Enha: atomic global vars instead of mutexes

.gitignore

@@ -3,6 +3,8 @@
 testlog
 history/
 *.db
+*.db-shm
+*.db-wal
 config.toml
 sysprompts/*
 !sysprompts/alice_bob_carl.json
@@ -15,3 +17,4 @@ gflt
 chat_exports/*.json
 ragimport
 .env
+onnx/

Makefile

@@ -1,4 +1,4 @@
-.PHONY: setconfig run lint lintall install-linters setup-whisper build-whisper download-whisper-model docker-up docker-down docker-logs noextra-run installdelve checkdelve
+.PHONY: setconfig run lint lintall install-linters setup-whisper build-whisper download-whisper-model docker-up docker-down docker-logs noextra-run installdelve checkdelve fetch-onnx install-onnx-deps
 
 run: setconfig
 	go build -tags extra -o gf-lt && ./gf-lt
@@ -30,6 +30,105 @@ lint: ## Run linters. Use make install-linters first.
 lintall: lint
 	noblanks ./...
 
+fetch-onnx:
+	mkdir -p onnx/embedgemma && curl -o onnx/embedgemma/config.json -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/config.json && curl -o onnx/embedgemma/tokenizer.json -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/tokenizer.json && curl -o onnx/embedgemma/model_q4.onnx -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/onnx/model_q4.onnx && curl -o onnx/embedgemma/model_q4.onnx_data -L https://huggingface.co/onnx-community/embeddinggemma-300m-ONNX/resolve/main/onnx/model_q4.onnx_data?download=true
+
+install-onnx-deps: ## Install ONNX Runtime with CUDA support (or CPU fallback)
+	@echo "=== ONNX Runtime Installer ===" && \
+	echo "" && \
+	echo "Checking for existing ONNX Runtime..." && \
+	if ldconfig -p 2>/dev/null | grep -q libonnxruntime.so.1; then \
+		echo "ONNX Runtime is already installed:" && \
+		ldconfig -p 2>/dev/null | grep libonnxruntime && \
+		echo "" && \
+		echo "Skipping installation. To reinstall, remove existing libs first:" && \
+		echo "  sudo rm -f /usr/local/lib/libonnxruntime*.so*" && \
+		exit 0; \
+	fi && \
+	echo "No ONNX Runtime found. Proceeding with installation..." && \
+	echo "" && \
+	echo "Detecting CUDA version..." && \
+	HAS_CUDA=0 && \
+	if command -v nvidia-smi >/dev/null 2>&1; then \
+		CUDA_INFO=$$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1) && \
+		if [ -n "$$CUDA_INFO" ]; then \
+			echo "Found NVIDIA GPU with driver: $$CUDA_INFO" && \
+			HAS_CUDA=1; \
+		else \
+			echo "NVIDIA driver found but could not detect CUDA version"; \
+		fi; \
+	else \
+		echo "No NVIDIA GPU detected (nvidia-smi not found)"; \
+	fi && \
+	echo "" && \
+	echo "Determining ONNX Runtime version..." && \
+	ARCH=$$(uname -m) && \
+	if [ "$$ARCH" = "x86_64" ]; then \
+		ONNX_ARCH="x64"; \
+	elif [ "$$ARCH" = "aarch64" ] || [ "$$ARCH" = "arm64" ]; then \
+		ONNX_ARCH="aarch64"; \
+	else \
+		echo "Unsupported architecture: $$ARCH" && \
+		exit 1; \
+	fi && \
+	echo "Detected architecture: $$ARCH (ONNX runtime: $$ONNX_ARCH)" && \
+	if [ "$$HAS_CUDA" = "1" ]; then \
+		echo "Installing ONNX Runtime with CUDA support..."; \
+		ONNX_VERSION="1.24.2"; \
+	else \
+		echo "Installing ONNX Runtime (CPU version)..."; \
+		ONNX_VERSION="1.24.2"; \
+	fi && \
+	FILENAME="onnxruntime-linux-$${ONNX_ARCH}-${ONNX_VERSION}.tgz" && \
+	URL="https://github.com/microsoft/onnxruntime/releases/download/v$${ONNX_VERSION}/$${FILENAME}" && \
+	echo "Downloading $${URL}..." && \
+	mkdir -p /tmp/onnx-install && \
+	curl -L -o /tmp/onnx-install/$${FILENAME} "$${URL}" || { \
+		echo "Failed to download ONNX Runtime v$${ONNX_VERSION}. Trying v1.18.0..." && \
+		ONNX_VERSION="1.18.0" && \
+		FILENAME="onnxruntime-linux-$${ONNX_ARCH}-${ONNX_VERSION}.tgz" && \
+		URL="https://github.com/microsoft/onnxruntime/releases/download/v$${ONNX_VERSION}/$${FILENAME}" && \
+		curl -L -o /tmp/onnx-install/$${FILENAME} "$${URL}" || { \
+			echo "ERROR: Failed to download ONNX Runtime from GitHub" && \
+			echo "" && \
+			echo "Please install manually:" && \
+			echo "  1. Go to https://github.com/microsoft/onnxruntime/releases" && \
+			echo "  2. Download onnxruntime-linux-$${ONNX_ARCH}-VERSION.tgz" && \
+			echo "  3. Extract and copy to /usr/local/lib:" && \
+			echo "     tar -xzf onnxruntime-linux-$${ONNX_ARCH}-VERSION.tgz" && \
+			echo "     sudo cp -r onnxruntime-linux-$${ONNX_ARCH}-VERSION/lib/* /usr/local/lib/" && \
+			echo "     sudo ldconfig" && \
+			exit 1; \
+		}; \
+	} && \
+	echo "Extracting..." && \
+	cd /tmp/onnx-install && tar -xzf $${FILENAME} && \
+	echo "Installing to /usr/local/lib..." && \
+	ONNX_DIR=$$(find /tmp/onnx-install -maxdepth 1 -type d -name "onnxruntime-linux-*") && \
+	if [ -d "$${ONNX_DIR}/lib" ]; then \
+		cp -r $${ONNX_DIR}/lib/* /usr/local/lib/ 2>/dev/null || sudo cp -r $${ONNX_DIR}/lib/* /usr/local/lib/; \
+	else \
+		echo "ERROR: Could not find lib directory in extracted archive" && \
+		exit 1; \
+	fi && \
+	echo "Updating library cache..." && \
+	sudo ldconfig 2>/dev/null || ldconfig && \
+	echo "" && \
+	echo "=== Installation complete! ===" && \
+	echo "" && \
+	echo "Installed libraries:" && \
+	ldconfig -p | grep libonnxruntime || echo "(libraries may require logout/relogin to appear)" && \
+	echo "" && \
+	if [ "$$HAS_CUDA" = "1" ]; then \
+		echo "NOTE: CUDA-enabled ONNX Runtime installed."; \
+		echo "Ensure you also have CUDA libraries installed:"; \
+		echo "  - libcudnn, libcublas, libcurand"; \
+	else \
+		echo "NOTE: CPU-only ONNX Runtime installed."; \
+		echo "For GPU support, install CUDA and re-run this script."; \
+	fi && \
+	rm -rf /tmp/onnx-install
+
 # Whisper STT Setup (in batteries directory)
 setup-whisper: build-whisper download-whisper-model
 

bot.go

@@ -22,7 +22,7 @@ import (
 	"slices"
 	"strconv"
 	"strings"
-	"sync"
+	"sync/atomic"
 	"time"
 )
 
@@ -37,7 +37,7 @@ var (
 	chunkChan       = make(chan string, 10)
 	openAIToolChan  = make(chan string, 10)
 	streamDone      = make(chan bool, 1)
-	chatBody        *models.ChatBody
+	chatBody        *models.SafeChatBody
 	store           storage.FullRepo
 	defaultFirstMsg = "Hello! What can I do for you?"
 	defaultStarter  = []models.RoleMsg{}
@@ -49,7 +49,6 @@ var (
 	//nolint:unused // TTS_ENABLED conditionally uses this
 	orator          Orator
 	asr             STT
-	localModelsMu   sync.RWMutex
 	defaultLCPProps = map[string]float32{
 		"temperature":    0.8,
 		"dry_multiplier": 0.0,
@@ -64,11 +63,17 @@ var (
 		"google/gemma-3-27b-it:free",
 		"meta-llama/llama-3.3-70b-instruct:free",
 	}
-	LocalModels     = []string{}
+	LocalModels     atomic.Value // stores []string
-	localModelsData *models.LCPModels
+	localModelsData atomic.Value // stores *models.LCPModels
-	orModelsData    *models.ORModels
+	orModelsData    atomic.Value // stores *models.ORModels
 )
 
+func init() {
+	LocalModels.Store([]string{})
+	localModelsData.Store((*models.LCPModels)(nil))
+	orModelsData.Store((*models.ORModels)(nil))
+}
+
 var thinkBlockRE = regexp.MustCompile(`(?s)<think>.*?</think>`)
 
 // parseKnownToTag extracts known_to list from content using configured tag.
@@ -262,13 +267,13 @@ func warmUpModel() {
 		return
 	}
 	// Check if model is already loaded
-	loaded, err := isModelLoaded(chatBody.Model)
+	loaded, err := isModelLoaded(chatBody.GetModel())
 	if err != nil {
-		logger.Debug("failed to check model status", "model", chatBody.Model, "error", err)
+		logger.Debug("failed to check model status", "model", chatBody.GetModel(), "error", err)
 		// Continue with warmup attempt anyway
 	}
 	if loaded {
-		showToast("model already loaded", "Model "+chatBody.Model+" is already loaded.")
+		showToast("model already loaded", "Model "+chatBody.GetModel()+" is already loaded.")
 		return
 	}
 	go func() {
@@ -277,7 +282,7 @@ func warmUpModel() {
 		switch {
 		case strings.HasSuffix(cfg.CurrentAPI, "/completion"):
 			// Old completion endpoint
-			req := models.NewLCPReq(".", chatBody.Model, nil, map[string]float32{
+			req := models.NewLCPReq(".", chatBody.GetModel(), nil, map[string]float32{
 				"temperature":    0.8,
 				"dry_multiplier": 0.0,
 				"min_p":          0.05,
@@ -289,7 +294,7 @@ func warmUpModel() {
 			// OpenAI-compatible chat endpoint
 			req := models.OpenAIReq{
 				ChatBody: &models.ChatBody{
-					Model: chatBody.Model,
+					Model: chatBody.GetModel(),
 					Messages: []models.RoleMsg{
 						{Role: "system", Content: "."},
 					},
@@ -313,7 +318,7 @@ func warmUpModel() {
 		}
 		resp.Body.Close()
 		// Start monitoring for model load completion
-		monitorModelLoad(chatBody.Model)
+		monitorModelLoad(chatBody.GetModel())
 	}()
 }
 
@@ -356,7 +361,7 @@ func fetchORModels(free bool) ([]string, error) {
 	if err := json.NewDecoder(resp.Body).Decode(data); err != nil {
 		return nil, err
 	}
-	orModelsData = data
+	orModelsData.Store(data)
 	freeModels := data.ListModels(free)
 	return freeModels, nil
 }
@@ -418,7 +423,7 @@ func fetchLCPModelsWithStatus() (*models.LCPModels, error) {
 	if err := json.NewDecoder(resp.Body).Decode(data); err != nil {
 		return nil, err
 	}
-	localModelsData = data
+	localModelsData.Store(data)
 	return data, nil
 }
 
@@ -821,10 +826,10 @@ func chatRound(r *models.ChatRoundReq) error {
 	}
 	go sendMsgToLLM(reader)
 	logger.Debug("looking at vars in chatRound", "msg", r.UserMsg, "regen", r.Regen, "resume", r.Resume)
-	msgIdx := len(chatBody.Messages)
+	msgIdx := chatBody.GetMessageCount()
 	if !r.Resume {
 		// Add empty message to chatBody immediately so it persists during Alt+T toggle
-		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{
+		chatBody.AppendMessage(models.RoleMsg{
 			Role: botPersona, Content: "",
 		})
 		nl := "\n\n"
@@ -836,7 +841,7 @@ func chatRound(r *models.ChatRoundReq) error {
 		}
 		fmt.Fprintf(textView, "%s[-:-:b](%d) %s[-:-:-]\n", nl, msgIdx, roleToIcon(botPersona))
 	} else {
-		msgIdx = len(chatBody.Messages) - 1
+		msgIdx = chatBody.GetMessageCount() - 1
 	}
 	respText := strings.Builder{}
 	toolResp := strings.Builder{}
@@ -893,7 +898,10 @@ out:
 			fmt.Fprint(textView, chunk)
 			respText.WriteString(chunk)
 			// Update the message in chatBody.Messages so it persists during Alt+T
-			chatBody.Messages[msgIdx].Content = respText.String()
+			chatBody.UpdateMessageFunc(msgIdx, func(msg models.RoleMsg) models.RoleMsg {
+				msg.Content = respText.String()
+				return msg
+			})
 			if scrollToEndEnabled {
 				textView.ScrollToEnd()
 			}
@@ -936,29 +944,32 @@ out:
 	}
 	botRespMode = false
 	if r.Resume {
-		chatBody.Messages[len(chatBody.Messages)-1].Content += respText.String()
+		chatBody.UpdateMessageFunc(chatBody.GetMessageCount()-1, func(msg models.RoleMsg) models.RoleMsg {
-		updatedMsg := chatBody.Messages[len(chatBody.Messages)-1]
+			msg.Content += respText.String()
-		processedMsg := processMessageTag(&updatedMsg)
+			processedMsg := processMessageTag(&msg)
-		chatBody.Messages[len(chatBody.Messages)-1] = *processedMsg
+			if msgStats != nil && processedMsg.Role != cfg.ToolRole {
-		if msgStats != nil && chatBody.Messages[len(chatBody.Messages)-1].Role != cfg.ToolRole {
+				processedMsg.Stats = msgStats
-			chatBody.Messages[len(chatBody.Messages)-1].Stats = msgStats
 			}
+			return *processedMsg
+		})
 	} else {
-		chatBody.Messages[msgIdx].Content = respText.String()
+		chatBody.UpdateMessageFunc(msgIdx, func(msg models.RoleMsg) models.RoleMsg {
-		processedMsg := processMessageTag(&chatBody.Messages[msgIdx])
+			msg.Content = respText.String()
-		chatBody.Messages[msgIdx] = *processedMsg
+			processedMsg := processMessageTag(&msg)
-		if msgStats != nil && chatBody.Messages[msgIdx].Role != cfg.ToolRole {
+			if msgStats != nil && processedMsg.Role != cfg.ToolRole {
-			chatBody.Messages[msgIdx].Stats = msgStats
+				processedMsg.Stats = msgStats
 			}
-		stopTTSIfNotForUser(&chatBody.Messages[msgIdx])
+			return *processedMsg
+		})
+		stopTTSIfNotForUser(&chatBody.GetMessages()[msgIdx])
 	}
 	cleanChatBody()
 	refreshChatDisplay()
 	updateStatusLine()
 	// bot msg is done;
 	// now check it for func call
-	// logChat(activeChatName, chatBody.Messages)
+	// logChat(activeChatName, chatBody.GetMessages())
-	if err := updateStorageChat(activeChatName, chatBody.Messages); err != nil {
+	if err := updateStorageChat(activeChatName, chatBody.GetMessages()); err != nil {
 		logger.Warn("failed to update storage", "error", err, "name", activeChatName)
 	}
 	// Strip think blocks before parsing for tool calls
@@ -973,8 +984,8 @@ out:
 	// If so, trigger those characters to respond if that char is not controlled by user
 	// perhaps we should have narrator role to determine which char is next to act
 	if cfg.AutoTurn {
-		lastMsg := chatBody.Messages[len(chatBody.Messages)-1]
+		lastMsg, ok := chatBody.GetLastMessage()
-		if len(lastMsg.KnownTo) > 0 {
+		if ok && len(lastMsg.KnownTo) > 0 {
 			triggerPrivateMessageResponses(&lastMsg)
 		}
 	}
@@ -983,13 +994,15 @@ out:
 
 // cleanChatBody removes messages with null or empty content to prevent API issues
 func cleanChatBody() {
-	if chatBody == nil || chatBody.Messages == nil {
+	if chatBody == nil || chatBody.GetMessageCount() == 0 {
 		return
 	}
 	// Tool request cleaning is now configurable via AutoCleanToolCallsFromCtx (default false)
 	// /completion msg where part meant for user and other part tool call
 	// chatBody.Messages = cleanToolCalls(chatBody.Messages)
-	chatBody.Messages = consolidateAssistantMessages(chatBody.Messages)
+	chatBody.WithLock(func(cb *models.ChatBody) {
+		cb.Messages = consolidateAssistantMessages(cb.Messages)
+	})
 }
 
 // convertJSONToMapStringString unmarshals JSON into map[string]interface{} and converts all values to strings.
@@ -1089,7 +1102,7 @@ func findCall(msg, toolCall string) bool {
 				Content:    fmt.Sprintf("Error processing tool call: %v. Please check the JSON format and try again.", err),
 				ToolCallID: lastToolCall.ID, // Use the stored tool call ID
 			}
-			chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
+			chatBody.AppendMessage(toolResponseMsg)
 			// Clear the stored tool call ID after using it (no longer needed)
 			// Trigger the assistant to continue processing with the error message
 			crr := &models.ChatRoundReq{
@@ -1126,7 +1139,7 @@ func findCall(msg, toolCall string) bool {
 				Role:    cfg.ToolRole,
 				Content: "Error processing tool call: no valid JSON found. Please check the JSON format.",
 			}
-			chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
+			chatBody.AppendMessage(toolResponseMsg)
 			crr := &models.ChatRoundReq{
 				Role: cfg.AssistantRole,
 			}
@@ -1143,8 +1156,8 @@ func findCall(msg, toolCall string) bool {
 				Role:    cfg.ToolRole,
 				Content: fmt.Sprintf("Error processing tool call: %v. Please check the JSON format and try again.", err),
 			}
-			chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
+			chatBody.AppendMessage(toolResponseMsg)
-			logger.Debug("findCall: added tool error response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "message_count_after_add", len(chatBody.Messages))
+			logger.Debug("findCall: added tool error response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "message_count_after_add", chatBody.GetMessageCount())
 			// Trigger the assistant to continue processing with the error message
 			// chatRound("", cfg.AssistantRole, tv, false, false)
 			crr := &models.ChatRoundReq{
@@ -1162,17 +1175,23 @@ func findCall(msg, toolCall string) bool {
 	// we got here => last msg recognized as a tool call (correct or not)
 	// Use the tool call ID from streaming response (lastToolCall.ID)
 	// Don't generate random ID - the ID should match between assistant message and tool response
-	lastMsgIdx := len(chatBody.Messages) - 1
+	lastMsgIdx := chatBody.GetMessageCount() - 1
 	if lastToolCall.ID != "" {
-		chatBody.Messages[lastMsgIdx].ToolCallID = lastToolCall.ID
+		chatBody.UpdateMessageFunc(lastMsgIdx, func(msg models.RoleMsg) models.RoleMsg {
+			msg.ToolCallID = lastToolCall.ID
+			return msg
+		})
 	}
 	// Store tool call info in the assistant message
 	// Convert Args map to JSON string for storage
-	chatBody.Messages[lastMsgIdx].ToolCall = &models.ToolCall{
+	chatBody.UpdateMessageFunc(lastMsgIdx, func(msg models.RoleMsg) models.RoleMsg {
+		msg.ToolCall = &models.ToolCall{
 			ID:   lastToolCall.ID,
 			Name: lastToolCall.Name,
 			Args: mapToString(lastToolCall.Args),
 		}
+		return msg
+	})
 	// call a func
 	_, ok := fnMap[fc.Name]
 	if !ok {
@@ -1183,8 +1202,8 @@ func findCall(msg, toolCall string) bool {
 			Content:    m,
 			ToolCallID: lastToolCall.ID, // Use the stored tool call ID
 		}
-		chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
+		chatBody.AppendMessage(toolResponseMsg)
-		logger.Debug("findCall: added tool not implemented response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "tool_call_id", toolResponseMsg.ToolCallID, "message_count_after_add", len(chatBody.Messages))
+		logger.Debug("findCall: added tool not implemented response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "tool_call_id", toolResponseMsg.ToolCallID, "message_count_after_add", chatBody.GetMessageCount())
 		// Clear the stored tool call ID after using it
 		lastToolCall.ID = ""
 		// Trigger the assistant to continue processing with the new tool response
@@ -1255,9 +1274,9 @@ func findCall(msg, toolCall string) bool {
 		}
 	}
 	fmt.Fprintf(textView, "%s[-:-:b](%d) <%s>: [-:-:-]\n%s\n",
-		"\n\n", len(chatBody.Messages), cfg.ToolRole, toolResponseMsg.GetText())
+		"\n\n", chatBody.GetMessageCount(), cfg.ToolRole, toolResponseMsg.GetText())
-	chatBody.Messages = append(chatBody.Messages, toolResponseMsg)
+	chatBody.AppendMessage(toolResponseMsg)
-	logger.Debug("findCall: added actual tool response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "tool_call_id", toolResponseMsg.ToolCallID, "message_count_after_add", len(chatBody.Messages))
+	logger.Debug("findCall: added actual tool response", "role", toolResponseMsg.Role, "content_len", len(toolResponseMsg.Content), "tool_call_id", toolResponseMsg.ToolCallID, "message_count_after_add", chatBody.GetMessageCount())
 	// Clear the stored tool call ID after using it
 	lastToolCall.ID = ""
 	// Trigger the assistant to continue processing with the new tool response
@@ -1387,28 +1406,29 @@ func charToStart(agentName string, keepSysP bool) bool {
 func updateModelLists() {
 	var err error
 	if cfg.OpenRouterToken != "" {
-		ORFreeModels, err = fetchORModels(true)
+		_, err := fetchORModels(true)
 		if err != nil {
 			logger.Warn("failed to fetch or models", "error", err)
 		}
 	}
 	// if llama.cpp started after gf-lt?
-	localModelsMu.Lock()
+	ml, err := fetchLCPModelsWithLoadStatus()
-	LocalModels, err = fetchLCPModelsWithLoadStatus()
-	localModelsMu.Unlock()
 	if err != nil {
 		logger.Warn("failed to fetch llama.cpp models", "error", err)
 	}
+	LocalModels.Store(ml)
+	for statusLineWidget == nil {
+		time.Sleep(time.Millisecond * 100)
+	}
 	// set already loaded model in llama.cpp
 	if strings.Contains(cfg.CurrentAPI, "localhost") || strings.Contains(cfg.CurrentAPI, "127.0.0.1") {
-		localModelsMu.Lock()
+		modelList := LocalModels.Load().([]string)
-		defer localModelsMu.Unlock()
+		for i := range modelList {
-		for i := range LocalModels {
+			if strings.Contains(modelList[i], models.LoadedMark) {
-			if strings.Contains(LocalModels[i], models.LoadedMark) {
+				m := strings.TrimPrefix(modelList[i], models.LoadedMark)
-				m := strings.TrimPrefix(LocalModels[i], models.LoadedMark)
 				cfg.CurrentModel = m
 				chatBody.Model = m
-				cachedModelColor = "green"
+				cachedModelColor.Store("green")
 				updateStatusLine()
 				updateToolCapabilities()
 				app.Draw()
@@ -1419,21 +1439,17 @@ func updateModelLists() {
 }
 
 func refreshLocalModelsIfEmpty() {
-	localModelsMu.RLock()
+	models := LocalModels.Load().([]string)
-	if len(LocalModels) > 0 {
+	if len(models) > 0 {
-		localModelsMu.RUnlock()
 		return
 	}
-	localModelsMu.RUnlock()
 	// try to fetch
 	models, err := fetchLCPModels()
 	if err != nil {
 		logger.Warn("failed to fetch llama.cpp models", "error", err)
 		return
 	}
-	localModelsMu.Lock()
+	LocalModels.Store(models)
-	LocalModels = models
-	localModelsMu.Unlock()
 }
 
 func summarizeAndStartNewChat() {
@@ -1493,14 +1509,20 @@ func init() {
 	// load cards
 	basicCard.Role = cfg.AssistantRole
 	logLevel.Set(slog.LevelInfo)
-	logger = slog.New(slog.NewTextHandler(logfile, &slog.HandlerOptions{Level: logLevel}))
+	logger = slog.New(slog.NewTextHandler(logfile, &slog.HandlerOptions{Level: logLevel, AddSource: true}))
 	store = storage.NewProviderSQL(cfg.DBPATH, logger)
 	if store == nil {
 		cancel()
 		os.Exit(1)
 		return
 	}
-	ragger = rag.New(logger, store, cfg)
+	ragger, err = rag.New(logger, store, cfg)
+	if err != nil {
+		logger.Error("failed to create RAG", "error", err)
+	}
+	if ragger != nil && ragger.FallbackMessage() != "" && app != nil {
+		showToast("RAG", "ONNX unavailable, using API: "+ragger.FallbackMessage())
+	}
 	// https://github.com/coreydaley/ggerganov-llama.cpp/blob/master/examples/server/README.md
 	// load all chats in memory
 	if _, err := loadHistoryChats(); err != nil {
@@ -1511,11 +1533,11 @@ func init() {
 	}
 	lastToolCall = &models.FuncCall{}
 	lastChat := loadOldChatOrGetNew()
-	chatBody = &models.ChatBody{
+	chatBody = models.NewSafeChatBody(&models.ChatBody{
 		Model:    "modelname",
 		Stream:   true,
 		Messages: lastChat,
-	}
+	})
 	choseChunkParser()
 	httpClient = createClient(time.Second * 90)
 	if cfg.TTS_ENABLED {

config.example.toml

@@ -13,6 +13,9 @@ OpenRouterChatAPI = "https://openrouter.ai/api/v1/chat/completions"
 # embeddings
 EmbedURL = "http://localhost:8082/v1/embeddings"
 HFToken = ""
+EmbedModelPath = "onnx/embedgemma/model_q4.onnx"
+EmbedTokenizerPath = "onnx/embedgemma/tokenizer.json"
+EmbedDims = 768
 #
 ShowSys = true
 LogFile = "log.txt"
@@ -24,9 +27,9 @@ ChunkLimit = 100000
 AutoScrollEnabled = true
 AutoCleanToolCallsFromCtx = false
 # rag settings
-RAGEnabled = false
 RAGBatchSize = 1
 RAGWordLimit = 80
+RAGOverlapWords = 16
 RAGDir = "ragimport"
 # extra tts
 TTS_ENABLED = false

config/config.go

@@ -36,11 +36,14 @@ type Config struct {
 	// embeddings
 	EmbedURL           string `toml:"EmbedURL"`
 	HFToken            string `toml:"HFToken"`
+	EmbedModelPath     string `toml:"EmbedModelPath"`
+	EmbedTokenizerPath string `toml:"EmbedTokenizerPath"`
+	EmbedDims          int    `toml:"EmbedDims"`
 	// rag settings
-	RAGEnabled   bool   `toml:"RAGEnabled"`
 	RAGDir          string `toml:"RAGDir"`
 	RAGBatchSize    int    `toml:"RAGBatchSize"`
 	RAGWordLimit    uint32 `toml:"RAGWordLimit"`
+	RAGOverlapWords uint32 `toml:"RAGOverlapWords"`
 	// deepseek
 	DeepSeekChatAPI       string `toml:"DeepSeekChatAPI"`
 	DeepSeekCompletionAPI string `toml:"DeepSeekCompletionAPI"`

docs/config.md

@@ -71,9 +71,6 @@ This document explains how to set up and configure the application using the `co
 #### EmbedURL (`"http://localhost:8082/v1/embeddings"`)
 - The endpoint for embedding API, used for RAG (Retrieval Augmented Generation) functionality.
 
-#### RAGEnabled (`false`)
-- Enable or disable RAG functionality for enhanced context retrieval.
-
 #### RAGBatchSize (`1`)
 - Number of documents to process in each RAG batch.
 

go.mod

@@ -7,7 +7,6 @@ require (
 	github.com/GrailFinder/google-translate-tts v0.1.3
 	github.com/GrailFinder/searchagent v0.2.0
 	github.com/PuerkitoBio/goquery v1.11.0
-	github.com/deckarep/golang-set/v2 v2.8.0
 	github.com/gdamore/tcell/v2 v2.13.2
 	github.com/glebarez/go-sqlite v1.22.0
 	github.com/gopxl/beep/v2 v2.1.1
@@ -17,14 +16,18 @@ require (
 	github.com/neurosnap/sentences v1.1.2
 	github.com/playwright-community/playwright-go v0.5700.1
 	github.com/rivo/tview v0.42.0
+	github.com/sugarme/tokenizer v0.3.0
+	github.com/yalue/onnxruntime_go v1.27.0
 	github.com/yuin/goldmark v1.4.13
 )
 
 require (
 	github.com/andybalholm/cascadia v1.3.3 // indirect
+	github.com/deckarep/golang-set/v2 v2.8.0 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/ebitengine/oto/v3 v3.4.0 // indirect
 	github.com/ebitengine/purego v0.9.1 // indirect
+	github.com/emirpasic/gods v1.18.1 // indirect
 	github.com/gdamore/encoding v1.0.1 // indirect
 	github.com/go-jose/go-jose/v3 v3.0.4 // indirect
 	github.com/go-stack/stack v1.8.1 // indirect
@@ -33,10 +36,14 @@ require (
 	github.com/hajimehoshi/oto/v2 v2.3.1 // indirect
 	github.com/lucasb-eyer/go-colorful v1.3.0 // indirect
 	github.com/mattn/go-isatty v0.0.20 // indirect
+	github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db // indirect
 	github.com/ncruces/go-strftime v1.0.0 // indirect
+	github.com/patrickmn/go-cache v2.1.0+incompatible // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
 	github.com/rivo/uniseg v0.4.7 // indirect
+	github.com/schollz/progressbar/v2 v2.15.0 // indirect
+	github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c // indirect
 	golang.org/x/exp v0.0.0-20251209150349-8475f28825e9 // indirect
 	golang.org/x/net v0.48.0 // indirect
 	golang.org/x/sys v0.39.0 // indirect

go.sum

@@ -21,6 +21,8 @@ github.com/ebitengine/oto/v3 v3.4.0 h1:br0PgASsEWaoWn38b2Goe7m1GKFYfNgnsjSd5Gg+/
 github.com/ebitengine/oto/v3 v3.4.0/go.mod h1:IOleLVD0m+CMak3mRVwsYY8vTctQgOM0iiL6S7Ar7eI=
 github.com/ebitengine/purego v0.9.1 h1:a/k2f2HQU3Pi399RPW1MOaZyhKJL9w/xFpKAg4q1s0A=
 github.com/ebitengine/purego v0.9.1/go.mod h1:iIjxzd6CiRiOG0UyXP+V1+jWqUXVjPKLAI0mRfJZTmQ=
+github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc=
+github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/gdamore/encoding v1.0.1 h1:YzKZckdBL6jVt2Gc+5p82qhrGiqMdG/eNs6Wy0u3Uhw=
 github.com/gdamore/encoding v1.0.1/go.mod h1:0Z0cMFinngz9kS1QfMjCP8TY7em3bZYeeklsSDPivEo=
 github.com/gdamore/tcell/v2 v2.13.2 h1:5j4srfF8ow3HICOv/61/sOhQtA25qxEB2XR3Q/Bhx2g=
@@ -61,10 +63,14 @@ github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWE
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/mattn/go-sqlite3 v1.14.22 h1:2gZY6PC6kBnID23Tichd1K+Z0oS6nE/XwU+Vz/5o4kU=
 github.com/mattn/go-sqlite3 v1.14.22/go.mod h1:Uh1q+B4BYcTPb+yiD3kU8Ct7aC0hY9fxUwlHK0RXw+Y=
+github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db h1:62I3jR2EmQ4l5rM/4FEfDWcRD+abF5XlKShorW5LRoQ=
+github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db/go.mod h1:l0dey0ia/Uv7NcFFVbCLtqEBQbrT4OCwCSKTEv6enCw=
 github.com/ncruces/go-strftime v1.0.0 h1:HMFp8mLCTPp341M/ZnA4qaf7ZlsbTc+miZjCLOFAw7w=
 github.com/ncruces/go-strftime v1.0.0/go.mod h1:Fwc5htZGVVkseilnfgOVb9mKy6w1naJmn9CehxcKcls=
 github.com/neurosnap/sentences v1.1.2 h1:iphYOzx/XckXeBiLIUBkPu2EKMJ+6jDbz/sLJZ7ZoUw=
 github.com/neurosnap/sentences v1.1.2/go.mod h1:/pwU4E9XNL21ygMIkOIllv/SMy2ujHwpf8GQPu1YPbQ=
+github.com/patrickmn/go-cache v2.1.0+incompatible h1:HRMgzkcYKYpi3C8ajMPV8OFXaaRUnok+kx1WdO15EQc=
+github.com/patrickmn/go-cache v2.1.0+incompatible/go.mod h1:3Qf8kWWT7OJRJbdiICTKqZju1ZixQ/KpMGzzAfe6+WQ=
 github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
 github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/playwright-community/playwright-go v0.5700.1 h1:PNFb1byWqrTT720rEO0JL88C6Ju0EmUnR5deFLvtP/U=
@@ -77,10 +83,19 @@ github.com/rivo/tview v0.42.0 h1:b/ftp+RxtDsHSaynXTbJb+/n/BxDEi+W3UfF5jILK6c=
 github.com/rivo/tview v0.42.0/go.mod h1:cSfIYfhpSGCjp3r/ECJb+GKS7cGJnqV8vfjQPwoXyfY=
 github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=
 github.com/rivo/uniseg v0.4.7/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUcx88=
+github.com/schollz/progressbar/v2 v2.15.0 h1:dVzHQ8fHRmtPjD3K10jT3Qgn/+H+92jhPrhmxIJfDz8=
+github.com/schollz/progressbar/v2 v2.15.0/go.mod h1:UdPq3prGkfQ7MOzZKlDRpYKcFqEMczbD7YmbPgpzKMI=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
 github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.10.0 h1:Xv5erBjTwe/5IxqUQTdXv5kgmIvbHo3QQyRwhJsOfJA=
 github.com/stretchr/testify v1.10.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
+github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c h1:pwb4kNSHb4K89ymCaN+5lPH/MwnfSVg4rzGDh4d+iy4=
+github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c/go.mod h1:2gwkXLWbDGUQWeL3RtpCmcY4mzCtU13kb9UsAg9xMaw=
+github.com/sugarme/tokenizer v0.3.0 h1:FE8DYbNSz/kSbgEo9l/RjgYHkIJYEdskumitFQBE9FE=
+github.com/sugarme/tokenizer v0.3.0/go.mod h1:VJ+DLK5ZEZwzvODOWwY0cw+B1dabTd3nCB5HuFCItCc=
+github.com/yalue/onnxruntime_go v1.27.0 h1:c1YSgDNtpf0WGtxj3YeRIb8VC5LmM1J+Ve3uHdteC1U=
+github.com/yalue/onnxruntime_go v1.27.0/go.mod h1:b4X26A8pekNb1ACJ58wAXgNKeUCGEAQ9dmACut9Sm/4=
 github.com/yuin/goldmark v1.4.13 h1:fVcFKWvrslecOb/tg+Cc05dkeYx540o0FuFt3nUVDoE=
 github.com/yuin/goldmark v1.4.13/go.mod h1:6yULJ656Px+3vBD8DxQVa3kxgyrAnzto9xy5taEt/CY=
 golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=

helpfuncs.go

@@ -16,11 +16,17 @@ import (
 	"time"
 	"unicode"
 
+	"sync/atomic"
+
 	"github.com/rivo/tview"
 )
 
 // Cached model color - updated by background goroutine
-var cachedModelColor string = "orange"
+var cachedModelColor atomic.Value // stores string
+
+func init() {
+	cachedModelColor.Store("orange")
+}
 
 // startModelColorUpdater starts a background goroutine that periodically updates
 // the cached model color. Only runs HTTP requests for local llama.cpp APIs.
@@ -39,20 +45,20 @@ func startModelColorUpdater() {
 // updateCachedModelColor updates the global cachedModelColor variable
 func updateCachedModelColor() {
 	if !isLocalLlamacpp() {
-		cachedModelColor = "orange"
+		cachedModelColor.Store("orange")
 		return
 	}
 	// Check if model is loaded
-	loaded, err := isModelLoaded(chatBody.Model)
+	loaded, err := isModelLoaded(chatBody.GetModel())
 	if err != nil {
 		// On error, assume not loaded (red)
-		cachedModelColor = "red"
+		cachedModelColor.Store("red")
 		return
 	}
 	if loaded {
-		cachedModelColor = "green"
+		cachedModelColor.Store("green")
 	} else {
-		cachedModelColor = "red"
+		cachedModelColor.Store("red")
 	}
 }
 
@@ -103,7 +109,7 @@ func refreshChatDisplay() {
 		viewingAs = cfg.WriteNextMsgAs
 	}
 	// Filter messages for this character
-	filteredMessages := filterMessagesForCharacter(chatBody.Messages, viewingAs)
+	filteredMessages := filterMessagesForCharacter(chatBody.GetMessages(), viewingAs)
 	displayText := chatToText(filteredMessages, cfg.ShowSys)
 	textView.SetText(displayText)
 	colorText()
@@ -217,8 +223,8 @@ func startNewChat(keepSysP bool) {
 		logger.Warn("no such sys msg", "name", cfg.AssistantRole)
 	}
 	// set chat body
-	chatBody.Messages = chatBody.Messages[:2]
+	chatBody.TruncateMessages(2)
-	textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+	textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 	newChat := &models.Chat{
 		ID:        id + 1,
 		Name:      fmt.Sprintf("%d_%s", id+1, cfg.AssistantRole),
@@ -335,7 +341,7 @@ func isLocalLlamacpp() bool {
 // The cached value is updated by a background goroutine every 5 seconds.
 // For non-local models, returns orange. For local llama.cpp models, returns green if loaded, red if not.
 func getModelColor() string {
-	return cachedModelColor
+	return cachedModelColor.Load().(string)
 }
 
 func makeStatusLine() string {
@@ -370,7 +376,7 @@ func makeStatusLine() string {
 	// Get model color based on load status for local llama.cpp models
 	modelColor := getModelColor()
 	statusLine := fmt.Sprintf(statusLineTempl, activeChatName,
-		boolColors[cfg.ToolUse], modelColor, chatBody.Model, boolColors[cfg.SkipLLMResp],
+		boolColors[cfg.ToolUse], modelColor, chatBody.GetModel(), boolColors[cfg.SkipLLMResp],
 		cfg.CurrentAPI, persona, botPersona)
 	if cfg.STT_ENABLED {
 		recordingS := fmt.Sprintf(" | [%s:-:b]voice recording[-:-:-] (ctrl+r)",
@@ -396,11 +402,11 @@ func makeStatusLine() string {
 }
 
 func getContextTokens() int {
-	if chatBody == nil || chatBody.Messages == nil {
+	if chatBody == nil {
 		return 0
 	}
 	total := 0
-	messages := chatBody.Messages
+	messages := chatBody.GetMessages()
 	for i := range messages {
 		msg := &messages[i]
 		if msg.Stats != nil && msg.Stats.Tokens > 0 {
@@ -415,26 +421,33 @@ func getContextTokens() int {
 const deepseekContext = 128000
 
 func getMaxContextTokens() int {
-	if chatBody == nil || chatBody.Model == "" {
+	if chatBody == nil || chatBody.GetModel() == "" {
 		return 0
 	}
-	modelName := chatBody.Model
+	modelName := chatBody.GetModel()
 	switch {
 	case strings.Contains(cfg.CurrentAPI, "openrouter"):
-		if orModelsData != nil {
+		ord := orModelsData.Load()
-			for i := range orModelsData.Data {
+		if ord != nil {
-				m := &orModelsData.Data[i]
+			data := ord.(*models.ORModels)
+			if data != nil {
+				for i := range data.Data {
+					m := &data.Data[i]
 					if m.ID == modelName {
 						return m.ContextLength
 					}
 				}
 			}
+		}
 	case strings.Contains(cfg.CurrentAPI, "deepseek"):
 		return deepseekContext
 	default:
-		if localModelsData != nil {
+		lmd := localModelsData.Load()
-			for i := range localModelsData.Data {
+		if lmd != nil {
-				m := &localModelsData.Data[i]
+			data := lmd.(*models.LCPModels)
+			if data != nil {
+				for i := range data.Data {
+					m := &data.Data[i]
 					if m.ID == modelName {
 						for _, arg := range m.Status.Args {
 							if strings.HasPrefix(arg, "--ctx-size") {
@@ -463,6 +476,7 @@ func getMaxContextTokens() int {
 				}
 			}
 		}
+	}
 	return 0
 }
 
@@ -490,7 +504,7 @@ func listChatRoles() []string {
 
 func deepseekModelValidator() error {
 	if cfg.CurrentAPI == cfg.DeepSeekChatAPI || cfg.CurrentAPI == cfg.DeepSeekCompletionAPI {
-		if chatBody.Model != "deepseek-chat" && chatBody.Model != "deepseek-reasoner" {
+		if chatBody.GetModel() != "deepseek-chat" && chatBody.GetModel() != "deepseek-reasoner" {
 			showToast("bad request", "wrong deepseek model name")
 			return nil
 		}
@@ -521,7 +535,7 @@ func updateFlexLayout() {
 	if shellMode {
 		flex.AddItem(shellInput, 0, 10, false)
 	} else {
-		flex.AddItem(textArea, 0, 10, false)
+		flex.AddItem(bottomFlex, 0, 10, true)
 	}
 	if positionVisible {
 		flex.AddItem(statusLineWidget, 0, 2, false)
@@ -567,13 +581,13 @@ func executeCommandAndDisplay(cmdText string) {
 			outputContent := workingDir
 			// Add the command being executed to the chat
 			fmt.Fprintf(textView, "\n[-:-:b](%d) <%s>: [-:-:-]\n$ %s\n",
-				len(chatBody.Messages), cfg.ToolRole, cmdText)
+				chatBody.GetMessageCount(), cfg.ToolRole, cmdText)
 			fmt.Fprintf(textView, "%s\n", outputContent)
 			combinedMsg := models.RoleMsg{
 				Role:    cfg.ToolRole,
 				Content: "$ " + cmdText + "\n\n" + outputContent,
 			}
-			chatBody.Messages = append(chatBody.Messages, combinedMsg)
+			chatBody.AppendMessage(combinedMsg)
 			if scrollToEndEnabled {
 				textView.ScrollToEnd()
 			}
@@ -582,13 +596,13 @@ func executeCommandAndDisplay(cmdText string) {
 		} else {
 			outputContent := "cd: " + newDir + ": No such file or directory"
 			fmt.Fprintf(textView, "\n[-:-:b](%d) <%s>: [-:-:-]\n$ %s\n",
-				len(chatBody.Messages), cfg.ToolRole, cmdText)
+				chatBody.GetMessageCount(), cfg.ToolRole, cmdText)
 			fmt.Fprintf(textView, "[red]%s[-:-:-]\n", outputContent)
 			combinedMsg := models.RoleMsg{
 				Role:    cfg.ToolRole,
 				Content: "$ " + cmdText + "\n\n" + outputContent,
 			}
-			chatBody.Messages = append(chatBody.Messages, combinedMsg)
+			chatBody.AppendMessage(combinedMsg)
 			if scrollToEndEnabled {
 				textView.ScrollToEnd()
 			}
@@ -604,7 +618,7 @@ func executeCommandAndDisplay(cmdText string) {
 	output, err := cmd.CombinedOutput()
 	// Add the command being executed to the chat
 	fmt.Fprintf(textView, "\n[-:-:b](%d) <%s>: [-:-:-]\n$ %s\n",
-		len(chatBody.Messages), cfg.ToolRole, cmdText)
+		chatBody.GetMessageCount(), cfg.ToolRole, cmdText)
 	var outputContent string
 	if err != nil {
 		// Include both output and error
@@ -635,7 +649,7 @@ func executeCommandAndDisplay(cmdText string) {
 		Role:    cfg.ToolRole,
 		Content: combinedContent,
 	}
-	chatBody.Messages = append(chatBody.Messages, combinedMsg)
+	chatBody.AppendMessage(combinedMsg)
 	// Scroll to end and update colors
 	if scrollToEndEnabled {
 		textView.ScrollToEnd()
@@ -665,7 +679,7 @@ func performSearch(term string) {
 		searchResultLengths = nil
 		originalTextForSearch = ""
 		// Re-render text without highlights
-		textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+		textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 		colorText()
 		return
 	}

llm.go

@@ -13,8 +13,9 @@ var lastImg string             // for ctrl+j
 
 // containsToolSysMsg checks if the toolSysMsg already exists in the chat body
 func containsToolSysMsg() bool {
-	for i := range chatBody.Messages {
+	messages := chatBody.GetMessages()
-		if chatBody.Messages[i].Role == cfg.ToolRole && chatBody.Messages[i].Content == toolSysMsg {
+	for i := range messages {
+		if messages[i].Role == cfg.ToolRole && messages[i].Content == toolSysMsg {
 			return true
 		}
 	}
@@ -135,13 +136,13 @@ func (lcp LCPCompletion) FormMsg(msg, role string, resume bool) (io.Reader, erro
 			newMsg = models.RoleMsg{Role: role, Content: msg}
 		}
 		newMsg = *processMessageTag(&newMsg)
-		chatBody.Messages = append(chatBody.Messages, newMsg)
+		chatBody.AppendMessage(newMsg)
 	}
 	// sending description of the tools and how to use them
 	if cfg.ToolUse && !resume && role == cfg.UserRole && !containsToolSysMsg() {
-		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
+		chatBody.AppendMessage(models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
 	}
-	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.GetMessages())
 	// Build prompt and extract images inline as we process each message
 	messages := make([]string, len(filteredMessages))
 	for i := range filteredMessages {
@@ -183,7 +184,7 @@ func (lcp LCPCompletion) FormMsg(msg, role string, resume bool) (io.Reader, erro
 	}
 	logger.Debug("checking prompt for /completion", "tool_use", cfg.ToolUse,
 		"msg", msg, "resume", resume, "prompt", prompt, "multimodal_data_count", len(multimodalData))
-	payload := models.NewLCPReq(prompt, chatBody.Model, multimodalData,
+	payload := models.NewLCPReq(prompt, chatBody.GetModel(), multimodalData,
 		defaultLCPProps, chatBody.MakeStopSliceExcluding("", listChatRoles()))
 	data, err := json.Marshal(payload)
 	if err != nil {
@@ -289,17 +290,17 @@ func (op LCPChat) FormMsg(msg, role string, resume bool) (io.Reader, error) {
 			newMsg = models.NewRoleMsg(role, msg)
 		}
 		newMsg = *processMessageTag(&newMsg)
-		chatBody.Messages = append(chatBody.Messages, newMsg)
+		chatBody.AppendMessage(newMsg)
 		logger.Debug("LCPChat FormMsg: added message to chatBody", "role", newMsg.Role,
-			"content_len", len(newMsg.Content), "message_count_after_add", len(chatBody.Messages))
+			"content_len", len(newMsg.Content), "message_count_after_add", chatBody.GetMessageCount())
 	}
-	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.GetMessages())
 	// openai /v1/chat does not support custom roles; needs to be user, assistant, system
 	// Add persona suffix to the last user message to indicate who the assistant should reply as
 	bodyCopy := &models.ChatBody{
 		Messages: make([]models.RoleMsg, len(filteredMessages)),
-		Model:    chatBody.Model,
+		Model:    chatBody.GetModel(),
-		Stream:   chatBody.Stream,
+		Stream:   chatBody.GetStream(),
 	}
 	for i := range filteredMessages {
 		strippedMsg := *stripThinkingFromMsg(&filteredMessages[i])
@@ -375,13 +376,13 @@ func (ds DeepSeekerCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 	if msg != "" { // otherwise let the bot to continue
 		newMsg := models.RoleMsg{Role: role, Content: msg}
 		newMsg = *processMessageTag(&newMsg)
-		chatBody.Messages = append(chatBody.Messages, newMsg)
+		chatBody.AppendMessage(newMsg)
 	}
 	// sending description of the tools and how to use them
 	if cfg.ToolUse && !resume && role == cfg.UserRole && !containsToolSysMsg() {
-		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
+		chatBody.AppendMessage(models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
 	}
-	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.GetMessages())
 	messages := make([]string, len(filteredMessages))
 	for i := range filteredMessages {
 		messages[i] = stripThinkingFromMsg(&filteredMessages[i]).ToPrompt()
@@ -394,7 +395,7 @@ func (ds DeepSeekerCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 	}
 	logger.Debug("checking prompt for /completion", "tool_use", cfg.ToolUse,
 		"msg", msg, "resume", resume, "prompt", prompt)
-	payload := models.NewDSCompletionReq(prompt, chatBody.Model,
+	payload := models.NewDSCompletionReq(prompt, chatBody.GetModel(),
 		defaultLCPProps["temp"],
 		chatBody.MakeStopSliceExcluding("", listChatRoles()))
 	data, err := json.Marshal(payload)
@@ -448,15 +449,15 @@ func (ds DeepSeekerChat) FormMsg(msg, role string, resume bool) (io.Reader, erro
 	if msg != "" { // otherwise let the bot continue
 		newMsg := models.RoleMsg{Role: role, Content: msg}
 		newMsg = *processMessageTag(&newMsg)
-		chatBody.Messages = append(chatBody.Messages, newMsg)
+		chatBody.AppendMessage(newMsg)
 	}
 	// Create copy of chat body with standardized user role
-	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.GetMessages())
 	// Add persona suffix to the last user message to indicate who the assistant should reply as
 	bodyCopy := &models.ChatBody{
 		Messages: make([]models.RoleMsg, len(filteredMessages)),
-		Model:    chatBody.Model,
+		Model:    chatBody.GetModel(),
-		Stream:   chatBody.Stream,
+		Stream:   chatBody.GetStream(),
 	}
 	for i := range filteredMessages {
 		strippedMsg := *stripThinkingFromMsg(&filteredMessages[i])
@@ -527,13 +528,13 @@ func (or OpenRouterCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 	if msg != "" { // otherwise let the bot to continue
 		newMsg := models.RoleMsg{Role: role, Content: msg}
 		newMsg = *processMessageTag(&newMsg)
-		chatBody.Messages = append(chatBody.Messages, newMsg)
+		chatBody.AppendMessage(newMsg)
 	}
 	// sending description of the tools and how to use them
 	if cfg.ToolUse && !resume && role == cfg.UserRole && !containsToolSysMsg() {
-		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
+		chatBody.AppendMessage(models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
 	}
-	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, botPersona := filterMessagesForCurrentCharacter(chatBody.GetMessages())
 	messages := make([]string, len(filteredMessages))
 	for i := range filteredMessages {
 		messages[i] = stripThinkingFromMsg(&filteredMessages[i]).ToPrompt()
@@ -547,7 +548,7 @@ func (or OpenRouterCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 	stopSlice := chatBody.MakeStopSliceExcluding("", listChatRoles())
 	logger.Debug("checking prompt for /completion", "tool_use", cfg.ToolUse,
 		"msg", msg, "resume", resume, "prompt", prompt, "stop_strings", stopSlice)
-	payload := models.NewOpenRouterCompletionReq(chatBody.Model, prompt,
+	payload := models.NewOpenRouterCompletionReq(chatBody.GetModel(), prompt,
 		defaultLCPProps, stopSlice)
 	data, err := json.Marshal(payload)
 	if err != nil {
@@ -633,15 +634,15 @@ func (or OpenRouterChat) FormMsg(msg, role string, resume bool) (io.Reader, erro
 			newMsg = models.NewRoleMsg(role, msg)
 		}
 		newMsg = *processMessageTag(&newMsg)
-		chatBody.Messages = append(chatBody.Messages, newMsg)
+		chatBody.AppendMessage(newMsg)
 	}
 	// Create copy of chat body with standardized user role
-	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
+	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.GetMessages())
 	// Add persona suffix to the last user message to indicate who the assistant should reply as
 	bodyCopy := &models.ChatBody{
 		Messages: make([]models.RoleMsg, len(filteredMessages)),
-		Model:    chatBody.Model,
+		Model:    chatBody.GetModel(),
-		Stream:   chatBody.Stream,
+		Stream:   chatBody.GetStream(),
 	}
 	for i := range filteredMessages {
 		strippedMsg := *stripThinkingFromMsg(&filteredMessages[i])

models/models.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"os"
 	"strings"
+	"sync"
 )
 
 type FuncCall struct {
@@ -639,3 +640,253 @@ type MultimodalToolResp struct {
 	Type  string              `json:"type"`
 	Parts []map[string]string `json:"parts"`
 }
+
+// SafeChatBody is a thread-safe wrapper around ChatBody using RWMutex.
+// This allows safe concurrent access to chat state from multiple goroutines.
+type SafeChatBody struct {
+	mu sync.RWMutex
+	ChatBody
+}
+
+// NewSafeChatBody creates a new SafeChatBody from an existing ChatBody.
+// If cb is nil, creates an empty ChatBody.
+func NewSafeChatBody(cb *ChatBody) *SafeChatBody {
+	if cb == nil {
+		return &SafeChatBody{
+			ChatBody: ChatBody{
+				Messages: []RoleMsg{},
+			},
+		}
+	}
+	return &SafeChatBody{
+		ChatBody: *cb,
+	}
+}
+
+// GetModel returns the model name (thread-safe read).
+func (s *SafeChatBody) GetModel() string {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	return s.Model
+}
+
+// SetModel sets the model name (thread-safe write).
+func (s *SafeChatBody) SetModel(model string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.Model = model
+}
+
+// GetStream returns the stream flag (thread-safe read).
+func (s *SafeChatBody) GetStream() bool {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	return s.Stream
+}
+
+// SetStream sets the stream flag (thread-safe write).
+func (s *SafeChatBody) SetStream(stream bool) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.Stream = stream
+}
+
+// GetMessages returns a copy of all messages (thread-safe read).
+// Returns a copy to prevent race conditions after the lock is released.
+func (s *SafeChatBody) GetMessages() []RoleMsg {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	// Return a copy to prevent external modification
+	messagesCopy := make([]RoleMsg, len(s.Messages))
+	copy(messagesCopy, s.Messages)
+	return messagesCopy
+}
+
+// SetMessages replaces all messages (thread-safe write).
+func (s *SafeChatBody) SetMessages(messages []RoleMsg) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.Messages = messages
+}
+
+// AppendMessage adds a message to the end (thread-safe write).
+func (s *SafeChatBody) AppendMessage(msg RoleMsg) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.Messages = append(s.Messages, msg)
+}
+
+// GetMessageAt returns a message at a specific index (thread-safe read).
+// Returns the message and a boolean indicating if the index was valid.
+func (s *SafeChatBody) GetMessageAt(index int) (RoleMsg, bool) {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	if index < 0 || index >= len(s.Messages) {
+		return RoleMsg{}, false
+	}
+	return s.Messages[index], true
+}
+
+// SetMessageAt updates a message at a specific index (thread-safe write).
+// Returns false if index is out of bounds.
+func (s *SafeChatBody) SetMessageAt(index int, msg RoleMsg) bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if index < 0 || index >= len(s.Messages) {
+		return false
+	}
+	s.Messages[index] = msg
+	return true
+}
+
+// GetLastMessage returns the last message (thread-safe read).
+// Returns the message and a boolean indicating if the chat has messages.
+func (s *SafeChatBody) GetLastMessage() (RoleMsg, bool) {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	if len(s.Messages) == 0 {
+		return RoleMsg{}, false
+	}
+	return s.Messages[len(s.Messages)-1], true
+}
+
+// GetMessageCount returns the number of messages (thread-safe read).
+func (s *SafeChatBody) GetMessageCount() int {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	return len(s.Messages)
+}
+
+// RemoveLastMessage removes the last message (thread-safe write).
+// Returns false if there are no messages.
+func (s *SafeChatBody) RemoveLastMessage() bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if len(s.Messages) == 0 {
+		return false
+	}
+	s.Messages = s.Messages[:len(s.Messages)-1]
+	return true
+}
+
+// TruncateMessages keeps only the first n messages (thread-safe write).
+func (s *SafeChatBody) TruncateMessages(n int) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if n < len(s.Messages) {
+		s.Messages = s.Messages[:n]
+	}
+}
+
+// ClearMessages removes all messages (thread-safe write).
+func (s *SafeChatBody) ClearMessages() {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.Messages = []RoleMsg{}
+}
+
+// Rename renames all occurrences of oldname to newname in messages (thread-safe read-modify-write).
+func (s *SafeChatBody) Rename(oldname, newname string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for i := range s.Messages {
+		s.Messages[i].Content = strings.ReplaceAll(s.Messages[i].Content, oldname, newname)
+		s.Messages[i].Role = strings.ReplaceAll(s.Messages[i].Role, oldname, newname)
+	}
+}
+
+// ListRoles returns all unique roles in messages (thread-safe read).
+func (s *SafeChatBody) ListRoles() []string {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	namesMap := make(map[string]struct{})
+	for i := range s.Messages {
+		namesMap[s.Messages[i].Role] = struct{}{}
+	}
+	resp := make([]string, len(namesMap))
+	i := 0
+	for k := range namesMap {
+		resp[i] = k
+		i++
+	}
+	return resp
+}
+
+// MakeStopSlice returns stop strings for all roles (thread-safe read).
+func (s *SafeChatBody) MakeStopSlice() []string {
+	return s.MakeStopSliceExcluding("", s.ListRoles())
+}
+
+// MakeStopSliceExcluding returns stop strings excluding a specific role (thread-safe read).
+func (s *SafeChatBody) MakeStopSliceExcluding(excludeRole string, roleList []string) []string {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	ss := []string{}
+	for _, role := range roleList {
+		if role == excludeRole {
+			continue
+		}
+		ss = append(ss,
+			role+":\n",
+			role+":",
+			role+": ",
+			role+":  ",
+			role+":  \n",
+			role+":   ",
+		)
+	}
+	return ss
+}
+
+// UpdateMessageFunc updates a message at index using a provided function.
+// The function receives the current message and returns the updated message.
+// This is atomic and thread-safe (read-modify-write under single lock).
+// Returns false if index is out of bounds.
+func (s *SafeChatBody) UpdateMessageFunc(index int, updater func(RoleMsg) RoleMsg) bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if index < 0 || index >= len(s.Messages) {
+		return false
+	}
+	s.Messages[index] = updater(s.Messages[index])
+	return true
+}
+
+// AppendMessageFunc appends a new message created by a provided function.
+// The function receives the current message count and returns the new message.
+// This is atomic and thread-safe.
+func (s *SafeChatBody) AppendMessageFunc(creator func(count int) RoleMsg) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	msg := creator(len(s.Messages))
+	s.Messages = append(s.Messages, msg)
+}
+
+// GetMessagesForLLM returns a filtered copy of messages for sending to LLM.
+// This is thread-safe and returns a copy safe for external modification.
+func (s *SafeChatBody) GetMessagesForLLM(filterFunc func([]RoleMsg) []RoleMsg) []RoleMsg {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	if filterFunc == nil {
+		messagesCopy := make([]RoleMsg, len(s.Messages))
+		copy(messagesCopy, s.Messages)
+		return messagesCopy
+	}
+	return filterFunc(s.Messages)
+}
+
+// WithLock executes a function while holding the write lock.
+// Use this for complex operations that need to be atomic.
+func (s *SafeChatBody) WithLock(fn func(*ChatBody)) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	fn(&s.ChatBody)
+}
+
+// WithRLock executes a function while holding the read lock.
+// Use this for complex read-only operations.
+func (s *SafeChatBody) WithRLock(fn func(*ChatBody)) {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	fn(&s.ChatBody)
+}

popups.go

@@ -22,7 +22,7 @@ func showModelSelectionPopup() {
 		models, err := fetchLCPModelsWithLoadStatus()
 		if err != nil {
 			logger.Error("failed to fetch models with load status", "error", err)
-			return LocalModels
+			return LocalModels.Load().([]string)
 		}
 		return models
 	}
@@ -30,7 +30,8 @@ func showModelSelectionPopup() {
 	modelList := getModelListForAPI(cfg.CurrentAPI)
 	// Check for empty options list
 	if len(modelList) == 0 {
-		logger.Warn("empty model list for", "api", cfg.CurrentAPI, "localModelsLen", len(LocalModels), "orModelsLen", len(ORFreeModels))
+		localModels := LocalModels.Load().([]string)
+		logger.Warn("empty model list for", "api", cfg.CurrentAPI, "localModelsLen", len(localModels), "orModelsLen", len(ORFreeModels))
 		var message string
 		switch {
 		case strings.Contains(cfg.CurrentAPI, "openrouter.ai"):
@@ -50,7 +51,7 @@ func showModelSelectionPopup() {
 	// Find the current model index to set as selected
 	currentModelIndex := -1
 	for i, model := range modelList {
-		if strings.TrimPrefix(model, models.LoadedMark) == chatBody.Model {
+		if strings.TrimPrefix(model, models.LoadedMark) == chatBody.GetModel() {
 			currentModelIndex = i
 		}
 		modelListWidget.AddItem(model, "", 0, nil)
@@ -61,8 +62,8 @@ func showModelSelectionPopup() {
 	}
 	modelListWidget.SetSelectedFunc(func(index int, mainText string, secondaryText string, shortcut rune) {
 		modelName := strings.TrimPrefix(mainText, models.LoadedMark)
-		chatBody.Model = modelName
+		chatBody.SetModel(modelName)
-		cfg.CurrentModel = chatBody.Model
+		cfg.CurrentModel = chatBody.GetModel()
 		pages.RemovePage("modelSelectionPopup")
 		app.SetFocus(textArea)
 		updateCachedModelColor()
@@ -150,15 +151,13 @@ func showAPILinkSelectionPopup() {
 			}
 			// Assume local llama.cpp
 			refreshLocalModelsIfEmpty()
-			localModelsMu.RLock()
+			return LocalModels.Load().([]string)
-			defer localModelsMu.RUnlock()
-			return LocalModels
 		}
 		newModelList := getModelListForAPI(cfg.CurrentAPI)
 		// Ensure chatBody.Model is in the new list; if not, set to first available model
-		if len(newModelList) > 0 && !slices.Contains(newModelList, chatBody.Model) {
+		if len(newModelList) > 0 && !slices.Contains(newModelList, chatBody.GetModel()) {
-			chatBody.Model = strings.TrimPrefix(newModelList[0], models.LoadedMark)
+			chatBody.SetModel(strings.TrimPrefix(newModelList[0], models.LoadedMark))
-			cfg.CurrentModel = chatBody.Model
+			cfg.CurrentModel = chatBody.GetModel()
 			updateToolCapabilities()
 		}
 		pages.RemovePage("apiLinkSelectionPopup")
@@ -229,7 +228,7 @@ func showUserRoleSelectionPopup() {
 		// Update the user role in config
 		cfg.WriteNextMsgAs = mainText
 		// role got switch, update textview with character specific context for user
-		filtered := filterMessagesForCharacter(chatBody.Messages, mainText)
+		filtered := filterMessagesForCharacter(chatBody.GetMessages(), mainText)
 		textView.SetText(chatToText(filtered, cfg.ShowSys))
 		// Remove the popup page
 		pages.RemovePage("userRoleSelectionPopup")

props_table.go

@@ -4,14 +4,11 @@ import (
 	"fmt"
 	"strconv"
 	"strings"
-	"sync"
 
 	"github.com/gdamore/tcell/v2"
 	"github.com/rivo/tview"
 )
 
-var _ = sync.RWMutex{}
-
 // Define constants for cell types
 const (
 	CellTypeCheckbox  = "checkbox"
@@ -115,9 +112,6 @@ func makePropsTable(props map[string]float32) *tview.Table {
 		row++
 	}
 	// Add checkboxes
-	addCheckboxRow("RAG use", cfg.RAGEnabled, func(checked bool) {
-		cfg.RAGEnabled = checked
-	})
 	addCheckboxRow("Inject role", injectRole, func(checked bool) {
 		injectRole = checked
 	})
@@ -160,9 +154,7 @@ func makePropsTable(props map[string]float32) *tview.Table {
 		}
 		// Assume local llama.cpp
 		refreshLocalModelsIfEmpty()
-		localModelsMu.RLock()
+		return LocalModels.Load().([]string)
-		defer localModelsMu.RUnlock()
-		return LocalModels
 	}
 	// Add input fields
 	addInputRow("New char to write msg as", "", func(text string) {
@@ -265,7 +257,8 @@ func makePropsTable(props map[string]float32) *tview.Table {
 
 				// Check for empty options list
 				if len(data.Options) == 0 {
-					logger.Warn("empty options list for", "label", label, "api", cfg.CurrentAPI, "localModelsLen", len(LocalModels), "orModelsLen", len(ORFreeModels))
+					localModels := LocalModels.Load().([]string)
+					logger.Warn("empty options list for", "label", label, "api", cfg.CurrentAPI, "localModelsLen", len(localModels), "orModelsLen", len(ORFreeModels))
 					message := "No options available for " + label
 					if label == "Select a model" {
 						switch {

rag/embedder.go

@@ -9,6 +9,13 @@ import (
 	"gf-lt/models"
 	"log/slog"
 	"net/http"
+	"os"
+	"sync"
+	"time"
+
+	"github.com/sugarme/tokenizer"
+	"github.com/sugarme/tokenizer/pretrained"
+	"github.com/yalue/onnxruntime_go"
 )
 
 // Embedder defines the interface for embedding text
@@ -27,7 +34,9 @@ type APIEmbedder struct {
 func NewAPIEmbedder(l *slog.Logger, cfg *config.Config) *APIEmbedder {
 	return &APIEmbedder{
 		logger: l,
-		client: &http.Client{},
+		client: &http.Client{
+			Timeout: 30 * time.Second,
+		},
 		cfg: cfg,
 	}
 }
@@ -134,11 +143,302 @@ func (a *APIEmbedder) EmbedSlice(lines []string) ([][]float32, error) {
 	return embeddings, nil
 }
 
-// TODO: ONNXEmbedder implementation would go here
-// This would require:
 // 1. Loading ONNX models locally
 // 2. Using a Go ONNX runtime (like gorgonia/onnx or similar)
 // 3. Converting text to embeddings without external API calls
-//
+type ONNXEmbedder struct {
-// For now, we'll focus on the API implementation which is already working in the current system,
+	session       *onnxruntime_go.DynamicAdvancedSession
-// and can be extended later when we have ONNX runtime integration
+	tokenizer     *tokenizer.Tokenizer
+	tokenizerPath string
+	dims          int
+	logger        *slog.Logger
+	mu            sync.Mutex
+	modelPath     string
+}
+
+var onnxInitOnce sync.Once
+var onnxReady bool
+var onnxLibPath string
+var cudaLibPath string
+
+var onnxLibPaths = []string{
+	"/usr/lib/libonnxruntime.so",
+	"/usr/lib/libonnxruntime.so.1.24.2",
+	"/usr/local/lib/libonnxruntime.so",
+	"/usr/lib/x86_64-linux-gnu/libonnxruntime.so",
+	"/opt/onnxruntime/lib/libonnxruntime.so",
+}
+
+var cudaLibPaths = []string{
+	"/usr/lib/libonnxruntime_providers_cuda.so",
+	"/usr/local/lib/libonnxruntime_providers_cuda.so",
+	"/opt/onnxruntime/lib/libonnxruntime_providers_cuda.so",
+}
+
+func findONNXLibrary() string {
+	for _, path := range onnxLibPaths {
+		if _, err := os.Stat(path); err == nil {
+			return path
+		}
+	}
+	return ""
+}
+
+func findCUDALibrary() string {
+	for _, path := range cudaLibPaths {
+		if _, err := os.Stat(path); err == nil {
+			return path
+		}
+	}
+	return ""
+}
+
+func NewONNXEmbedder(modelPath, tokenizerPath string, dims int, logger *slog.Logger) (*ONNXEmbedder, error) {
+	// Check if model and tokenizer files exist
+	if _, err := os.Stat(modelPath); err != nil {
+		return nil, fmt.Errorf("ONNX model not found: %w", err)
+	}
+	if _, err := os.Stat(tokenizerPath); err != nil {
+		return nil, fmt.Errorf("tokenizer not found: %w", err)
+	}
+
+	// Find ONNX library
+	onnxLibPath = findONNXLibrary()
+	if onnxLibPath == "" {
+		return nil, errors.New("ONNX runtime library not found in standard locations")
+	}
+
+	// Find CUDA provider library (optional)
+	cudaLibPath = findCUDALibrary()
+	if cudaLibPath == "" {
+		fmt.Println("WARNING: CUDA provider library not found, will use CPU")
+	}
+	emb := &ONNXEmbedder{
+		tokenizerPath: tokenizerPath,
+		dims:          dims,
+		logger:        logger,
+		modelPath:     modelPath,
+	}
+	return emb, nil
+}
+
+func (e *ONNXEmbedder) ensureInitialized() error {
+	if e.session != nil {
+		return nil
+	}
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	if e.session != nil {
+		return nil
+	}
+	// Load tokenizer lazily
+	if e.tokenizer == nil {
+		tok, err := pretrained.FromFile(e.tokenizerPath)
+		if err != nil {
+			return fmt.Errorf("failed to load tokenizer: %w", err)
+		}
+		e.tokenizer = tok
+	}
+	onnxInitOnce.Do(func() {
+		onnxruntime_go.SetSharedLibraryPath(onnxLibPath)
+		if err := onnxruntime_go.InitializeEnvironment(); err != nil {
+			e.logger.Error("failed to initialize ONNX runtime", "error", err)
+			onnxReady = false
+			return
+		}
+		// Register CUDA provider if available
+		if cudaLibPath != "" {
+			if err := onnxruntime_go.RegisterExecutionProviderLibrary("CUDA", cudaLibPath); err != nil {
+				e.logger.Warn("failed to register CUDA provider", "error", err)
+			}
+		}
+		onnxReady = true
+	})
+	if !onnxReady {
+		return errors.New("ONNX runtime not ready")
+	}
+	// Create session options
+	opts, err := onnxruntime_go.NewSessionOptions()
+	if err != nil {
+		return fmt.Errorf("failed to create session options: %w", err)
+	}
+	defer func() {
+		_ = opts.Destroy()
+	}()
+
+	// Try to add CUDA provider
+	useCUDA := cudaLibPath != ""
+	if useCUDA {
+		cudaOpts, err := onnxruntime_go.NewCUDAProviderOptions()
+		if err != nil {
+			e.logger.Warn("failed to create CUDA provider options, falling back to CPU", "error", err)
+			useCUDA = false
+		} else {
+			defer func() {
+				_ = cudaOpts.Destroy()
+			}()
+			if err := cudaOpts.Update(map[string]string{"device_id": "0"}); err != nil {
+				e.logger.Warn("failed to update CUDA options, falling back to CPU", "error", err)
+				useCUDA = false
+			} else if err := opts.AppendExecutionProviderCUDA(cudaOpts); err != nil {
+				e.logger.Warn("failed to append CUDA provider, falling back to CPU", "error", err)
+				useCUDA = false
+			}
+		}
+	}
+	if useCUDA {
+		e.logger.Info("Using CUDA for ONNX inference")
+	} else {
+		e.logger.Info("Using CPU for ONNX inference")
+	}
+
+	// Create session with options
+	session, err := onnxruntime_go.NewDynamicAdvancedSession(
+		e.getModelPath(),
+		[]string{"input_ids", "attention_mask"},
+		[]string{"sentence_embedding"},
+		opts,
+	)
+	if err != nil {
+		return fmt.Errorf("failed to create ONNX session: %w", err)
+	}
+	e.session = session
+	return nil
+}
+
+func (e *ONNXEmbedder) getModelPath() string {
+	return e.modelPath
+}
+
+func (e *ONNXEmbedder) Destroy() error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	if e.session != nil {
+		if err := e.session.Destroy(); err != nil {
+			return fmt.Errorf("failed to destroy ONNX session: %w", err)
+		}
+		e.session = nil
+		e.logger.Info("ONNX session destroyed, VRAM freed")
+	}
+	return nil
+}
+
+func (e *ONNXEmbedder) Embed(text string) ([]float32, error) {
+	if err := e.ensureInitialized(); err != nil {
+		return nil, err
+	}
+	// 1. Tokenize
+	encoding, err := e.tokenizer.EncodeSingle(text)
+	if err != nil {
+		return nil, fmt.Errorf("tokenization failed: %w", err)
+	}
+	// 2. Convert to int64 and create attention mask
+	ids := encoding.Ids
+	inputIDs := make([]int64, len(ids))
+	attentionMask := make([]int64, len(ids))
+	for i, id := range ids {
+		inputIDs[i] = int64(id)
+		attentionMask[i] = 1
+	}
+	// 3. Create input tensors (shape: [1, seq_len])
+	seqLen := int64(len(inputIDs))
+	inputIDsTensor, err := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(1, seqLen),
+		inputIDs,
+	)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create input_ids tensor: %w", err)
+	}
+	defer func() { _ = inputIDsTensor.Destroy() }()
+	maskTensor, err := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(1, seqLen),
+		attentionMask,
+	)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create attention_mask tensor: %w", err)
+	}
+	defer func() { _ = maskTensor.Destroy() }()
+	// 4. Create output tensor
+	outputTensor, err := onnxruntime_go.NewEmptyTensor[float32](
+		onnxruntime_go.NewShape(1, int64(e.dims)),
+	)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create output tensor: %w", err)
+	}
+	defer func() { _ = outputTensor.Destroy() }()
+	// 5. Run inference
+	err = e.session.Run(
+		[]onnxruntime_go.Value{inputIDsTensor, maskTensor},
+		[]onnxruntime_go.Value{outputTensor},
+	)
+	if err != nil {
+		return nil, fmt.Errorf("inference failed: %w", err)
+	}
+	// 6. Copy output data
+	outputData := outputTensor.GetData()
+	embedding := make([]float32, len(outputData))
+	copy(embedding, outputData)
+	return embedding, nil
+}
+
+func (e *ONNXEmbedder) EmbedSlice(texts []string) ([][]float32, error) {
+	if err := e.ensureInitialized(); err != nil {
+		return nil, err
+	}
+	encodings := make([]*tokenizer.Encoding, len(texts))
+	maxLen := 0
+	for i, txt := range texts {
+		enc, err := e.tokenizer.EncodeSingle(txt)
+		if err != nil {
+			return nil, err
+		}
+		encodings[i] = enc
+		if l := len(enc.Ids); l > maxLen {
+			maxLen = l
+		}
+	}
+	batchSize := len(texts)
+	inputIDs := make([]int64, batchSize*maxLen)
+	attentionMask := make([]int64, batchSize*maxLen)
+	for i, enc := range encodings {
+		ids := enc.Ids
+		offset := i * maxLen
+		for j, id := range ids {
+			inputIDs[offset+j] = int64(id)
+			attentionMask[offset+j] = 1
+		}
+		// Remaining positions are already zero (padding)
+	}
+	// Create tensors with shape [batchSize, maxLen]
+	inputTensor, _ := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(int64(batchSize), int64(maxLen)),
+		inputIDs,
+	)
+	defer func() { _ = inputTensor.Destroy() }()
+	maskTensor, _ := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(int64(batchSize), int64(maxLen)),
+		attentionMask,
+	)
+	defer func() { _ = maskTensor.Destroy() }()
+	outputTensor, _ := onnxruntime_go.NewEmptyTensor[float32](
+		onnxruntime_go.NewShape(int64(batchSize), int64(e.dims)),
+	)
+	defer func() { _ = outputTensor.Destroy() }()
+	err := e.session.Run(
+		[]onnxruntime_go.Value{inputTensor, maskTensor},
+		[]onnxruntime_go.Value{outputTensor},
+	)
+	if err != nil {
+		return nil, err
+	}
+	// Extract embeddings per batch item
+	data := outputTensor.GetData()
+	embeddings := make([][]float32, batchSize)
+	for i := 0; i < batchSize; i++ {
+		start := i * e.dims
+		emb := make([]float32, e.dims)
+		copy(emb, data[start:start+e.dims])
+		embeddings[i] = emb
+	}
+	return embeddings, nil
+}

rag/rag.go

@@ -1,6 +1,7 @@
 package rag
 
 import (
+	"context"
 	"errors"
 	"fmt"
 	"gf-lt/config"
@@ -9,16 +10,20 @@ import (
 	"log/slog"
 	"path"
 	"regexp"
+	"runtime"
 	"sort"
 	"strings"
 	"sync"
+	"time"
 
 	"github.com/neurosnap/sentences/english"
 )
 
+const ()
+
 var (
 	// Status messages for TUI integration
-	LongJobStatusCh     = make(chan string, 10) // Increased buffer size to prevent blocking
+	LongJobStatusCh     = make(chan string, 100) // Increased buffer size for parallel batch updates
 	FinishedRAGStatus   = "finished loading RAG file; press Enter"
 	LoadedFileRAGStatus = "loaded file"
 	ErrRAGStatus        = "some error occurred; failed to transfer data to vector db"
@@ -30,30 +35,143 @@ type RAG struct {
 	cfg         *config.Config
 	embedder    Embedder
 	storage     *VectorStorage
-	mu       sync.Mutex
+	mu          sync.RWMutex
+	idleMu      sync.Mutex
+	fallbackMsg string
+	idleTimer   *time.Timer
+	idleTimeout time.Duration
 }
 
-func New(l *slog.Logger, s storage.FullRepo, cfg *config.Config) *RAG {
+// batchTask represents a single batch to be embedded
-	// Initialize with API embedder by default, could be configurable later
+type batchTask struct {
-	embedder := NewAPIEmbedder(l, cfg)
+	batchIndex   int
+	paragraphs   []string
+	filename     string
+	totalBatches int
+}
+
+// batchResult represents the result of embedding a batch
+type batchResult struct {
+	batchIndex int
+	embeddings [][]float32
+	paragraphs []string
+	filename   string
+}
+
+// sendStatusNonBlocking sends a status message without blocking
+func (r *RAG) sendStatusNonBlocking(status string) {
+	select {
+	case LongJobStatusCh <- status:
+	default:
+		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", status)
+	}
+}
+
+func New(l *slog.Logger, s storage.FullRepo, cfg *config.Config) (*RAG, error) {
+	var embedder Embedder
+	var fallbackMsg string
+	if cfg.EmbedModelPath != "" && cfg.EmbedTokenizerPath != "" {
+		emb, err := NewONNXEmbedder(cfg.EmbedModelPath, cfg.EmbedTokenizerPath, cfg.EmbedDims, l)
+		if err != nil {
+			l.Error("failed to create ONNX embedder, falling back to API", "error", err)
+			fallbackMsg = err.Error()
+			embedder = NewAPIEmbedder(l, cfg)
+		} else {
+			embedder = emb
+			l.Info("using ONNX embedder", "model", cfg.EmbedModelPath, "dims", cfg.EmbedDims)
+		}
+	} else {
+		embedder = NewAPIEmbedder(l, cfg)
+		l.Info("using API embedder", "url", cfg.EmbedURL)
+	}
 	rag := &RAG{
 		logger:      l,
 		store:       s,
 		cfg:         cfg,
 		embedder:    embedder,
 		storage:     NewVectorStorage(l, s),
+		fallbackMsg: fallbackMsg,
+		idleTimeout: 30 * time.Second,
 	}
 
 	// Note: Vector tables are created via database migrations, not at runtime
 
-	return rag
+	return rag, nil
 }
 
-func wordCounter(sentence string) int {
+func createChunks(sentences []string, wordLimit, overlapWords uint32) []string {
-	return len(strings.Split(strings.TrimSpace(sentence), " "))
+	if len(sentences) == 0 {
+		return nil
+	}
+	if overlapWords >= wordLimit {
+		overlapWords = wordLimit / 2
+	}
+	var chunks []string
+	i := 0
+	for i < len(sentences) {
+		var chunkWords []string
+		wordCount := 0
+		j := i
+		for j < len(sentences) && wordCount <= int(wordLimit) {
+			sentence := sentences[j]
+			words := strings.Fields(sentence)
+			chunkWords = append(chunkWords, sentence)
+			wordCount += len(words)
+			j++
+			// If this sentence alone exceeds limit, still include it and stop
+			if wordCount > int(wordLimit) {
+				break
+			}
+		}
+		if len(chunkWords) == 0 {
+			break
+		}
+		chunk := strings.Join(chunkWords, " ")
+		chunks = append(chunks, chunk)
+		if j >= len(sentences) {
+			break
+		}
+		// Move i forward by skipping overlap
+		if overlapWords == 0 {
+			i = j
+			continue
+		}
+		// Calculate how many sentences to skip to achieve overlapWords
+		overlapRemaining := int(overlapWords)
+		newI := i
+		for newI < j && overlapRemaining > 0 {
+			words := len(strings.Fields(sentences[newI]))
+			overlapRemaining -= words
+			if overlapRemaining >= 0 {
+				newI++
+			}
+		}
+		if newI == i {
+			newI = j
+		}
+		i = newI
+	}
+	return chunks
+}
+
+func sanitizeFTSQuery(query string) string {
+	// Remove double quotes and other problematic characters for FTS5
+	query = strings.ReplaceAll(query, "\"", " ")
+	query = strings.ReplaceAll(query, "'", " ")
+	query = strings.ReplaceAll(query, ";", " ")
+	query = strings.ReplaceAll(query, "\\", " ")
+	query = strings.TrimSpace(query)
+	if query == "" {
+		return "*" // match all
+	}
+	return query
 }
 
 func (r *RAG) LoadRAG(fpath string) error {
+	return r.LoadRAGWithContext(context.Background(), fpath)
+}
+
+func (r *RAG) LoadRAGWithContext(ctx context.Context, fpath string) error {
 	r.mu.Lock()
 	defer r.mu.Unlock()
 	fileText, err := ExtractText(fpath)
@@ -61,11 +179,9 @@ func (r *RAG) LoadRAG(fpath string) error {
 		return err
 	}
 	r.logger.Debug("rag: loaded file", "fp", fpath)
-	select {
+
-	case LongJobStatusCh <- LoadedFileRAGStatus:
+	// Send initial status (non-blocking with retry)
-	default:
+	r.sendStatusNonBlocking(LoadedFileRAGStatus)
-		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", LoadedFileRAGStatus)
-	}
 	tokenizer, err := english.NewSentenceTokenizer(nil)
 	if err != nil {
 		return err
@@ -75,31 +191,9 @@ func (r *RAG) LoadRAG(fpath string) error {
 	for i, s := range sentences {
 		sents[i] = s.Text
 	}
-	// Group sentences into paragraphs based on word limit
+
-	paragraphs := []string{}
+	// Create chunks with overlap
-	par := strings.Builder{}
+	paragraphs := createChunks(sents, r.cfg.RAGWordLimit, r.cfg.RAGOverlapWords)
-	for i := 0; i < len(sents); i++ {
-		if strings.TrimSpace(sents[i]) != "" {
-			if par.Len() > 0 {
-				par.WriteString(" ")
-			}
-			par.WriteString(sents[i])
-		}
-		if wordCounter(par.String()) > int(r.cfg.RAGWordLimit) {
-			paragraph := strings.TrimSpace(par.String())
-			if paragraph != "" {
-				paragraphs = append(paragraphs, paragraph)
-			}
-			par.Reset()
-		}
-	}
-	// Handle any remaining content in the paragraph buffer
-	if par.Len() > 0 {
-		paragraph := strings.TrimSpace(par.String())
-		if paragraph != "" {
-			paragraphs = append(paragraphs, paragraph)
-		}
-	}
 	// Adjust batch size if needed
 	if len(paragraphs) < r.cfg.RAGBatchSize && len(paragraphs) > 0 {
 		r.cfg.RAGBatchSize = len(paragraphs)
@@ -107,15 +201,61 @@ func (r *RAG) LoadRAG(fpath string) error {
 	if len(paragraphs) == 0 {
 		return errors.New("no valid paragraphs found in file")
 	}
-	// Process paragraphs in batches synchronously
+	totalBatches := (len(paragraphs) + r.cfg.RAGBatchSize - 1) / r.cfg.RAGBatchSize
-	batchCount := 0
+	r.logger.Debug("starting parallel embedding", "total_batches", totalBatches, "batch_size", r.cfg.RAGBatchSize)
-	for i := 0; i < len(paragraphs); i += r.cfg.RAGBatchSize {
+
-		end := i + r.cfg.RAGBatchSize
+	// Determine concurrency level
+	concurrency := runtime.NumCPU()
+	if concurrency > totalBatches {
+		concurrency = totalBatches
+	}
+	if concurrency < 1 {
+		concurrency = 1
+	}
+	// If using ONNX embedder, limit concurrency to 1 due to mutex serialization
+	var isONNX bool
+	if _, isONNX = r.embedder.(*ONNXEmbedder); isONNX {
+		concurrency = 1
+	}
+	embedderType := "API"
+	if isONNX {
+		embedderType = "ONNX"
+	}
+	r.logger.Debug("parallel embedding setup",
+		"total_batches", totalBatches,
+		"concurrency", concurrency,
+		"embedder", embedderType,
+		"batch_size", r.cfg.RAGBatchSize)
+
+	// Create context with timeout (30 minutes) and cancellation for error handling
+	ctx, cancel := context.WithTimeout(ctx, 30*time.Minute)
+	defer cancel()
+
+	// Channels for task distribution and results
+	taskCh := make(chan batchTask, totalBatches)
+	resultCh := make(chan batchResult, totalBatches)
+	errorCh := make(chan error, totalBatches)
+
+	// Start worker goroutines
+	var wg sync.WaitGroup
+	for w := 0; w < concurrency; w++ {
+		wg.Add(1)
+		go r.embeddingWorker(ctx, w, taskCh, resultCh, errorCh, &wg)
+	}
+
+	// Close task channel after all tasks are sent (by separate goroutine)
+	go func() {
+		// Ensure task channel is closed when this goroutine exits
+		defer close(taskCh)
+		r.logger.Debug("task distributor started", "total_batches", totalBatches)
+		for i := 0; i < totalBatches; i++ {
+			start := i * r.cfg.RAGBatchSize
+			end := start + r.cfg.RAGBatchSize
 			if end > len(paragraphs) {
 				end = len(paragraphs)
 			}
-		batch := paragraphs[i:end]
+			batch := paragraphs[start:end]
-		batchCount++
+
 			// Filter empty paragraphs
 			nonEmptyBatch := make([]string, 0, len(batch))
 			for _, p := range batch {
@@ -123,75 +263,286 @@ func (r *RAG) LoadRAG(fpath string) error {
 					nonEmptyBatch = append(nonEmptyBatch, strings.TrimSpace(p))
 				}
 			}
-		if len(nonEmptyBatch) == 0 {
+
+			task := batchTask{
+				batchIndex:   i,
+				paragraphs:   nonEmptyBatch,
+				filename:     path.Base(fpath),
+				totalBatches: totalBatches,
+			}
+
+			select {
+			case taskCh <- task:
+				r.logger.Debug("task distributor sent batch", "batch", i, "paragraphs", len(nonEmptyBatch))
+			case <-ctx.Done():
+				r.logger.Debug("task distributor cancelled", "batches_sent", i+1, "total_batches", totalBatches)
+				return
+			}
+		}
+		r.logger.Debug("task distributor finished", "batches_sent", totalBatches)
+	}()
+
+	// Wait for workers to finish and close result channel
+	go func() {
+		wg.Wait()
+		close(resultCh)
+	}()
+
+	// Process results in order and write to database
+	nextExpectedBatch := 0
+	resultsBuffer := make(map[int]batchResult)
+	filename := path.Base(fpath)
+	batchesProcessed := 0
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+
+		case err := <-errorCh:
+			// First error from any worker, cancel everything
+			cancel()
+			r.logger.Error("embedding worker failed", "error", err)
+			r.sendStatusNonBlocking(ErrRAGStatus)
+			return fmt.Errorf("embedding failed: %w", err)
+
+		case result, ok := <-resultCh:
+			if !ok {
+				// All results processed
+				resultCh = nil
+				r.logger.Debug("result channel closed", "batches_processed", batchesProcessed, "total_batches", totalBatches)
 				continue
 			}
-		// Embed the batch
+
-		embeddings, err := r.embedder.EmbedSlice(nonEmptyBatch)
+			// Store result in buffer
-		if err != nil {
+			resultsBuffer[result.batchIndex] = result
-			r.logger.Error("failed to embed batch", "error", err, "batch", batchCount)
+
-			select {
+			// Process buffered results in order
-			case LongJobStatusCh <- ErrRAGStatus:
+			for {
-			default:
+				if res, exists := resultsBuffer[nextExpectedBatch]; exists {
-				r.logger.Warn("LongJobStatusCh channel full, dropping message")
+					// Write this batch to database
-			}
+					if err := r.writeBatchToStorage(ctx, res, filename); err != nil {
-			return fmt.Errorf("failed to embed batch %d: %w", batchCount, err)
+						cancel()
-		}
-		if len(embeddings) != len(nonEmptyBatch) {
-			err := errors.New("embedding count mismatch")
-			r.logger.Error("embedding mismatch", "expected", len(nonEmptyBatch), "got", len(embeddings))
 						return err
 					}
-		// Write vectors to storage
+
-		filename := path.Base(fpath)
+					batchesProcessed++
-		for j, text := range nonEmptyBatch {
+					// Send progress update
-			vector := models.VectorRow{
+					statusMsg := fmt.Sprintf("processed batch %d/%d", batchesProcessed, totalBatches)
-				Embeddings: embeddings[j],
+					r.sendStatusNonBlocking(statusMsg)
+
+					delete(resultsBuffer, nextExpectedBatch)
+					nextExpectedBatch++
+				} else {
+					break
+				}
+			}
+
+		default:
+			// No channels ready, check for deadlock conditions
+			if resultCh == nil && nextExpectedBatch < totalBatches {
+				// Missing batch results after result channel closed
+				r.logger.Error("missing batch results",
+					"expected", totalBatches,
+					"received", nextExpectedBatch,
+					"missing", totalBatches-nextExpectedBatch)
+
+				// Wait a short time for any delayed errors, then cancel
+				select {
+				case <-time.After(5 * time.Second):
+					cancel()
+					return fmt.Errorf("missing batch results: expected %d, got %d", totalBatches, nextExpectedBatch)
+				case <-ctx.Done():
+					return ctx.Err()
+				case err := <-errorCh:
+					cancel()
+					r.logger.Error("embedding worker failed after result channel closed", "error", err)
+					r.sendStatusNonBlocking(ErrRAGStatus)
+					return fmt.Errorf("embedding failed: %w", err)
+				}
+			}
+			// If we reach here, no deadlock yet, just busy loop prevention
+			time.Sleep(100 * time.Millisecond)
+		}
+
+		// Check if we're done
+		if resultCh == nil && nextExpectedBatch >= totalBatches {
+			r.logger.Debug("all batches processed successfully", "total", totalBatches)
+			break
+		}
+	}
+	r.logger.Debug("finished writing vectors", "batches", batchesProcessed)
+	r.resetIdleTimer()
+	r.sendStatusNonBlocking(FinishedRAGStatus)
+	return nil
+}
+
+// embeddingWorker processes batch embedding tasks
+func (r *RAG) embeddingWorker(ctx context.Context, workerID int, taskCh <-chan batchTask, resultCh chan<- batchResult, errorCh chan<- error, wg *sync.WaitGroup) {
+	defer wg.Done()
+	r.logger.Debug("embedding worker started", "worker", workerID)
+
+	// Panic recovery to ensure worker doesn't crash silently
+	defer func() {
+		if rec := recover(); rec != nil {
+			r.logger.Error("embedding worker panicked", "worker", workerID, "panic", rec)
+			// Try to send error, but don't block if channel is full
+			select {
+			case errorCh <- fmt.Errorf("worker %d panicked: %v", workerID, rec):
+			default:
+				r.logger.Warn("error channel full, dropping panic error", "worker", workerID)
+			}
+		}
+	}()
+	for task := range taskCh {
+		select {
+		case <-ctx.Done():
+			r.logger.Debug("embedding worker cancelled", "worker", workerID)
+			return
+		default:
+		}
+		r.logger.Debug("worker processing batch", "worker", workerID, "batch", task.batchIndex, "paragraphs", len(task.paragraphs), "total_batches", task.totalBatches)
+
+		// Skip empty batches
+		if len(task.paragraphs) == 0 {
+			select {
+			case resultCh <- batchResult{
+				batchIndex: task.batchIndex,
+				embeddings: nil,
+				paragraphs: nil,
+				filename:   task.filename,
+			}:
+			case <-ctx.Done():
+				r.logger.Debug("embedding worker cancelled while sending empty batch", "worker", workerID)
+				return
+			}
+			r.logger.Debug("worker sent empty batch", "worker", workerID, "batch", task.batchIndex)
+			continue
+		}
+		// Embed with retry for API embedder
+		embeddings, err := r.embedWithRetry(ctx, task.paragraphs, 3)
+		if err != nil {
+			// Try to send error, but don't block indefinitely
+			select {
+			case errorCh <- fmt.Errorf("worker %d batch %d: %w", workerID, task.batchIndex, err):
+			case <-ctx.Done():
+				r.logger.Debug("embedding worker cancelled while sending error", "worker", workerID)
+			}
+			return
+		}
+		// Send result with context awareness
+		select {
+		case resultCh <- batchResult{
+			batchIndex: task.batchIndex,
+			embeddings: embeddings,
+			paragraphs: task.paragraphs,
+			filename:   task.filename,
+		}:
+		case <-ctx.Done():
+			r.logger.Debug("embedding worker cancelled while sending result", "worker", workerID)
+			return
+		}
+		r.logger.Debug("worker completed batch", "worker", workerID, "batch", task.batchIndex, "embeddings", len(embeddings))
+	}
+	r.logger.Debug("embedding worker finished", "worker", workerID)
+}
+
+// embedWithRetry attempts embedding with exponential backoff for API embedder
+func (r *RAG) embedWithRetry(ctx context.Context, paragraphs []string, maxRetries int) ([][]float32, error) {
+	var lastErr error
+	for attempt := 0; attempt < maxRetries; attempt++ {
+		if attempt > 0 {
+			// Exponential backoff
+			backoff := time.Duration(attempt*attempt) * time.Second
+			if backoff > 10*time.Second {
+				backoff = 10 * time.Second
+			}
+			select {
+			case <-time.After(backoff):
+			case <-ctx.Done():
+				return nil, ctx.Err()
+			}
+			r.logger.Debug("retrying embedding", "attempt", attempt, "max_retries", maxRetries)
+		}
+
+		embeddings, err := r.embedder.EmbedSlice(paragraphs)
+		if err == nil {
+			// Validate embedding count
+			if len(embeddings) != len(paragraphs) {
+				return nil, fmt.Errorf("embedding count mismatch: expected %d, got %d", len(paragraphs), len(embeddings))
+			}
+			return embeddings, nil
+		}
+
+		lastErr = err
+		// Only retry for API embedder errors (network/timeout)
+		// For ONNX embedder, fail fast
+		if _, isAPI := r.embedder.(*APIEmbedder); !isAPI {
+			break
+		}
+	}
+	return nil, fmt.Errorf("embedding failed after %d attempts: %w", maxRetries, lastErr)
+}
+
+// writeBatchToStorage writes a single batch of vectors to the database
+func (r *RAG) writeBatchToStorage(ctx context.Context, result batchResult, filename string) error {
+	if len(result.embeddings) == 0 {
+		// Empty batch, skip
+		return nil
+	}
+	// Check context before starting
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	default:
+	}
+
+	// Build all vectors for batch write
+	vectors := make([]*models.VectorRow, 0, len(result.paragraphs))
+	for j, text := range result.paragraphs {
+		vectors = append(vectors, &models.VectorRow{
+			Embeddings: result.embeddings[j],
 			RawText:    text,
-				Slug:       fmt.Sprintf("%s_%d_%d", filename, batchCount, j),
+			Slug:       fmt.Sprintf("%s_%d_%d", filename, result.batchIndex+1, j),
 			FileName:   filename,
+		})
 	}
-			if err := r.storage.WriteVector(&vector); err != nil {
+
-				r.logger.Error("failed to write vector to DB", "error", err, "slug", vector.Slug)
+	// Write all vectors in a single transaction
-				select {
+	if err := r.storage.WriteVectors(vectors); err != nil {
-				case LongJobStatusCh <- ErrRAGStatus:
+		r.logger.Error("failed to write vectors batch to DB", "error", err, "batch", result.batchIndex+1, "size", len(vectors))
-				default:
+		r.sendStatusNonBlocking(ErrRAGStatus)
-					r.logger.Warn("LongJobStatusCh channel full, dropping message")
+		return fmt.Errorf("failed to write vectors batch: %w", err)
-				}
-				return fmt.Errorf("failed to write vector: %w", err)
-			}
-		}
-		r.logger.Debug("wrote batch to db", "batch", batchCount, "size", len(nonEmptyBatch))
-		// Send progress status
-		statusMsg := fmt.Sprintf("processed batch %d/%d", batchCount, (len(paragraphs)+r.cfg.RAGBatchSize-1)/r.cfg.RAGBatchSize)
-		select {
-		case LongJobStatusCh <- statusMsg:
-		default:
-			r.logger.Warn("LongJobStatusCh channel full, dropping message")
-		}
-	}
-	r.logger.Debug("finished writing vectors", "batches", batchCount)
-	select {
-	case LongJobStatusCh <- FinishedRAGStatus:
-	default:
-		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", FinishedRAGStatus)
 	}
+	r.logger.Debug("wrote batch to db", "batch", result.batchIndex+1, "size", len(result.paragraphs))
 	return nil
 }
 
 func (r *RAG) LineToVector(line string) ([]float32, error) {
+	r.resetIdleTimer()
 	return r.embedder.Embed(line)
 }
 
-func (r *RAG) SearchEmb(emb *models.EmbeddingResp) ([]models.VectorRow, error) {
+func (r *RAG) searchEmb(emb *models.EmbeddingResp, limit int) ([]models.VectorRow, error) {
-	return r.storage.SearchClosest(emb.Embedding)
+	r.resetIdleTimer()
+	return r.storage.SearchClosest(emb.Embedding, limit)
+}
+
+func (r *RAG) searchKeyword(query string, limit int) ([]models.VectorRow, error) {
+	r.resetIdleTimer()
+	sanitized := sanitizeFTSQuery(query)
+	return r.storage.SearchKeyword(sanitized, limit)
 }
 
 func (r *RAG) ListLoaded() ([]string, error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
 	return r.storage.ListFiles()
 }
 
 func (r *RAG) RemoveFile(filename string) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.resetIdleTimer()
 	return r.storage.RemoveEmbByFileName(filename)
 }
 
@@ -211,11 +562,14 @@ func (r *RAG) RefineQuery(query string) string {
 		return original
 	}
 	query = strings.ToLower(query)
+	words := strings.Fields(query)
+	if len(words) >= 3 {
 		for _, stopWord := range stopWords {
 			wordPattern := `\b` + stopWord + `\b`
 			re := regexp.MustCompile(wordPattern)
 			query = re.ReplaceAllString(query, "")
 		}
+	}
 	query = strings.TrimSpace(query)
 	if len(query) < 5 {
 		return original
@@ -246,7 +600,7 @@ func (r *RAG) extractImportantPhrases(query string) string {
 				break
 			}
 		}
-		if isImportant || len(word) > 3 {
+		if isImportant || len(word) >= 3 {
 			important = append(important, word)
 		}
 	}
@@ -265,6 +619,36 @@ func (r *RAG) GenerateQueryVariations(query string) []string {
 	if len(parts) == 0 {
 		return variations
 	}
+	// Get loaded filenames to filter out filename terms
+	filenames, err := r.storage.ListFiles()
+	if err == nil && len(filenames) > 0 {
+		// Convert to lowercase for case-insensitive matching
+		lowerFilenames := make([]string, len(filenames))
+		for i, f := range filenames {
+			lowerFilenames[i] = strings.ToLower(f)
+		}
+		filteredParts := make([]string, 0, len(parts))
+		for _, part := range parts {
+			partLower := strings.ToLower(part)
+			skip := false
+			for _, fn := range lowerFilenames {
+				if strings.Contains(fn, partLower) || strings.Contains(partLower, fn) {
+					skip = true
+					break
+				}
+			}
+			if !skip {
+				filteredParts = append(filteredParts, part)
+			}
+		}
+		// If filteredParts not empty and different from original, add filtered query
+		if len(filteredParts) > 0 && len(filteredParts) != len(parts) {
+			filteredQuery := strings.Join(filteredParts, " ")
+			if len(filteredQuery) >= 5 {
+				variations = append(variations, filteredQuery)
+			}
+		}
+	}
 	if len(parts) >= 2 {
 		trimmed := strings.Join(parts[:len(parts)-1], " ")
 		if len(trimmed) >= 5 {
@@ -328,9 +712,14 @@ func (r *RAG) RerankResults(results []models.VectorRow, query string) []models.V
 	})
 	unique := make([]models.VectorRow, 0)
 	seen := make(map[string]bool)
+	fileCounts := make(map[string]int)
 	for i := range scored {
 		if !seen[scored[i].row.Slug] {
+			if fileCounts[scored[i].row.FileName] >= 2 {
+				continue
+			}
 			seen[scored[i].row.Slug] = true
+			fileCounts[scored[i].row.FileName]++
 			unique = append(unique, scored[i].row)
 		}
 	}
@@ -341,6 +730,9 @@ func (r *RAG) RerankResults(results []models.VectorRow, query string) []models.V
 }
 
 func (r *RAG) SynthesizeAnswer(results []models.VectorRow, query string) (string, error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	r.resetIdleTimer()
 	if len(results) == 0 {
 		return "No relevant information found in the vector database.", nil
 	}
@@ -369,7 +761,7 @@ func (r *RAG) SynthesizeAnswer(results []models.VectorRow, query string) (string
 		Embedding: emb,
 		Index:     0,
 	}
-	topResults, err := r.SearchEmb(embResp)
+	topResults, err := r.searchEmb(embResp, 1)
 	if err != nil {
 		r.logger.Error("failed to search for synthesis context", "error", err)
 		return "", err
@@ -396,9 +788,14 @@ func truncateString(s string, maxLen int) string {
 }
 
 func (r *RAG) Search(query string, limit int) ([]models.VectorRow, error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	r.resetIdleTimer()
 	refined := r.RefineQuery(query)
 	variations := r.GenerateQueryVariations(refined)
-	allResults := make([]models.VectorRow, 0)
+
+	// Collect embedding search results from all variations
+	var embResults []models.VectorRow
 	seen := make(map[string]bool)
 	for _, q := range variations {
 		emb, err := r.LineToVector(q)
@@ -406,29 +803,78 @@ func (r *RAG) Search(query string, limit int) ([]models.VectorRow, error) {
 			r.logger.Error("failed to embed query variation", "error", err, "query", q)
 			continue
 		}
-
 		embResp := &models.EmbeddingResp{
 			Embedding: emb,
 			Index:     0,
 		}
-
+		results, err := r.searchEmb(embResp, limit*2) // Get more candidates
-		results, err := r.SearchEmb(embResp)
 		if err != nil {
 			r.logger.Error("failed to search embeddings", "error", err, "query", q)
 			continue
 		}
-
 		for _, row := range results {
 			if !seen[row.Slug] {
 				seen[row.Slug] = true
-				allResults = append(allResults, row)
+				embResults = append(embResults, row)
 			}
 		}
 	}
-	reranked := r.RerankResults(allResults, query)
+	// Sort embedding results by distance (lower is better)
-	if len(reranked) > limit {
+	sort.Slice(embResults, func(i, j int) bool {
-		reranked = reranked[:limit]
+		return embResults[i].Distance < embResults[j].Distance
+	})
+
+	// Perform keyword search
+	kwResults, err := r.searchKeyword(refined, limit*2)
+	if err != nil {
+		r.logger.Warn("keyword search failed, using only embeddings", "error", err)
+		kwResults = nil
 	}
+	// Sort keyword results by distance (already sorted by BM25 score)
+	// kwResults already sorted by distance (lower is better)
+
+	// Combine using Reciprocal Rank Fusion (RRF)
+	const rrfK = 60
+	type scoredRow struct {
+		row   models.VectorRow
+		score float64
+	}
+	scoreMap := make(map[string]float64)
+	// Add embedding results
+	for rank, row := range embResults {
+		score := 1.0 / (float64(rank) + rrfK)
+		scoreMap[row.Slug] += score
+	}
+	// Add keyword results
+	for rank, row := range kwResults {
+		score := 1.0 / (float64(rank) + rrfK)
+		scoreMap[row.Slug] += score
+		// Ensure row exists in combined results
+		if _, exists := seen[row.Slug]; !exists {
+			embResults = append(embResults, row)
+		}
+	}
+	// Create slice of scored rows
+	scoredRows := make([]scoredRow, 0, len(embResults))
+	for _, row := range embResults {
+		score := scoreMap[row.Slug]
+		scoredRows = append(scoredRows, scoredRow{row: row, score: score})
+	}
+	// Sort by descending RRF score
+	sort.Slice(scoredRows, func(i, j int) bool {
+		return scoredRows[i].score > scoredRows[j].score
+	})
+	// Take top limit
+	if len(scoredRows) > limit {
+		scoredRows = scoredRows[:limit]
+	}
+	// Convert back to VectorRow
+	finalResults := make([]models.VectorRow, len(scoredRows))
+	for i, sr := range scoredRows {
+		finalResults[i] = sr.row
+	}
+	// Apply reranking heuristics
+	reranked := r.RerankResults(finalResults, query)
 	return reranked, nil
 }
 
@@ -437,16 +883,58 @@ var (
 	ragOnce     sync.Once
 )
 
+func (r *RAG) FallbackMessage() string {
+	return r.fallbackMsg
+}
+
 func Init(c *config.Config, l *slog.Logger, s storage.FullRepo) error {
+	var err error
 	ragOnce.Do(func() {
 		if c == nil || l == nil || s == nil {
 			return
 		}
-		ragInstance = New(l, s, c)
+		ragInstance, err = New(l, s, c)
 	})
-	return nil
+	return err
 }
 
 func GetInstance() *RAG {
 	return ragInstance
 }
+
+func (r *RAG) resetIdleTimer() {
+	r.idleMu.Lock()
+	defer r.idleMu.Unlock()
+	if r.idleTimer != nil {
+		r.idleTimer.Stop()
+	}
+	r.idleTimer = time.AfterFunc(r.idleTimeout, func() {
+		r.freeONNXMemory()
+	})
+}
+
+func (r *RAG) freeONNXMemory() {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if onnx, ok := r.embedder.(*ONNXEmbedder); ok {
+		if err := onnx.Destroy(); err != nil {
+			r.logger.Error("failed to free ONNX memory", "error", err)
+		} else {
+			r.logger.Info("freed ONNX VRAM after idle timeout")
+		}
+	}
+}
+
+func (r *RAG) Destroy() {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if r.idleTimer != nil {
+		r.idleTimer.Stop()
+		r.idleTimer = nil
+	}
+	if onnx, ok := r.embedder.(*ONNXEmbedder); ok {
+		if err := onnx.Destroy(); err != nil {
+			r.logger.Error("failed to destroy ONNX embedder", "error", err)
+		}
+	}
+}

rag/storage.go

@@ -1,6 +1,7 @@
 package rag
 
 import (
+	"database/sql"
 	"encoding/binary"
 	"fmt"
 	"gf-lt/models"
@@ -62,6 +63,17 @@ func (vs *VectorStorage) WriteVector(row *models.VectorRow) error {
 	if err != nil {
 		return err
 	}
+	embeddingSize := len(row.Embeddings)
+	// Start transaction
+	tx, err := vs.sqlxDB.Beginx()
+	if err != nil {
+		return err
+	}
+	defer func() {
+		if err != nil {
+			_ = tx.Rollback()
+		}
+	}()
 
 	// Serialize the embeddings to binary
 	serializedEmbeddings := SerializeVector(row.Embeddings)
@@ -69,10 +81,102 @@ func (vs *VectorStorage) WriteVector(row *models.VectorRow) error {
 		"INSERT INTO %s (embeddings, slug, raw_text, filename) VALUES (?, ?, ?, ?)",
 		tableName,
 	)
-	if _, err := vs.sqlxDB.Exec(query, serializedEmbeddings, row.Slug, row.RawText, row.FileName); err != nil {
+	if _, err := tx.Exec(query, serializedEmbeddings, row.Slug, row.RawText, row.FileName); err != nil {
 		vs.logger.Error("failed to write vector", "error", err, "slug", row.Slug)
 		return err
 	}
+	// Insert into FTS table
+	ftsQuery := `INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size) VALUES (?, ?, ?, ?)`
+	if _, err := tx.Exec(ftsQuery, row.Slug, row.RawText, row.FileName, embeddingSize); err != nil {
+		vs.logger.Error("failed to write to FTS table", "error", err, "slug", row.Slug)
+		return err
+	}
+	err = tx.Commit()
+	if err != nil {
+		vs.logger.Error("failed to commit transaction", "error", err)
+		return err
+	}
+	return nil
+}
+
+// WriteVectors stores multiple embedding vectors in a single transaction
+func (vs *VectorStorage) WriteVectors(rows []*models.VectorRow) error {
+	if len(rows) == 0 {
+		return nil
+	}
+	// SQLite has limit of 999 parameters per statement, each row uses 4 parameters
+	const maxBatchSize = 200 // 200 * 4 = 800 < 999
+	if len(rows) > maxBatchSize {
+		// Process in chunks
+		for i := 0; i < len(rows); i += maxBatchSize {
+			end := i + maxBatchSize
+			if end > len(rows) {
+				end = len(rows)
+			}
+			if err := vs.WriteVectors(rows[i:end]); err != nil {
+				return err
+			}
+		}
+		return nil
+	}
+	// All rows should have same embedding size (same model)
+	firstSize := len(rows[0].Embeddings)
+	for i, row := range rows {
+		if len(row.Embeddings) != firstSize {
+			return fmt.Errorf("embedding size mismatch: row %d has size %d, expected %d", i, len(row.Embeddings), firstSize)
+		}
+	}
+	tableName, err := vs.getTableName(rows[0].Embeddings)
+	if err != nil {
+		return err
+	}
+	// Start transaction
+	tx, err := vs.sqlxDB.Beginx()
+	if err != nil {
+		return err
+	}
+	defer func() {
+		if err != nil {
+			_ = tx.Rollback()
+		}
+	}()
+
+	// Build batch insert for embeddings table
+	embeddingPlaceholders := make([]string, 0, len(rows))
+	embeddingArgs := make([]any, 0, len(rows)*4)
+	for _, row := range rows {
+		embeddingPlaceholders = append(embeddingPlaceholders, "(?, ?, ?, ?)")
+		embeddingArgs = append(embeddingArgs, SerializeVector(row.Embeddings), row.Slug, row.RawText, row.FileName)
+	}
+	embeddingQuery := fmt.Sprintf(
+		"INSERT INTO %s (embeddings, slug, raw_text, filename) VALUES %s",
+		tableName,
+		strings.Join(embeddingPlaceholders, ", "),
+	)
+	if _, err := tx.Exec(embeddingQuery, embeddingArgs...); err != nil {
+		vs.logger.Error("failed to write vectors batch", "error", err, "batch_size", len(rows))
+		return err
+	}
+	// Build batch insert for FTS table
+	ftsPlaceholders := make([]string, 0, len(rows))
+	ftsArgs := make([]any, 0, len(rows)*4)
+	embeddingSize := len(rows[0].Embeddings)
+	for _, row := range rows {
+		ftsPlaceholders = append(ftsPlaceholders, "(?, ?, ?, ?)")
+		ftsArgs = append(ftsArgs, row.Slug, row.RawText, row.FileName, embeddingSize)
+	}
+	ftsQuery := "INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size) VALUES " +
+		strings.Join(ftsPlaceholders, ", ")
+	if _, err := tx.Exec(ftsQuery, ftsArgs...); err != nil {
+		vs.logger.Error("failed to write FTS batch", "error", err, "batch_size", len(rows))
+		return err
+	}
+	err = tx.Commit()
+	if err != nil {
+		vs.logger.Error("failed to commit transaction", "error", err)
+		return err
+	}
+	vs.logger.Debug("wrote vectors batch", "batch_size", len(rows))
 	return nil
 }
 
@@ -98,30 +202,25 @@ func (vs *VectorStorage) getTableName(emb []float32) (string, error) {
 }
 
 // SearchClosest finds vectors closest to the query vector using efficient cosine similarity calculation
-func (vs *VectorStorage) SearchClosest(query []float32) ([]models.VectorRow, error) {
+func (vs *VectorStorage) SearchClosest(query []float32, limit int) ([]models.VectorRow, error) {
+	if limit <= 0 {
+		limit = 10
+	}
 	tableName, err := vs.getTableName(query)
 	if err != nil {
 		return nil, err
 	}
-
-	// For better performance, instead of loading all vectors at once,
-	// we'll implement batching and potentially add L2 distance-based pre-filtering
-	// since cosine similarity is related to L2 distance for normalized vectors
-
 	querySQL := "SELECT embeddings, slug, raw_text, filename FROM " + tableName
 	rows, err := vs.sqlxDB.Query(querySQL)
 	if err != nil {
 		return nil, err
 	}
 	defer rows.Close()
-
-	// Use a min-heap or simple slice to keep track of top 3 closest vectors
 	type SearchResult struct {
 		vector   models.VectorRow
 		distance float32
 	}
 	var topResults []SearchResult
-	// Process vectors one by one to avoid loading everything into memory
 	for rows.Next() {
 		var (
 			embeddingsBlob          []byte
@@ -132,12 +231,9 @@ func (vs *VectorStorage) SearchClosest(query []float32) ([]models.VectorRow, err
 			vs.logger.Error("failed to scan row", "error", err)
 			continue
 		}
-
 		storedEmbeddings := DeserializeVector(embeddingsBlob)
-
-		// Calculate cosine similarity (returns value between -1 and 1, where 1 is most similar)
 		similarity := cosineSimilarity(query, storedEmbeddings)
-		distance := 1 - similarity // Convert to distance where 0 is most similar
+		distance := 1 - similarity
 
 		result := SearchResult{
 			vector: models.VectorRow{
@@ -149,20 +245,14 @@ func (vs *VectorStorage) SearchClosest(query []float32) ([]models.VectorRow, err
 			distance: distance,
 		}
 
-		// Add to top results and maintain only top 3
 		topResults = append(topResults, result)
-
-		// Sort and keep only top 3
 		sort.Slice(topResults, func(i, j int) bool {
 			return topResults[i].distance < topResults[j].distance
 		})
-
+		if len(topResults) > limit {
-		if len(topResults) > 3 {
+			topResults = topResults[:limit]
-			topResults = topResults[:3] // Keep only closest 3
 		}
 	}
-
-	// Convert back to VectorRow slice
 	results := make([]models.VectorRow, 0, len(topResults))
 	for _, result := range topResults {
 		result.vector.Distance = result.distance
@@ -171,6 +261,100 @@ func (vs *VectorStorage) SearchClosest(query []float32) ([]models.VectorRow, err
 	return results, nil
 }
 
+// GetVectorBySlug retrieves a vector row by its slug
+func (vs *VectorStorage) GetVectorBySlug(slug string) (*models.VectorRow, error) {
+	embeddingSizes := []int{384, 768, 1024, 1536, 2048, 3072, 4096, 5120}
+	for _, size := range embeddingSizes {
+		table := fmt.Sprintf("embeddings_%d", size)
+		query := fmt.Sprintf("SELECT embeddings, slug, raw_text, filename FROM %s WHERE slug = ?", table)
+		row := vs.sqlxDB.QueryRow(query, slug)
+		var (
+			embeddingsBlob                   []byte
+			retrievedSlug, rawText, fileName string
+		)
+		if err := row.Scan(&embeddingsBlob, &retrievedSlug, &rawText, &fileName); err != nil {
+			// No row in this table, continue to next size
+			continue
+		}
+		storedEmbeddings := DeserializeVector(embeddingsBlob)
+		return &models.VectorRow{
+			Embeddings: storedEmbeddings,
+			Slug:       retrievedSlug,
+			RawText:    rawText,
+			FileName:   fileName,
+		}, nil
+	}
+	return nil, fmt.Errorf("vector with slug %s not found", slug)
+}
+
+// SearchKeyword performs full-text search using FTS5
+func (vs *VectorStorage) SearchKeyword(query string, limit int) ([]models.VectorRow, error) {
+	// Use FTS5 bm25 ranking. bm25 returns negative values where more negative is better.
+	// We'll order by bm25 (ascending) and limit.
+	ftsQuery := `SELECT slug, raw_text, filename, bm25(fts_embeddings) as score 
+				 FROM fts_embeddings 
+				 WHERE fts_embeddings MATCH ? 
+				 ORDER BY score 
+				 LIMIT ?`
+
+	// Try original query first
+	rows, err := vs.sqlxDB.Query(ftsQuery, query, limit)
+	if err != nil {
+		return nil, fmt.Errorf("FTS search failed: %w", err)
+	}
+	results, err := vs.scanRows(rows)
+	rows.Close()
+	if err != nil {
+		return nil, err
+	}
+
+	// If no results and query contains multiple terms, try OR fallback
+	if len(results) == 0 && strings.Contains(query, " ") && !strings.Contains(strings.ToUpper(query), " OR ") {
+		// Build OR query: term1 OR term2 OR term3
+		terms := strings.Fields(query)
+		if len(terms) > 1 {
+			orQuery := strings.Join(terms, " OR ")
+			rows, err := vs.sqlxDB.Query(ftsQuery, orQuery, limit)
+			if err != nil {
+				// Return original empty results rather than error
+				return results, nil
+			}
+			orResults, err := vs.scanRows(rows)
+			rows.Close()
+			if err == nil {
+				results = orResults
+			}
+		}
+	}
+	return results, nil
+}
+
+// scanRows converts SQL rows to VectorRow slice
+func (vs *VectorStorage) scanRows(rows *sql.Rows) ([]models.VectorRow, error) {
+	var results []models.VectorRow
+	for rows.Next() {
+		var slug, rawText, fileName string
+		var score float64
+		if err := rows.Scan(&slug, &rawText, &fileName, &score); err != nil {
+			vs.logger.Error("failed to scan FTS row", "error", err)
+			continue
+		}
+		// Convert BM25 score to distance-like metric (lower is better)
+		// BM25 is negative, more negative is better. We'll normalize to positive distance.
+		distance := float32(-score) // Make positive (since score is negative)
+		if distance < 0 {
+			distance = 0
+		}
+		results = append(results, models.VectorRow{
+			Slug:     slug,
+			RawText:  rawText,
+			FileName: fileName,
+			Distance: distance,
+		})
+	}
+	return results, nil
+}
+
 // ListFiles returns a list of all loaded files
 func (vs *VectorStorage) ListFiles() ([]string, error) {
 	fileLists := make([][]string, 0)
@@ -215,6 +399,10 @@ func (vs *VectorStorage) ListFiles() ([]string, error) {
 // RemoveEmbByFileName removes all embeddings associated with a specific filename
 func (vs *VectorStorage) RemoveEmbByFileName(filename string) error {
 	var errors []string
+	// Delete from FTS table first
+	if _, err := vs.sqlxDB.Exec("DELETE FROM fts_embeddings WHERE filename = ?", filename); err != nil {
+		errors = append(errors, err.Error())
+	}
 	embeddingSizes := []int{384, 768, 1024, 1536, 2048, 3072, 4096, 5120}
 	for _, size := range embeddingSizes {
 		table := fmt.Sprintf("embeddings_%d", size)

session.go

@@ -29,7 +29,7 @@ func historyToSJSON(msgs []models.RoleMsg) (string, error) {
 }
 
 func exportChat() error {
-	data, err := json.MarshalIndent(chatBody.Messages, "", "  ")
+	data, err := json.MarshalIndent(chatBody.GetMessages(), "", "  ")
 	if err != nil {
 		return err
 	}
@@ -54,7 +54,7 @@ func importChat(filename string) error {
 	if _, ok := chatMap[activeChatName]; !ok {
 		addNewChat(activeChatName)
 	}
-	chatBody.Messages = messages
+	chatBody.SetMessages(messages)
 	cfg.AssistantRole = messages[1].Role
 	if cfg.AssistantRole == cfg.UserRole {
 		cfg.AssistantRole = messages[2].Role

storage/migrations/003_add_fts.down.sql

@@ -0,0 +1,2 @@
+-- Drop FTS5 virtual table
+DROP TABLE IF EXISTS fts_embeddings;

storage/migrations/003_add_fts.up.sql

@@ -0,0 +1,15 @@
+-- Create FTS5 virtual table for full-text search
+CREATE VIRTUAL TABLE IF NOT EXISTS fts_embeddings USING fts5(
+    slug UNINDEXED,
+    raw_text,
+    filename UNINDEXED,
+    embedding_size UNINDEXED,
+    tokenize='porter unicode61'  -- Use porter stemmer and unicode61 tokenizer
+);
+
+-- Create triggers to maintain FTS table when embeddings are inserted/deleted
+-- Note: We'll handle inserts/deletes programmatically for simplicity
+-- but triggers could be added here if needed.
+
+-- Indexes for performance (FTS5 manages its own indexes)
+-- No additional indexes needed for FTS5 virtual table.

storage/migrations/004_populate_fts.down.sql

@@ -0,0 +1,2 @@
+-- Clear FTS table (optional)
+DELETE FROM fts_embeddings;

storage/migrations/004_populate_fts.up.sql

@@ -0,0 +1,26 @@
+-- Populate FTS table with existing embeddings
+DELETE FROM fts_embeddings;
+
+INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size)
+SELECT slug, raw_text, filename, 384 FROM embeddings_384;
+
+INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size)
+SELECT slug, raw_text, filename, 768 FROM embeddings_768;
+
+INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size)
+SELECT slug, raw_text, filename, 1024 FROM embeddings_1024;
+
+INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size)
+SELECT slug, raw_text, filename, 1536 FROM embeddings_1536;
+
+INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size)
+SELECT slug, raw_text, filename, 2048 FROM embeddings_2048;
+
+INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size)
+SELECT slug, raw_text, filename, 3072 FROM embeddings_3072;
+
+INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size)
+SELECT slug, raw_text, filename, 4096 FROM embeddings_4096;
+
+INSERT INTO fts_embeddings (slug, raw_text, filename, embedding_size)
+SELECT slug, raw_text, filename, 5120 FROM embeddings_5120;

storage/storage.go

@@ -102,6 +102,22 @@ func NewProviderSQL(dbPath string, logger *slog.Logger) FullRepo {
 		logger.Error("failed to open db connection", "error", err)
 		return nil
 	}
+	// Enable WAL mode for better concurrency and performance
+	if _, err := db.Exec("PRAGMA journal_mode = WAL;"); err != nil {
+		logger.Warn("failed to enable WAL mode", "error", err)
+	}
+	if _, err := db.Exec("PRAGMA synchronous = NORMAL;"); err != nil {
+		logger.Warn("failed to set synchronous mode", "error", err)
+	}
+	// Increase cache size for better performance
+	if _, err := db.Exec("PRAGMA cache_size = -2000;"); err != nil {
+		logger.Warn("failed to set cache size", "error", err)
+	}
+	// Log actual journal mode for debugging
+	var journalMode string
+	if err := db.QueryRow("PRAGMA journal_mode;").Scan(&journalMode); err == nil {
+		logger.Debug("SQLite journal mode", "mode", journalMode)
+	}
 	p := ProviderSQL{db: db, logger: logger}
 	if err := p.Migrate(); err != nil {
 		logger.Error("migration failed, app cannot start", "error", err)

storage/vector.go

@@ -4,6 +4,7 @@ import (
 	"encoding/binary"
 	"fmt"
 	"gf-lt/models"
+	"sort"
 	"unsafe"
 
 	"github.com/jmoiron/sqlx"
@@ -11,7 +12,7 @@ import (
 
 type VectorRepo interface {
 	WriteVector(*models.VectorRow) error
-	SearchClosest(q []float32) ([]models.VectorRow, error)
+	SearchClosest(q []float32, limit int) ([]models.VectorRow, error)
 	ListFiles() ([]string, error)
 	RemoveEmbByFileName(filename string) error
 	DB() *sqlx.DB
@@ -79,7 +80,7 @@ func (p ProviderSQL) WriteVector(row *models.VectorRow) error {
 	return err
 }
 
-func (p ProviderSQL) SearchClosest(q []float32) ([]models.VectorRow, error) {
+func (p ProviderSQL) SearchClosest(q []float32, limit int) ([]models.VectorRow, error) {
 	tableName, err := fetchTableName(q)
 	if err != nil {
 		return nil, err
@@ -94,7 +95,7 @@ func (p ProviderSQL) SearchClosest(q []float32) ([]models.VectorRow, error) {
 		vector   models.VectorRow
 		distance float32
 	}
-	var topResults []SearchResult
+	var allResults []SearchResult
 	for rows.Next() {
 		var (
 			embeddingsBlob          []byte
@@ -119,28 +120,19 @@ func (p ProviderSQL) SearchClosest(q []float32) ([]models.VectorRow, error) {
 			},
 			distance: distance,
 		}
-
+		allResults = append(allResults, result)
-		// Add to top results and maintain only top results
-		topResults = append(topResults, result)
-
-		// Sort and keep only top results
-		// We'll keep the top 3 closest vectors
-		if len(topResults) > 3 {
-			// Simple sort and truncate to maintain only 3 best matches
-			for i := 0; i < len(topResults); i++ {
-				for j := i + 1; j < len(topResults); j++ {
-					if topResults[i].distance > topResults[j].distance {
-						topResults[i], topResults[j] = topResults[j], topResults[i]
 	}
+	// Sort by distance
+	sort.Slice(allResults, func(i, j int) bool {
+		return allResults[i].distance < allResults[j].distance
+	})
+	// Truncate to limit
+	if len(allResults) > limit {
+		allResults = allResults[:limit]
 	}
-			}
-			topResults = topResults[:3]
-		}
-	}
-
 	// Convert back to VectorRow slice
-	results := make([]models.VectorRow, len(topResults))
+	results := make([]models.VectorRow, len(allResults))
-	for i, result := range topResults {
+	for i, result := range allResults {
 		result.vector.Distance = result.distance
 		results[i] = result.vector
 	}

sysprompts/cluedo.json

@@ -1,7 +0,0 @@
-{
-  "sys_prompt": "A game of cluedo. Players are {{user}}, {{char}}, {{char2}};\n\nrooms: hall, lounge, dinning room kitchen, ballroom, conservatory, billiard room, library, study;\nweapons: candlestick, dagger, lead pipe, revolver, rope, spanner;\npeople: miss Scarlett, colonel Mustard, mrs. White, reverend Green, mrs. Peacock, professor Plum;\n\nA murder happened in a mansion with 9 rooms. Victim is dr. Black.\nPlayers goal is to find out who commited a murder, in what room and with what weapon.\nWeapons, people and rooms not involved in murder are distributed between players (as cards) by tool agent.\nThe objective of the game is to deduce the details of the murder. There are six characters, six murder weapons, and nine rooms, leaving the players with 324 possibilities. As soon as a player enters a room, they may make a suggestion as to the details, naming a suspect, the room they are in, and the weapon. For example: \"I suspect Professor Plum, in the Dining Room, with the candlestick\".\nOnce a player makes a suggestion, the others are called upon to disprove it.\nBefore the player's move, tool agent will remind that players their cards. There are two types of moves: making a suggestion (suggestion_move) and disproving other player suggestion (evidence_move);\nIn this version player wins when the correct details are named in the suggestion_move.\n\n<example_game>\n{{user}}:\nlet's start a game of cluedo!\ntool: cards of {{char}} are 'LEAD PIPE', 'BALLROOM', 'CONSERVATORY', 'STUDY', 'Mrs. White'; suggestion_move;\n{{char}}:\n(putting miss Scarlet into the Hall with the Revolver) \"I suspect miss Scarlett, in the Hall, with the revolver.\"\ntool: cards of {{char2}} are 'SPANNER', 'DAGGER', 'Professor Plum', 'LIBRARY', 'Mrs. Peacock'; evidence_move;\n{{char2}}:\n\"No objections.\" (no cards matching the suspicion of {{char}})\ntool: cards of {{user}} are 'Colonel Mustard', 'Miss Scarlett', 'DINNING ROOM', 'CANDLESTICK', 'HALL'; evidence_move;\n{{user}}:\n\"I object. Miss Scarlett is innocent.\" (shows card with 'Miss Scarlett')\ntool: cards of {{char2}} are 'SPANNER', 'DAGGER', 'Professor Plum', 'LIBRARY', 'Mrs. Peacock'; suggestion_move;\n{{char2}}:\n*So it was not Miss Scarlett, good to know.*\n(moves Mrs. White to the Billiard Room) \"It might have been Mrs. White, in the Billiard Room, with the Revolver.\"\ntool: cards of {{user}} are 'Colonel Mustard', 'Miss Scarlett', 'DINNING ROOM', 'CANDLESTICK', 'HALL'; evidence_move;\n{{user}}:\n(no matching cards for the assumption of {{char2}}) \"Sounds possible to me.\"\ntool: cards of {{char}} are 'LEAD PIPE', 'BALLROOM', 'CONSERVATORY', 'STUDY', 'Mrs. White'; evidence_move;\n{{char}}:\n(shows Mrs. White card) \"No. Was not Mrs. White\"\ntool: cards of {{user}} are 'Colonel Mustard', 'Miss Scarlett', 'DINNING ROOM', 'CANDLESTICK', 'HALL'; suggestion_move;\n{{user}}:\n*So not Mrs. White...* (moves Reverend Green into the Billiard Room) \"I suspect Reverend Green, in the Billiard Room, with the Revolver.\"\ntool: Correct. It was Reverend Green in the Billiard Room, with the revolver. {{user}} wins.\n</example_game>",
-  "role": "CluedoPlayer",
-  "role2": "CluedoEnjoyer",
-  "filepath": "sysprompts/cluedo.json",
-  "first_msg": "Hey guys! Want to play cluedo?"
-}

tables.go

@@ -128,8 +128,8 @@ func makeChatTable(chatMap map[string]models.Chat) *tview.Table {
 				pages.RemovePage(historyPage)
 				return
 			}
-			chatBody.Messages = history
+			chatBody.SetMessages(history)
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			activeChatName = selectedChat
 			pages.RemovePage(historyPage)
 			return
@@ -149,8 +149,8 @@ func makeChatTable(chatMap map[string]models.Chat) *tview.Table {
 			}
 			showToast("chat deleted", selectedChat+" was deleted")
 			// load last chat
-			chatBody.Messages = loadOldChatOrGetNew()
+			chatBody.SetMessages(loadOldChatOrGetNew())
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			pages.RemovePage(historyPage)
 			return
 		case "update card":
@@ -163,16 +163,24 @@ func makeChatTable(chatMap map[string]models.Chat) *tview.Table {
 				showToast("error", "no such card: "+agentName)
 				return
 			}
-			cc.SysPrompt = chatBody.Messages[0].Content
+			if msg0, ok := chatBody.GetMessageAt(0); ok {
-			cc.FirstMsg = chatBody.Messages[1].Content
+				cc.SysPrompt = msg0.Content
+			}
+			if msg1, ok := chatBody.GetMessageAt(1); ok {
+				cc.FirstMsg = msg1.Content
+			}
 			if err := pngmeta.WriteToPng(cc.ToSpec(cfg.UserRole), cc.FilePath, cc.FilePath); err != nil {
 				logger.Error("failed to write charcard", "error", err)
 			}
 			return
 		case "move sysprompt onto 1st msg":
-			chatBody.Messages[1].Content = chatBody.Messages[0].Content + chatBody.Messages[1].Content
+			chatBody.WithLock(func(cb *models.ChatBody) {
-			chatBody.Messages[0].Content = rpDefenitionSysMsg
+				if len(cb.Messages) >= 2 {
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+					cb.Messages[1].Content = cb.Messages[0].Content + cb.Messages[1].Content
+					cb.Messages[0].Content = rpDefenitionSysMsg
+				}
+			})
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			activeChatName = selectedChat
 			pages.RemovePage(historyPage)
 			return
@@ -563,7 +571,7 @@ func makeAgentTable(agentList []string) *tview.Table {
 				return
 			}
 			// replace textview
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			colorText()
 			updateStatusLine()
 			// sysModal.ClearButtons()
@@ -732,7 +740,7 @@ func makeImportChatTable(filenames []string) *tview.Table {
 			colorText()
 			updateStatusLine()
 			// redraw the text in text area
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			pages.RemovePage(historyPage)
 			app.SetFocus(textArea)
 			return

tools.go

@@ -1215,11 +1215,11 @@ func isCommandAllowed(command string, args ...string) bool {
 }
 
 func summarizeChat(args map[string]string) []byte {
-	if len(chatBody.Messages) == 0 {
+	if chatBody.GetMessageCount() == 0 {
 		return []byte("No chat history to summarize.")
 	}
 	// Format chat history for the agent
-	chatText := chatToText(chatBody.Messages, true) // include system and tool messages
+	chatText := chatToText(chatBody.GetMessages(), true) // include system and tool messages
 	return []byte(chatText)
 }
 

tui.go

@@ -29,6 +29,8 @@ var (
 	statusLineWidget   *tview.TextView
 	helpView           *tview.TextView
 	flex               *tview.Flex
+	bottomFlex         *tview.Flex
+	notificationWidget *tview.TextView
 	imgView            *tview.Image
 	defaultImage       = "sysprompts/llama.png"
 	indexPickWindow    *tview.InputField
@@ -36,6 +38,7 @@ var (
 	roleEditWindow     *tview.InputField
 	shellInput         *tview.InputField
 	confirmModal       *tview.Modal
+	toastTimer         *time.Timer
 	confirmPageName    = "confirm"
 	fullscreenMode     bool
 	positionVisible    bool = true
@@ -137,8 +140,8 @@ func setShellMode(enabled bool) {
 	}()
 }
 
-// showToast displays a temporary message in the top‑right corner.
+// showToast displays a temporary notification in the bottom-right corner.
-// It auto‑hides after 3 seconds and disappears when clicked.
+// It auto-hides after 3 seconds.
 func showToast(title, message string) {
 	sanitize := func(s string, maxLen int) string {
 		sanitized := strings.Map(func(r rune) rune {
@@ -154,6 +157,11 @@ func showToast(title, message string) {
 	}
 	title = sanitize(title, 50)
 	message = sanitize(message, 197)
+	if toastTimer != nil {
+		toastTimer.Stop()
+	}
+	// show blocking notification to not mess up flex
+	if fullscreenMode {
 		notification := tview.NewTextView().
 			SetTextAlign(tview.AlignCenter).
 			SetDynamicColors(true).
@@ -176,14 +184,44 @@ func showToast(title, message string) {
 		// Generate a unique page name (e.g., using timestamp) to allow multiple toasts.
 		pageName := fmt.Sprintf("toast-%d", time.Now().UnixNano())
 		pages.AddPage(pageName, background, true, true)
-	// Auto‑dismiss after 3 seconds.
+		// Auto‑dismiss after 2 seconds, since blocking is more annoying
-	time.AfterFunc(3*time.Second, func() {
+		time.AfterFunc(2*time.Second, func() {
 			app.QueueUpdateDraw(func() {
 				if pages.HasPage(pageName) {
 					pages.RemovePage(pageName)
 				}
 			})
 		})
+		return
+	}
+	notificationWidget.SetTitle(title)
+	notificationWidget.SetText(fmt.Sprintf("[yellow]%s[-]", message))
+	go func() {
+		app.QueueUpdateDraw(func() {
+			flex.RemoveItem(bottomFlex)
+			flex.RemoveItem(statusLineWidget)
+			bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+				AddItem(textArea, 0, 1, true).
+				AddItem(notificationWidget, 40, 1, false)
+			flex.AddItem(bottomFlex, 0, 10, true)
+			if positionVisible {
+				flex.AddItem(statusLineWidget, 0, 2, false)
+			}
+		})
+	}()
+	toastTimer = time.AfterFunc(3*time.Second, func() {
+		app.QueueUpdateDraw(func() {
+			flex.RemoveItem(bottomFlex)
+			flex.RemoveItem(statusLineWidget)
+			bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+				AddItem(textArea, 0, 1, true).
+				AddItem(notificationWidget, 0, 0, false)
+			flex.AddItem(bottomFlex, 0, 10, true)
+			if positionVisible {
+				flex.AddItem(statusLineWidget, 0, 2, false)
+			}
+		})
+	})
 }
 
 func init() {
@@ -235,7 +273,7 @@ func init() {
 			shellHistoryPos = -1
 		}
 		// Handle Tab key for @ file completion
-		if event.Key() == tcell.KeyTab {
+		if event.Key() == tcell.KeyTab && shellMode {
 			currentText := shellInput.GetText()
 			atIndex := strings.LastIndex(currentText, "@")
 			if atIndex >= 0 {
@@ -286,12 +324,26 @@ func init() {
 		SetDynamicColors(true).
 		SetRegions(true).
 		SetChangedFunc(func() {
+			// INFO:
+			// https://github.com/rivo/tview/wiki/Concurrency#event-handlers
+			// although already called by default per tview specs
+			// calling it explicitly makes text streaming to look more smooth
 			app.Draw()
 		})
+	notificationWidget = tview.NewTextView().
+		SetTextAlign(tview.AlignCenter).
+		SetDynamicColors(true).
+		SetRegions(true).
+		SetChangedFunc(func() {
+		})
+	notificationWidget.SetBorder(true).SetTitle("notification")
+	bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+		AddItem(textArea, 0, 1, true).
+		AddItem(notificationWidget, 0, 0, false)
 	//
 	flex = tview.NewFlex().SetDirection(tview.FlexRow).
 		AddItem(textView, 0, 40, false).
-		AddItem(textArea, 0, 10, true) // Restore original height
+		AddItem(bottomFlex, 0, 10, true)
 	if positionVisible {
 		flex.AddItem(statusLineWidget, 0, 2, false)
 	}
@@ -303,7 +355,7 @@ func init() {
 				searchResults = nil       // Clear search results
 				searchResultLengths = nil // Clear search result lengths
 				originalTextForSearch = ""
-				textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys)) // Reset text without search regions
+				textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys)) // Reset text without search regions
 				colorText()                                                  // Apply normal chat coloring
 			} else {
 				// Original logic if no search is active
@@ -360,10 +412,14 @@ func init() {
 	// 	y += h / 2
 	// 	return x, y, w, h
 	// })
+	notificationWidget.SetDrawFunc(func(screen tcell.Screen, x, y, w, h int) (int, int, int, int) {
+		y += h / 2
+		return x, y, w, h
+	})
 	// Initially set up flex without search bar
 	flex = tview.NewFlex().SetDirection(tview.FlexRow).
 		AddItem(textView, 0, 40, false).
-		AddItem(textArea, 0, 10, true) // Restore original height
+		AddItem(bottomFlex, 0, 10, true)
 	if positionVisible {
 		flex.AddItem(statusLineWidget, 0, 2, false)
 	}
@@ -380,9 +436,11 @@ func init() {
 				pages.RemovePage(editMsgPage)
 				return nil
 			}
-			chatBody.Messages[selectedIndex].SetText(editedMsg)
+			chatBody.WithLock(func(cb *models.ChatBody) {
+				cb.Messages[selectedIndex].SetText(editedMsg)
+			})
 			// change textarea
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			pages.RemovePage(editMsgPage)
 			editMode = false
 			return nil
@@ -410,9 +468,11 @@ func init() {
 					pages.RemovePage(roleEditPage)
 					return
 				}
-				if selectedIndex >= 0 && selectedIndex < len(chatBody.Messages) {
+				if selectedIndex >= 0 && selectedIndex < chatBody.GetMessageCount() {
-					chatBody.Messages[selectedIndex].Role = newRole
+					chatBody.WithLock(func(cb *models.ChatBody) {
-					textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+						cb.Messages[selectedIndex].Role = newRole
+					})
+					textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 					colorText()
 					pages.RemovePage(roleEditPage)
 				}
@@ -441,7 +501,7 @@ func init() {
 				return nil
 			}
 			selectedIndex = siInt
-			if len(chatBody.Messages)-1 < selectedIndex || selectedIndex < 0 {
+			if chatBody.GetMessageCount()-1 < selectedIndex || selectedIndex < 0 {
 				msg := "chosen index is out of bounds, will copy user input"
 				logger.Warn(msg, "index", selectedIndex)
 				showToast("error", msg)
@@ -451,7 +511,7 @@ func init() {
 				hideIndexBar() // Hide overlay instead of removing page directly
 				return nil
 			}
-			m := chatBody.Messages[selectedIndex]
+			m := chatBody.GetMessages()[selectedIndex]
 			switch {
 			case roleEditMode:
 				hideIndexBar() // Hide overlay first
@@ -518,7 +578,7 @@ func init() {
 					searchResults = nil
 					searchResultLengths = nil
 					originalTextForSearch = ""
-					textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+					textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 					colorText()
 					return
 				} else {
@@ -576,7 +636,7 @@ func init() {
 	//
 	textArea.SetMovedFunc(updateStatusLine)
 	updateStatusLine()
-	textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+	textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 	colorText()
 	if scrollToEndEnabled {
 		textView.ScrollToEnd()
@@ -590,7 +650,7 @@ func init() {
 		if event.Key() == tcell.KeyRune && event.Rune() == '5' && event.Modifiers()&tcell.ModAlt != 0 {
 			// switch cfg.ShowSys
 			cfg.ShowSys = !cfg.ShowSys
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			colorText()
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == '3' && event.Modifiers()&tcell.ModAlt != 0 {
@@ -623,7 +683,7 @@ func init() {
 		// Handle Alt+T to toggle thinking block visibility
 		if event.Key() == tcell.KeyRune && event.Rune() == 't' && event.Modifiers()&tcell.ModAlt != 0 {
 			thinkingCollapsed = !thinkingCollapsed
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			colorText()
 			status := "expanded"
 			if thinkingCollapsed {
@@ -635,7 +695,7 @@ func init() {
 		// Handle Ctrl+T to toggle tool call/response visibility
 		if event.Key() == tcell.KeyCtrlT {
 			toolCollapsed = !toolCollapsed
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			colorText()
 			status := "expanded"
 			if toolCollapsed {
@@ -678,14 +738,14 @@ func init() {
 		}
 		if event.Key() == tcell.KeyF2 && !botRespMode {
 			// regen last msg
-			if len(chatBody.Messages) == 0 {
+			if chatBody.GetMessageCount() == 0 {
 				showToast("info", "no messages to regenerate")
 				return nil
 			}
-			chatBody.Messages = chatBody.Messages[:len(chatBody.Messages)-1]
+			chatBody.TruncateMessages(chatBody.GetMessageCount() - 1)
 			// there is no case where user msg is regenerated
-			// lastRole := chatBody.Messages[len(chatBody.Messages)-1].Role
+			// lastRole := chatBody.GetMessages()[chatBody.GetMessageCount()-1].Role
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			// go chatRound("", cfg.UserRole, textView, true, false)
 			if cfg.TTS_ENABLED {
 				TTSDoneChan <- true
@@ -704,12 +764,12 @@ func init() {
 				colorText()
 				return nil
 			}
-			if len(chatBody.Messages) == 0 {
+			if chatBody.GetMessageCount() == 0 {
 				showToast("info", "no messages to delete")
 				return nil
 			}
-			chatBody.Messages = chatBody.Messages[:len(chatBody.Messages)-1]
+			chatBody.TruncateMessages(chatBody.GetMessageCount() - 1)
-			textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
+			textView.SetText(chatToText(chatBody.GetMessages(), cfg.ShowSys))
 			if cfg.TTS_ENABLED {
 				TTSDoneChan <- true
 			}
@@ -757,7 +817,7 @@ func init() {
 		if event.Key() == tcell.KeyF7 {
 			// copy msg to clipboard
 			editMode = false
-			m := chatBody.Messages[len(chatBody.Messages)-1]
+			m := chatBody.GetMessages()[chatBody.GetMessageCount()-1]
 			msgText := m.GetText()
 			if err := copyToClipboard(msgText); err != nil {
 				logger.Error("failed to copy to clipboard", "error", err)
@@ -941,10 +1001,10 @@ func init() {
 			TTSDoneChan <- true
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == '0' && event.Modifiers()&tcell.ModAlt != 0 && cfg.TTS_ENABLED {
-			if len(chatBody.Messages) > 0 {
+			if chatBody.GetMessageCount() > 0 {
 				// Stop any currently playing TTS first
 				TTSDoneChan <- true
-				lastMsg := chatBody.Messages[len(chatBody.Messages)-1]
+				lastMsg := chatBody.GetMessages()[chatBody.GetMessageCount()-1]
 				cleanedText := models.CleanText(lastMsg.GetText())
 				if cleanedText != "" {
 					// nolint: errcheck
@@ -956,7 +1016,7 @@ func init() {
 		if event.Key() == tcell.KeyCtrlW {
 			// INFO: continue bot/text message
 			// without new role
-			lastRole := chatBody.Messages[len(chatBody.Messages)-1].Role
+			lastRole := chatBody.GetMessages()[chatBody.GetMessageCount()-1].Role
 			// go chatRound("", lastRole, textView, false, true)
 			chatRoundChan <- &models.ChatRoundReq{Role: lastRole, Resume: true}
 			return nil
@@ -1042,7 +1102,7 @@ func init() {
 		if event.Key() == tcell.KeyRune && event.Modifiers() == tcell.ModAlt && event.Rune() == '9' {
 			// Warm up (load) the currently selected model
 			go warmUpModel()
-			showToast("model warmup", "loading model: "+chatBody.Model)
+			showToast("model warmup", "loading model: "+chatBody.GetModel())
 			return nil
 		}
 		// cannot send msg in editMode or botRespMode
@@ -1081,7 +1141,7 @@ func init() {
 				}
 				// add user icon before user msg
 				fmt.Fprintf(textView, "%s[-:-:b](%d) <%s>: [-:-:-]\n%s\n",
-					nl, len(chatBody.Messages), persona, msgText)
+					nl, chatBody.GetMessageCount(), persona, msgText)
 				textArea.SetText("", true)
 				if scrollToEndEnabled {
 					textView.ScrollToEnd()
@@ -1095,7 +1155,7 @@ func init() {
 			chatRoundChan <- &models.ChatRoundReq{Role: persona, UserMsg: msgText}
 			return nil
 		}
-		if event.Key() == tcell.KeyTab {
+		if event.Key() == tcell.KeyTab && !shellMode {
 			currentF := app.GetFocus()
 			if currentF == textArea {
 				currentText := textArea.GetText()