());
     if (!errors.isEmpty()) {
@@ -232,19 +243,31 @@ void LLMClientInterface::sendCompletionToClient(const QString &completion,
                                                 const QJsonObject &request,
                                                 bool isComplete)
 {
+    auto templateName = Settings::generalSettings().ccTemplate();
+    auto promptTemplate = LLMCore::PromptTemplateManager::instance().getFimTemplateByName(
+        templateName);
+
     QJsonObject position = request["params"].toObject()["doc"].toObject()["position"].toObject();
 
     QJsonObject response;
     response["jsonrpc"] = "2.0";
     response[LanguageServerProtocol::idKey] = request["id"];
+
     QJsonObject result;
     QJsonArray completions;
     QJsonObject completionItem;
-    completionItem[LanguageServerProtocol::textKey] = completion;
+
+    QString processedCompletion
+        = promptTemplate->type() == LLMCore::TemplateType::Chat
+                  && Settings::codeCompletionSettings().smartProcessInstuctText()
+              ? CodeHandler::processText(completion)
+              : completion;
+
+    completionItem[LanguageServerProtocol::textKey] = processedCompletion;
     QJsonObject range;
     range["start"] = position;
     QJsonObject end = position;
-    end["character"] = position["character"].toInt() + completion.length();
+    end["character"] = position["character"].toInt() + processedCompletion.length();
     range["end"] = end;
     completionItem[LanguageServerProtocol::rangeKey] = range;
     completionItem[LanguageServerProtocol::positionKey] = position;

llmcore/CMakeLists.txt

@@ -10,6 +10,7 @@ add_library(LLMCore STATIC
     OllamaMessage.hpp OllamaMessage.cpp
     OpenAIMessage.hpp OpenAIMessage.cpp
     ValidationUtils.hpp ValidationUtils.cpp
+    MessageBuilder.hpp MessageBuilder.cpp
 )
 
 target_link_libraries(LLMCore

llmcore/MessageBuilder.cpp

@@ -0,0 +1,3 @@
+#include "MessageBuilder.hpp"
+
+namespace QodeAssist::LLMCore {} // namespace QodeAssist::LLMCore

llmcore/MessageBuilder.hpp

@@ -0,0 +1,127 @@
+/* 
+ * 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 <QJsonArray>
+#include <QJsonObject>
+#include <QString>
+#include <QVector>
+
+#include "PromptTemplate.hpp"
+
+namespace QodeAssist::LLMCore {
+
+enum class MessageRole { System, User, Assistant };
+
+enum class OllamaFormat { Messages, Completions };
+
+enum class ProvidersApi { Ollama, OpenAI };
+
+static const QString ROLE_SYSTEM = "system";
+static const QString ROLE_USER = "user";
+static const QString ROLE_ASSISTANT = "assistant";
+
+struct Message
+{
+    MessageRole role;
+    QString content;
+};
+
+class MessageBuilder
+{
+public:
+    MessageBuilder &addSystemMessage(const QString &content)
+    {
+        m_systemMessage = content;
+        return *this;
+    }
+
+    MessageBuilder &addUserMessage(const QString &content)
+    {
+        m_messages.append({MessageRole::User, content});
+        return *this;
+    }
+
+    MessageBuilder &addSuffix(const QString &content)
+    {
+        m_suffix = content;
+        return *this;
+    }
+
+    MessageBuilder &addtTokenizer(PromptTemplate *promptTemplate)
+    {
+        m_promptTemplate = promptTemplate;
+        return *this;
+    }
+
+    QString roleToString(MessageRole role) const
+    {
+        switch (role) {
+        case MessageRole::System:
+            return ROLE_SYSTEM;
+        case MessageRole::User:
+            return ROLE_USER;
+        case MessageRole::Assistant:
+            return ROLE_ASSISTANT;
+        default:
+            return ROLE_USER;
+        }
+    }
+
+    void saveTo(QJsonObject &request, ProvidersApi api)
+    {
+        if (!m_promptTemplate) {
+            return;
+        }
+
+        if (api == ProvidersApi::Ollama) {
+            ContextData context{
+                m_messages.isEmpty() ? QString() : m_messages.last().content,
+                m_suffix,
+                m_systemMessage};
+
+            if (m_promptTemplate->type() == TemplateType::Fim) {
+                m_promptTemplate->prepareRequest(request, context);
+            } else {
+                QJsonArray messages;
+
+                messages.append(QJsonObject{{"role", "system"}, {"content", m_systemMessage}});
+                messages.append(
+                    QJsonObject{{"role", "user"}, {"content", m_messages.last().content}});
+                request["messages"] = messages;
+                m_promptTemplate->prepareRequest(request, {});
+            }
+        } else if (api == ProvidersApi::OpenAI) {
+            QJsonArray messages;
+
+            messages.append(QJsonObject{{"role", "system"}, {"content", m_systemMessage}});
+            messages.append(QJsonObject{{"role", "user"}, {"content", m_messages.last().content}});
+            request["messages"] = messages;
+            m_promptTemplate->prepareRequest(request, {});
+        }
+    }
+
+private:
+    QString m_systemMessage;
+    QString m_suffix;
+    QVector<Message> m_messages;
+    PromptTemplate *m_promptTemplate;
+};
+} // namespace QodeAssist::LLMCore

llmcore/PromptTemplateManager.cpp

@@ -40,7 +40,6 @@ QStringList PromptTemplateManager::chatTemplatesNames() const
 PromptTemplateManager::~PromptTemplateManager()
 {
     qDeleteAll(m_fimTemplates);
-    qDeleteAll(m_chatTemplates);
 }
 
 PromptTemplate *PromptTemplateManager::getFimTemplateByName(const QString &templateName)

llmcore/PromptTemplateManager.hpp

@@ -39,9 +39,8 @@ public:
                       "T must inherit from PromptTemplate");
         T *template_ptr = new T();
         QString name = template_ptr->name();
-        if (template_ptr->type() == TemplateType::Fim) {
-            m_fimTemplates[name] = template_ptr;
-        } else if (template_ptr->type() == TemplateType::Chat) {
+        m_fimTemplates[name] = template_ptr;
+        if (template_ptr->type() == TemplateType::Chat) {
             m_chatTemplates[name] = template_ptr;
         }
     }

llmcore/RequestHandler.cpp

@@ -75,7 +75,7 @@ void RequestHandler::handleLLMResponse(QNetworkReply *reply,
 
     bool isComplete = config.provider->handleResponse(reply, accumulatedResponse);
 
-    if (config.requestType == RequestType::Fim) {
+    if (config.requestType == RequestType::CodeCompletion) {
         if (!config.multiLineCompletion
             && processSingleLineCompletion(reply, request, accumulatedResponse, config)) {
             return;
@@ -84,7 +84,6 @@ void RequestHandler::handleLLMResponse(QNetworkReply *reply,
         if (isComplete) {
             auto cleanedCompletion = removeStopWords(accumulatedResponse,
                                                      config.promptTemplate->stopWords());
-            removeCodeBlockWrappers(cleanedCompletion);
 
             emit completionReceived(cleanedCompletion, request, true);
         }
@@ -126,7 +125,6 @@ bool RequestHandler::processSingleLineCompletion(
     const LLMConfig &config)
 {
     QString cleanedResponse = accumulatedResponse;
-    removeCodeBlockWrappers(cleanedResponse);
 
     int newlinePos = cleanedResponse.indexOf('\n');
     if (newlinePos != -1) {

llmcore/RequestType.hpp

@@ -21,5 +21,5 @@
 
 namespace QodeAssist::LLMCore {
 
-enum RequestType { Fim, Chat };
+enum RequestType { CodeCompletion, Chat };
 }

providers/LMStudioProvider.cpp

@@ -94,7 +94,7 @@ void LMStudioProvider::prepareRequest(QJsonObject &request, LLMCore::RequestType
         request["messages"] = std::move(messages);
     }
 
-    if (type == LLMCore::RequestType::Fim) {
+    if (type == LLMCore::RequestType::CodeCompletion) {
         applyModelParams(Settings::codeCompletionSettings());
     } else {
         applyModelParams(Settings::chatAssistantSettings());

providers/OllamaProvider.cpp

@@ -66,6 +66,7 @@ void OllamaProvider::prepareRequest(QJsonObject &request, LLMCore::RequestType t
         QJsonObject options;
         options["num_predict"] = settings.maxTokens();
         options["temperature"] = settings.temperature();
+        options["stop"] = request.take("stop");
 
         if (settings.useTopP())
             options["top_p"] = settings.topP();
@@ -80,7 +81,7 @@ void OllamaProvider::prepareRequest(QJsonObject &request, LLMCore::RequestType t
         request["keep_alive"] = settings.ollamaLivetime();
     };
 
-    if (type == LLMCore::RequestType::Fim) {
+    if (type == LLMCore::RequestType::CodeCompletion) {
         applySettings(Settings::codeCompletionSettings());
     } else {
         applySettings(Settings::chatAssistantSettings());
@@ -150,6 +151,7 @@ QList<QString> OllamaProvider::validateRequest(const QJsonObject &request, LLMCo
         {"options",
          QJsonObject{
              {"temperature", {}},
+             {"stop", {}},
              {"top_p", {}},
              {"top_k", {}},
              {"num_predict", {}},
@@ -164,6 +166,7 @@ QList<QString> OllamaProvider::validateRequest(const QJsonObject &request, LLMCo
         {"options",
          QJsonObject{
              {"temperature", {}},
+             {"stop", {}},
              {"top_p", {}},
              {"top_k", {}},
              {"num_predict", {}},

providers/OpenAICompatProvider.cpp

@@ -93,7 +93,7 @@ void OpenAICompatProvider::prepareRequest(QJsonObject &request, LLMCore::Request
         request["messages"] = std::move(messages);
     }
 
-    if (type == LLMCore::RequestType::Fim) {
+    if (type == LLMCore::RequestType::CodeCompletion) {
         applyModelParams(Settings::codeCompletionSettings());
     } else {
         applyModelParams(Settings::chatAssistantSettings());

providers/OpenRouterAIProvider.cpp

@@ -77,7 +77,7 @@ void OpenRouterProvider::prepareRequest(QJsonObject &request, LLMCore::RequestTy
         request["messages"] = std::move(messages);
     }
 
-    if (type == LLMCore::RequestType::Fim) {
+    if (type == LLMCore::RequestType::CodeCompletion) {
         applyModelParams(Settings::codeCompletionSettings());
     } else {
         applyModelParams(Settings::chatAssistantSettings());

settings/CodeCompletionSettings.cpp

@@ -48,13 +48,17 @@ CodeCompletionSettings::CodeCompletionSettings()
     autoCompletion.setDefaultValue(true);
 
     multiLineCompletion.setSettingsKey(Constants::CC_MULTILINE_COMPLETION);
-    multiLineCompletion.setDefaultValue(false);
-    multiLineCompletion.setLabelText(Tr::tr("Enable Multiline Completion(experimental)"));
+    multiLineCompletion.setDefaultValue(true);
+    multiLineCompletion.setLabelText(Tr::tr("Enable Multiline Completion"));
 
     stream.setSettingsKey(Constants::CC_STREAM);
     stream.setDefaultValue(true);
     stream.setLabelText(Tr::tr("Enable stream option"));
 
+    smartProcessInstuctText.setSettingsKey(Constants::CC_SMART_PROCESS_INSTRUCT_TEXT);
+    smartProcessInstuctText.setDefaultValue(true);
+    smartProcessInstuctText.setLabelText(Tr::tr("Enable smart process text from instruct model"));
+
     startSuggestionTimer.setSettingsKey(Constants::СС_START_SUGGESTION_TIMER);
     startSuggestionTimer.setLabelText(Tr::tr("with delay(ms)"));
     startSuggestionTimer.setRange(10, 10000);
@@ -147,8 +151,8 @@ 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. ANSWER "
-                                 "should be SHORT and in CODE");
+    systemPrompt.setDefaultValue("You are an expert C++, Qt, and QML code completion AI. Answer "
+                                 "should be ONLY in CODE and without repeating current.");
 
     useFilePathInContext.setSettingsKey(Constants::CC_USE_FILE_PATH_IN_CONTEXT);
     useFilePathInContext.setDefaultValue(true);
@@ -227,6 +231,7 @@ CodeCompletionSettings::CodeCompletionSettings()
                     Space{8},
                     multiLineCompletion,
                     stream,
+                    smartProcessInstuctText,
                     Row{autoCompletionCharThreshold,
                         autoCompletionTypingInterval,
                         startSuggestionTimer,

settings/CodeCompletionSettings.hpp

@@ -36,6 +36,7 @@ public:
     Utils::BoolAspect autoCompletion{this};
     Utils::BoolAspect multiLineCompletion{this};
     Utils::BoolAspect stream{this};
+    Utils::BoolAspect smartProcessInstuctText{this};
 
     Utils::IntegerAspect startSuggestionTimer{this};
     Utils::IntegerAspect autoCompletionCharThreshold{this};

settings/SettingsConstants.hpp

@@ -50,6 +50,7 @@ const char СС_AUTO_COMPLETION_TYPING_INTERVAL[] = "QodeAssist.autoCompletionTy
 const char MAX_FILE_THRESHOLD[] = "QodeAssist.maxFileThreshold";
 const char CC_MULTILINE_COMPLETION[] = "QodeAssist.ccMultilineCompletion";
 const char CC_STREAM[] = "QodeAssist.ccStream";
+const char CC_SMART_PROCESS_INSTRUCT_TEXT[] = "QodeAssist.ccSmartProcessInstructText";
 const char CUSTOM_JSON_TEMPLATE[] = "QodeAssist.customJsonTemplate";
 const char CA_TOKENS_THRESHOLD[] = "QodeAssist.caTokensThreshold";
 const char CA_SHARING_CURRENT_FILE[] = "QodeAssist.caSharingCurrentFile";