Enha: rag tuning and tests

Enha (rag): semantic hybrid search
Enha (rag): bigger default batch
2026-03-08 16:12:32 +03:00 · 2026-03-08 13:27:09 +03:00 · 2026-03-08 11:38:56 +03:00 · 2026-03-08 10:28:30 +03:00 · 2026-03-08 09:42:07 +03:00 · 2026-03-08 09:00:24 +03:00
36 changed files with 3354 additions and 1302 deletions
--- a/.gitignore
+++ b/.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/
--- a/108
+++ b/108
@@ -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

@@ -44,11 +143,10 @@ build-whisper: ## Build whisper.cpp from source in batteries directory

 download-whisper-model: ## Download Whisper model for STT in batteries directory
 	@echo "Downloading Whisper model for STT..."
-	@if [ ! -d "batteries/whisper.cpp" ]; then \
-		echo "Please run 'make setup-whisper' first to clone the repository."; \
-		exit 1; \
+	@if [ ! -d "batteries/whisper.cpp/models" ]; then \
+		mkdir -p "batteries/whisper.cpp/models" \
 	fi
-	@cd batteries/whisper.cpp && bash ./models/download-ggml-model.sh large-v3-turbo-q5_0
+	curl -o batteries/whisper.cpp/models/ggml-large-v3-turbo-q5_0.bin -L "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q5_0.bin?download=true"
 	@echo "Whisper model downloaded successfully!"

 # Docker targets for STT/TTS services (in batteries directory)
--- a/README.md
+++ b/README.md
@@ -13,6 +13,12 @@ made with use of [tview](https://github.com/rivo/tview)
 #### how it looks
 ![how it looks](assets/ex01.png)

+
+#### dependencies
+- make
+- go
+- ffmpeg (extra)
+
 #### how to install
 (requires golang)
 clone the project
--- a/batteries/docker-compose.yml
+++ b/batteries/docker-compose.yml
@@ -6,19 +6,27 @@ services:
    ports:
      - "8081:8081"
    volumes:
-      - whisper_models:/app/models
+      - ./whisper.cpp/models/ggml-large-v3-turbo-q5_0.bin:/app/models/ggml-large-v3-turbo-q5_0.bin
    working_dir: /app
    entrypoint: ""
    command: >
      sh -c "
-      if [ ! -f /app/models/ggml-large-v3-turbo.bin ]; then
-        echo 'Downloading ggml-large-v3-turbo model...'
-        ./download-ggml-model.sh large-v3-turbo /app/models
+      if [ ! -f /app/models/ggml-large-v3-turbo-q5_0.bin ]; then
+        echo 'Downloading ggml-large-v3-turboq5_0 model...'
+      	curl -o /app/models/ggml-large-v3-turbo-q5_0.bin -L "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q5_0.bin?download=true"
      fi &&
-      ./build/bin/whisper-server -m /app/models/ggml-large-v3-turbo.bin -t 4 -p 1 --port 8081 --host 0.0.0.0
+      ./build/bin/whisper-server -m /app/models/ggml-large-v3-turbo-q5_0.bin -t 4 -p 1 --port 8081 --host 0.0.0.0
      "
    environment:
      - WHISPER_LOG_LEVEL=3
+    # For GPU support, uncomment the following lines:
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
    # Restart policy in case the service fails
    restart: unless-stopped

@@ -45,7 +53,5 @@ services:
 volumes:
  models:
    driver: local
-  audio:
-    driver: local
  whisper_models:
    driver: local
--- a/bot.go
+++ b/bot.go
@@ -16,13 +16,13 @@ import (
 	"log/slog"
 	"net"
 	"net/http"
-	"net/url"
 	"os"
 	"regexp"
 	"slices"
 	"strconv"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 )

@@ -41,7 +41,7 @@ var (
 	store           storage.FullRepo
 	defaultFirstMsg = "Hello! What can I do for you?"
 	defaultStarter  = []models.RoleMsg{}
-	interruptResp   = false
+	interruptResp   atomic.Bool
 	ragger          *rag.RAG
 	chunkParser     ChunkParser
 	lastToolCall    *models.FuncCall
@@ -253,12 +253,7 @@ func createClient(connectTimeout time.Duration) *http.Client {
 }

 func warmUpModel() {
-	u, err := url.Parse(cfg.CurrentAPI)
-	if err != nil {
-		return
-	}
-	host := u.Hostname()
-	if host != "localhost" && host != "127.0.0.1" && host != "::1" {
+	if !isLocalLlamacpp() {
 		return
 	}
 	// Check if model is already loaded
@@ -649,7 +644,7 @@ func sendMsgToLLM(body io.Reader) {
 			// continue
 		}
 		if len(line) <= 1 {
-			if interruptResp {
+			if interruptResp.Load() {
 				goto interrupt // get unstuck from bad connection
 			}
 			continue // skip \n
@@ -742,8 +737,7 @@ func sendMsgToLLM(body io.Reader) {
 			lastToolCall.ID = chunk.ToolID
 		}
 	interrupt:
-		if interruptResp { // read bytes, so it would not get into beginning of the next req
-			// interruptResp = false
+		if interruptResp.Load() { // read bytes, so it would not get into beginning of the next req
 			logger.Info("interrupted bot response", "chunk_counter", counter)
 			streamDone <- true
 			break
@@ -776,14 +770,14 @@ func showSpinner() {
 	if cfg.WriteNextMsgAsCompletionAgent != "" {
 		botPersona = cfg.WriteNextMsgAsCompletionAgent
 	}
-	for botRespMode || toolRunningMode {
+	for botRespMode.Load() || toolRunningMode.Load() {
 		time.Sleep(400 * time.Millisecond)
 		spin := i % len(spinners)
 		app.QueueUpdateDraw(func() {
 			switch {
-			case toolRunningMode:
+			case toolRunningMode.Load():
 				textArea.SetTitle(spinners[spin] + " tool")
-			case botRespMode:
+			case botRespMode.Load():
 				textArea.SetTitle(spinners[spin] + " " + botPersona + " (F6 to interrupt)")
 			default:
 				textArea.SetTitle(spinners[spin] + " input")
@@ -797,8 +791,8 @@ func showSpinner() {
 }

 func chatRound(r *models.ChatRoundReq) error {
-	interruptResp = false
-	botRespMode = true
+	interruptResp.Store(false)
+	botRespMode.Store(true)
 	go showSpinner()
 	updateStatusLine()
 	botPersona := cfg.AssistantRole
@@ -806,7 +800,7 @@ func chatRound(r *models.ChatRoundReq) error {
 		botPersona = cfg.WriteNextMsgAsCompletionAgent
 	}
 	defer func() {
-		botRespMode = false
+		botRespMode.Store(false)
 		ClearImageAttachment()
 	}()
 	// check that there is a model set to use if is not local
@@ -857,7 +851,7 @@ out:
 				if thinkingCollapsed {
 					// Show placeholder immediately when thinking starts in collapsed mode
 					fmt.Fprint(textView, "[yellow::i][thinking... (press Alt+T to expand)][-:-:-]")
-					if scrollToEndEnabled {
+					if cfg.AutoScrollEnabled {
 						textView.ScrollToEnd()
 					}
 					respText.WriteString(chunk)
@@ -872,7 +866,7 @@ out:
 						// Thinking already displayed as placeholder, just update respText
 						respText.WriteString(chunk)
 						justExitedThinkingCollapsed = true
-						if scrollToEndEnabled {
+						if cfg.AutoScrollEnabled {
 							textView.ScrollToEnd()
 						}
 						continue
@@ -893,8 +887,10 @@ 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()
-			if scrollToEndEnabled {
+			if !r.Resume {
+				chatBody.Messages[msgIdx].Content += respText.String()
+			}
+			if cfg.AutoScrollEnabled {
 				textView.ScrollToEnd()
 			}
 			// Send chunk to audio stream handler
@@ -904,7 +900,7 @@ out:
 		case toolChunk := <-openAIToolChan:
 			fmt.Fprint(textView, toolChunk)
 			toolResp.WriteString(toolChunk)
-			if scrollToEndEnabled {
+			if cfg.AutoScrollEnabled {
 				textView.ScrollToEnd()
 			}
 		case <-streamDone:
@@ -912,7 +908,7 @@ out:
 				chunk := <-chunkChan
 				fmt.Fprint(textView, chunk)
 				respText.WriteString(chunk)
-				if scrollToEndEnabled {
+				if cfg.AutoScrollEnabled {
 					textView.ScrollToEnd()
 				}
 				if cfg.TTS_ENABLED {
@@ -934,7 +930,7 @@ out:
 		}
 		lastRespStats = nil
 	}
-	botRespMode = false
+	botRespMode.Store(false)
 	if r.Resume {
 		chatBody.Messages[len(chatBody.Messages)-1].Content += respText.String()
 		updatedMsg := chatBody.Messages[len(chatBody.Messages)-1]
@@ -963,7 +959,7 @@ out:
 	}
 	// Strip think blocks before parsing for tool calls
 	respTextNoThink := thinkBlockRE.ReplaceAllString(respText.String(), "")
-	if interruptResp {
+	if interruptResp.Load() {
 		return nil
 	}
 	if findCall(respTextNoThink, toolResp.String()) {
@@ -1198,9 +1194,9 @@ func findCall(msg, toolCall string) bool {
 	}
 	// Show tool call progress indicator before execution
 	fmt.Fprintf(textView, "\n[yellow::i][tool: %s...][-:-:-]", fc.Name)
-	toolRunningMode = true
+	toolRunningMode.Store(true)
 	resp := callToolWithAgent(fc.Name, fc.Args)
-	toolRunningMode = false
+	toolRunningMode.Store(false)
 	toolMsg := string(resp)
 	logger.Info("llm used a tool call", "tool_name", fc.Name, "too_args", fc.Args, "id", fc.ID, "tool_resp", toolMsg)
 	// Create tool response message with the proper tool_call_id
@@ -1393,27 +1389,29 @@ func updateModelLists() {
 		}
 	}
 	// if llama.cpp started after gf-lt?
-	localModelsMu.Lock()
-	LocalModels, err = fetchLCPModelsWithLoadStatus()
-	localModelsMu.Unlock()
+	ml, err := fetchLCPModelsWithLoadStatus()
 	if err != nil {
 		logger.Warn("failed to fetch llama.cpp models", "error", err)
 	}
+	localModelsMu.Lock()
+	LocalModels = ml
+	localModelsMu.Unlock()
 	// set already loaded model in llama.cpp
-	if strings.Contains(cfg.CurrentAPI, "localhost") || strings.Contains(cfg.CurrentAPI, "127.0.0.1") {
-		localModelsMu.Lock()
-		defer localModelsMu.Unlock()
-		for i := range LocalModels {
-			if strings.Contains(LocalModels[i], models.LoadedMark) {
-				m := strings.TrimPrefix(LocalModels[i], models.LoadedMark)
-				cfg.CurrentModel = m
-				chatBody.Model = m
-				cachedModelColor = "green"
-				updateStatusLine()
-				updateToolCapabilities()
-				app.Draw()
-				return
-			}
+	if !isLocalLlamacpp() {
+		return
+	}
+	localModelsMu.Lock()
+	defer localModelsMu.Unlock()
+	for i := range LocalModels {
+		if strings.Contains(LocalModels[i], models.LoadedMark) {
+			m := strings.TrimPrefix(LocalModels[i], models.LoadedMark)
+			cfg.CurrentModel = m
+			chatBody.Model = m
+			cachedModelColor.Store("green")
+			updateStatusLine()
+			updateToolCapabilities()
+			app.Draw()
+			return
 		}
 	}
 }
@@ -1500,7 +1498,13 @@ func init() {
 		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 {
@@ -1541,57 +1545,9 @@ func init() {
 			}
 		}
 	}
-	// Initialize scrollToEndEnabled based on config
-	scrollToEndEnabled = cfg.AutoScrollEnabled
-	go updateModelLists()
+	// atomic default values
+	cachedModelColor.Store("orange")
 	go chatWatcher(ctx)
-}
-
-func getValidKnowToRecipient(msg *models.RoleMsg) (string, bool) {
-	if cfg == nil || !cfg.CharSpecificContextEnabled {
-		return "", false
-	}
-	// case where all roles are in the tag => public message
-	cr := listChatRoles()
-	slices.Sort(cr)
-	slices.Sort(msg.KnownTo)
-	if slices.Equal(cr, msg.KnownTo) {
-		logger.Info("got msg with tag mentioning every role")
-		return "", false
-	}
-	// Check each character in the KnownTo list
-	for _, recipient := range msg.KnownTo {
-		if recipient == msg.Role || recipient == cfg.ToolRole {
-			// weird cases, skip
-			continue
-		}
-		// Skip if this is the user character (user handles their own turn)
-		// If user is in KnownTo, stop processing - it's the user's turn
-		if recipient == cfg.UserRole || recipient == cfg.WriteNextMsgAs {
-			return "", false
-		}
-		return recipient, true
-	}
-	return "", false
-}
-
-// triggerPrivateMessageResponses checks if a message was sent privately to specific characters
-// and triggers those non-user characters to respond
-func triggerPrivateMessageResponses(msg *models.RoleMsg) {
-	recipient, ok := getValidKnowToRecipient(msg)
-	if !ok || recipient == "" {
-		return
-	}
-	// Trigger the recipient character to respond
-	triggerMsg := recipient + ":\n"
-	// Send empty message so LLM continues naturally from the conversation
-	crr := &models.ChatRoundReq{
-		UserMsg: triggerMsg,
-		Role:    recipient,
-		Resume:  true,
-	}
-	fmt.Fprintf(textView, "\n[-:-:b](%d) ", len(chatBody.Messages))
-	fmt.Fprint(textView, roleToIcon(recipient))
-	fmt.Fprint(textView, "[-:-:-]\n")
-	chatRoundChan <- crr
+	initTUI()
+	initTools()
 }
--- a/config.example.toml
+++ b/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
+RAGWordLimit = 250
+RAGOverlapWords = 25
 RAGDir = "ragimport"
 # extra tts
 TTS_ENABLED = false
--- a/config/config.go
+++ b/config/config.go
@@ -27,20 +27,22 @@ type Config struct {
 	WriteNextMsgAs                string
 	WriteNextMsgAsCompletionAgent string
 	SkipLLMResp                   bool
-	AutoCleanToolCallsFromCtx     bool   `toml:"AutoCleanToolCallsFromCtx"`
 	DBPATH                        string `toml:"DBPATH"`
 	FilePickerDir                 string `toml:"FilePickerDir"`
 	FilePickerExts                string `toml:"FilePickerExts"`
 	ImagePreview                  bool   `toml:"ImagePreview"`
 	EnableMouse                   bool   `toml:"EnableMouse"`
 	// embeddings
-	EmbedURL string `toml:"EmbedURL"`
-	HFToken  string `toml:"HFToken"`
+	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"`
+	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"`
--- a/docs/config.md
+++ b/docs/config.md
@@ -63,17 +63,11 @@ This document explains how to set up and configure the application using the `co
 #### AutoScrollEnabled (`true`)
 - Whether to automatically scroll chat window while llm streams its repsonse.

-#### AutoCleanToolCallsFromCtx (`false`)
- Whether to automatically clean tool calls from the conversation context to manage token usage.
-
 ### RAG (Retrieval Augmented Generation) Settings

 #### 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.

--- a/extra/google_tts.go
+++ b/extra/google_tts.go
@@ -0,0 +1,218 @@
+//go:build extra
+// +build extra
+
+package extra
+
+import (
+	"fmt"
+	"gf-lt/models"
+	"io"
+	"log/slog"
+	"os/exec"
+	"strings"
+	"sync"
+
+	google_translate_tts "github.com/GrailFinder/google-translate-tts"
+	"github.com/neurosnap/sentences/english"
+)
+
+type GoogleTranslateOrator struct {
+	logger *slog.Logger
+	mu     sync.Mutex
+	speech *google_translate_tts.Speech
+	// fields for playback control
+	cmd    *exec.Cmd
+	cmdMu  sync.Mutex
+	stopCh chan struct{}
+	// text buffer and interrupt flag
+	textBuffer strings.Builder
+	interrupt  bool
+	Speed      float32
+}
+
+func (o *GoogleTranslateOrator) stoproutine() {
+	for {
+		<-TTSDoneChan
+		o.logger.Debug("orator got done signal")
+		o.Stop()
+		for len(TTSTextChan) > 0 {
+			<-TTSTextChan
+		}
+		o.mu.Lock()
+		o.textBuffer.Reset()
+		o.interrupt = true
+		o.mu.Unlock()
+	}
+}
+
+func (o *GoogleTranslateOrator) readroutine() {
+	tokenizer, _ := english.NewSentenceTokenizer(nil)
+	for {
+		select {
+		case chunk := <-TTSTextChan:
+			o.mu.Lock()
+			o.interrupt = false
+			_, err := o.textBuffer.WriteString(chunk)
+			if err != nil {
+				o.logger.Warn("failed to write to stringbuilder", "error", err)
+				o.mu.Unlock()
+				continue
+			}
+			text := o.textBuffer.String()
+			sentences := tokenizer.Tokenize(text)
+			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
+			if len(sentences) <= 1 {
+				o.mu.Unlock()
+				continue
+			}
+			completeSentences := sentences[:len(sentences)-1]
+			remaining := sentences[len(sentences)-1].Text
+			o.textBuffer.Reset()
+			o.textBuffer.WriteString(remaining)
+			o.mu.Unlock()
+			for _, sentence := range completeSentences {
+				o.mu.Lock()
+				interrupted := o.interrupt
+				o.mu.Unlock()
+				if interrupted {
+					return
+				}
+				cleanedText := models.CleanText(sentence.Text)
+				if cleanedText == "" {
+					continue
+				}
+				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
+				if err := o.Speak(cleanedText); err != nil {
+					o.logger.Error("tts failed", "sentence", cleanedText, "error", err)
+				}
+			}
+		case <-TTSFlushChan:
+			o.logger.Debug("got flushchan signal start")
+			// lln is done get the whole message out
+			if len(TTSTextChan) > 0 { // otherwise might get stuck
+				for chunk := range TTSTextChan {
+					o.mu.Lock()
+					_, err := o.textBuffer.WriteString(chunk)
+					o.mu.Unlock()
+					if err != nil {
+						o.logger.Warn("failed to write to stringbuilder", "error", err)
+						continue
+					}
+					if len(TTSTextChan) == 0 {
+						break
+					}
+				}
+			}
+			o.mu.Lock()
+			remaining := o.textBuffer.String()
+			remaining = models.CleanText(remaining)
+			o.textBuffer.Reset()
+			o.mu.Unlock()
+			if remaining == "" {
+				continue
+			}
+			o.logger.Debug("calling Speak with remainder", "rem", remaining)
+			sentencesRem := tokenizer.Tokenize(remaining)
+			for _, rs := range sentencesRem { // to avoid dumping large volume of text
+				o.mu.Lock()
+				interrupt := o.interrupt
+				o.mu.Unlock()
+				if interrupt {
+					break
+				}
+				if err := o.Speak(rs.Text); err != nil {
+					o.logger.Error("tts failed", "sentence", rs.Text, "error", err)
+				}
+			}
+		}
+	}
+}
+
+func (o *GoogleTranslateOrator) GetLogger() *slog.Logger {
+	return o.logger
+}
+
+func (o *GoogleTranslateOrator) Speak(text string) error {
+	o.logger.Debug("fn: Speak is called", "text-len", len(text))
+	// Generate MP3 data directly as an io.Reader
+	reader, err := o.speech.GenerateSpeech(text)
+	if err != nil {
+		return fmt.Errorf("generate speech failed: %w", err)
+	}
+	// Wrap in io.NopCloser since GenerateSpeech returns io.Reader (no close needed)
+	body := io.NopCloser(reader)
+	defer body.Close()
+	// Build ffplay command with optional speed filter
+	args := []string{"-nodisp", "-autoexit"}
+	if o.Speed > 0.1 && o.Speed != 1.0 {
+		// atempo range is 0.5 to 2.0; you might clamp it here
+		args = append(args, "-af", fmt.Sprintf("atempo=%.2f", o.Speed))
+	}
+	args = append(args, "-i", "pipe:0")
+	cmd := exec.Command("ffplay", args...)
+	stdin, err := cmd.StdinPipe()
+	if err != nil {
+		return fmt.Errorf("failed to get stdin pipe: %w", err)
+	}
+	o.cmdMu.Lock()
+	o.cmd = cmd
+	o.stopCh = make(chan struct{})
+	o.cmdMu.Unlock()
+	if err := cmd.Start(); err != nil {
+		return fmt.Errorf("failed to start ffplay: %w", err)
+	}
+	copyErr := make(chan error, 1)
+	go func() {
+		_, err := io.Copy(stdin, body)
+		stdin.Close()
+		copyErr <- err
+	}()
+	done := make(chan error, 1)
+	go func() {
+		done <- cmd.Wait()
+	}()
+	select {
+	case <-o.stopCh:
+		if o.cmd != nil && o.cmd.Process != nil {
+			o.cmd.Process.Kill()
+		}
+		<-done
+		return nil
+	case copyErrVal := <-copyErr:
+		if copyErrVal != nil {
+			if o.cmd != nil && o.cmd.Process != nil {
+				o.cmd.Process.Kill()
+			}
+			<-done
+			return copyErrVal
+		}
+		return <-done
+	case err := <-done:
+		return err
+	}
+}
+
+func (o *GoogleTranslateOrator) Stop() {
+	o.cmdMu.Lock()
+	defer o.cmdMu.Unlock()
+	// Signal any running Speak to stop
+	if o.stopCh != nil {
+		select {
+		case <-o.stopCh: // already closed
+		default:
+			close(o.stopCh)
+		}
+		o.stopCh = nil
+	}
+	// Kill the external player process if it's still running
+	if o.cmd != nil && o.cmd.Process != nil {
+		o.cmd.Process.Kill()
+		o.cmd.Wait() // clean up zombie process
+		o.cmd = nil
+	}
+	// Also reset text buffer and interrupt flag (with o.mu)
+	o.mu.Lock()
+	o.textBuffer.Reset()
+	o.interrupt = true
+	o.mu.Unlock()
+}
--- a/extra/kokoro.go
+++ b/extra/kokoro.go
@@ -0,0 +1,259 @@
+//go:build extra
+// +build extra
+
+package extra
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"gf-lt/models"
+	"io"
+	"log/slog"
+	"net/http"
+	"os/exec"
+	"strings"
+	"sync"
+
+	"github.com/neurosnap/sentences/english"
+)
+
+type KokoroOrator struct {
+	logger   *slog.Logger
+	mu       sync.Mutex
+	URL      string
+	Format   models.AudioFormat
+	Stream   bool
+	Speed    float32
+	Language string
+	Voice    string
+	// fields for playback control
+	cmd    *exec.Cmd
+	cmdMu  sync.Mutex
+	stopCh chan struct{}
+	// textBuffer, interrupt etc. remain the same
+	textBuffer strings.Builder
+	interrupt  bool
+}
+
+func (o *KokoroOrator) GetLogger() *slog.Logger {
+	return o.logger
+}
+
+func (o *KokoroOrator) Speak(text string) error {
+	o.logger.Debug("fn: Speak is called", "text-len", len(text))
+	body, err := o.requestSound(text)
+	if err != nil {
+		return fmt.Errorf("request failed: %w", err)
+	}
+	defer body.Close()
+	cmd := exec.Command("ffplay", "-nodisp", "-autoexit", "-i", "pipe:0")
+	stdin, err := cmd.StdinPipe()
+	if err != nil {
+		return fmt.Errorf("failed to get stdin pipe: %w", err)
+	}
+	o.cmdMu.Lock()
+	o.cmd = cmd
+	o.stopCh = make(chan struct{})
+	o.cmdMu.Unlock()
+	if err := cmd.Start(); err != nil {
+		return fmt.Errorf("failed to start ffplay: %w", err)
+	}
+	// Copy audio in background
+	copyErr := make(chan error, 1)
+	go func() {
+		_, err := io.Copy(stdin, body)
+		stdin.Close()
+		copyErr <- err
+	}()
+	// Wait for player in background
+	done := make(chan error, 1)
+	go func() {
+		done <- cmd.Wait()
+	}()
+	// Wait for BOTH copy and player, but ensure we block until done
+	select {
+	case <-o.stopCh:
+		// Stop requested: kill player and wait for it to exit
+		if o.cmd != nil && o.cmd.Process != nil {
+			o.cmd.Process.Kill()
+		}
+		<-done // Wait for process to actually exit
+		return nil
+	case copyErrVal := <-copyErr:
+		if copyErrVal != nil {
+			// Copy failed: kill player and wait
+			if o.cmd != nil && o.cmd.Process != nil {
+				o.cmd.Process.Kill()
+			}
+			<-done
+			return copyErrVal
+		}
+		// Copy succeeded, now wait for playback to complete
+		return <-done
+	case err := <-done:
+		// Playback finished normally (copy must have succeeded or player would have exited early)
+		return err
+	}
+}
+func (o *KokoroOrator) requestSound(text string) (io.ReadCloser, error) {
+	if o.URL == "" {
+		return nil, fmt.Errorf("TTS URL is empty")
+	}
+	payload := map[string]interface{}{
+		"input":           text,
+		"voice":           o.Voice,
+		"response_format": o.Format,
+		"download_format": o.Format,
+		"stream":          o.Stream,
+		"speed":           o.Speed,
+		// "return_download_link": true,
+		"lang_code": o.Language,
+	}
+	payloadBytes, err := json.Marshal(payload)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal payload: %w", err)
+	}
+	req, err := http.NewRequest("POST", o.URL, bytes.NewBuffer(payloadBytes)) //nolint:noctx
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+	req.Header.Set("accept", "application/json")
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("request failed: %w", err)
+	}
+	if resp.StatusCode != http.StatusOK {
+		defer resp.Body.Close()
+		return nil, fmt.Errorf("unexpected status code: %d", resp.StatusCode)
+	}
+	return resp.Body, nil
+}
+
+func (o *KokoroOrator) stoproutine() {
+	for {
+		<-TTSDoneChan
+		o.logger.Debug("orator got done signal")
+		// 1. Stop any ongoing playback (kills external player, closes stopCh)
+		o.Stop()
+		// 2. Drain any pending text chunks
+		for len(TTSTextChan) > 0 {
+			<-TTSTextChan
+		}
+		// 3. Reset internal state
+		o.mu.Lock()
+		o.textBuffer.Reset()
+		o.interrupt = true
+		o.mu.Unlock()
+	}
+}
+
+func (o *KokoroOrator) Stop() {
+	o.cmdMu.Lock()
+	defer o.cmdMu.Unlock()
+	// Signal any running Speak to stop
+	if o.stopCh != nil {
+		select {
+		case <-o.stopCh: // already closed
+		default:
+			close(o.stopCh)
+		}
+		o.stopCh = nil
+	}
+	// Kill the external player process if it's still running
+	if o.cmd != nil && o.cmd.Process != nil {
+		o.cmd.Process.Kill()
+		o.cmd.Wait() // clean up zombie process
+		o.cmd = nil
+	}
+	// Also reset text buffer and interrupt flag (with o.mu)
+	o.mu.Lock()
+	o.textBuffer.Reset()
+	o.interrupt = true
+	o.mu.Unlock()
+}
+
+func (o *KokoroOrator) readroutine() {
+	tokenizer, _ := english.NewSentenceTokenizer(nil)
+	for {
+		select {
+		case chunk := <-TTSTextChan:
+			o.mu.Lock()
+			o.interrupt = false
+			_, err := o.textBuffer.WriteString(chunk)
+			if err != nil {
+				o.logger.Warn("failed to write to stringbuilder", "error", err)
+				o.mu.Unlock()
+				continue
+			}
+			text := o.textBuffer.String()
+			sentences := tokenizer.Tokenize(text)
+			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
+			if len(sentences) <= 1 {
+				o.mu.Unlock()
+				continue
+			}
+			completeSentences := sentences[:len(sentences)-1]
+			remaining := sentences[len(sentences)-1].Text
+			o.textBuffer.Reset()
+			o.textBuffer.WriteString(remaining)
+			o.mu.Unlock()
+			for _, sentence := range completeSentences {
+				o.mu.Lock()
+				interrupted := o.interrupt
+				o.mu.Unlock()
+				if interrupted {
+					return
+				}
+				cleanedText := models.CleanText(sentence.Text)
+				if cleanedText == "" {
+					continue
+				}
+				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
+				if err := o.Speak(cleanedText); err != nil {
+					o.logger.Error("tts failed", "sentence", cleanedText, "error", err)
+				}
+			}
+		case <-TTSFlushChan:
+			o.logger.Debug("got flushchan signal start")
+			// lln is done get the whole message out
+			if len(TTSTextChan) > 0 { // otherwise might get stuck
+				for chunk := range TTSTextChan {
+					o.mu.Lock()
+					_, err := o.textBuffer.WriteString(chunk)
+					o.mu.Unlock()
+					if err != nil {
+						o.logger.Warn("failed to write to stringbuilder", "error", err)
+						continue
+					}
+					if len(TTSTextChan) == 0 {
+						break
+					}
+				}
+			}
+			// flush remaining text
+			o.mu.Lock()
+			remaining := o.textBuffer.String()
+			remaining = models.CleanText(remaining)
+			o.textBuffer.Reset()
+			o.mu.Unlock()
+			if remaining == "" {
+				continue
+			}
+			o.logger.Debug("calling Speak with remainder", "rem", remaining)
+			sentencesRem := tokenizer.Tokenize(remaining)
+			for _, rs := range sentencesRem { // to avoid dumping large volume of text
+				o.mu.Lock()
+				interrupt := o.interrupt
+				o.mu.Unlock()
+				if interrupt {
+					break
+				}
+				if err := o.Speak(rs.Text); err != nil {
+					o.logger.Error("tts failed", "sentence", rs, "error", err)
+				}
+			}
+		}
+	}
+}
--- a/extra/stt.go
+++ b/extra/stt.go
@@ -6,18 +6,10 @@ package extra
 import (
 	"bytes"
 	"encoding/binary"
-	"errors"
-	"fmt"
 	"gf-lt/config"
 	"io"
 	"log/slog"
-	"mime/multipart"
-	"net/http"
 	"regexp"
-	"strings"
-	"syscall"
-
-	"github.com/gordonklaus/portaudio"
 )

 var specialRE = regexp.MustCompile(`\[.*?\]`)
@@ -44,14 +36,6 @@ func NewSTT(logger *slog.Logger, cfg *config.Config) STT {
 	return NewWhisperServer(logger, cfg)
 }

-type WhisperServer struct {
-	logger      *slog.Logger
-	ServerURL   string
-	SampleRate  int
-	AudioBuffer *bytes.Buffer
-	recording   bool
-}
-
 func NewWhisperServer(logger *slog.Logger, cfg *config.Config) *WhisperServer {
 	return &WhisperServer{
 		logger:      logger,
@@ -61,69 +45,6 @@ func NewWhisperServer(logger *slog.Logger, cfg *config.Config) *WhisperServer {
 	}
 }

-func (stt *WhisperServer) StartRecording() error {
-	if err := stt.microphoneStream(stt.SampleRate); err != nil {
-		return fmt.Errorf("failed to init microphone: %w", err)
-	}
-	stt.recording = true
-	return nil
-}
-
-func (stt *WhisperServer) StopRecording() (string, error) {
-	stt.recording = false
-	// wait loop to finish?
-	if stt.AudioBuffer == nil {
-		err := errors.New("unexpected nil AudioBuffer")
-		stt.logger.Error(err.Error())
-		return "", err
-	}
-	// Create WAV header first
-	body := &bytes.Buffer{}
-	writer := multipart.NewWriter(body)
-	// Add audio file part
-	part, err := writer.CreateFormFile("file", "recording.wav")
-	if err != nil {
-		stt.logger.Error("fn: StopRecording", "error", err)
-		return "", err
-	}
-	// Stream directly to multipart writer: header + raw data
-	dataSize := stt.AudioBuffer.Len()
-	stt.writeWavHeader(part, dataSize)
-	if _, err := io.Copy(part, stt.AudioBuffer); err != nil {
-		stt.logger.Error("fn: StopRecording", "error", err)
-		return "", err
-	}
-	// Reset buffer for next recording
-	stt.AudioBuffer.Reset()
-	// Add response format field
-	err = writer.WriteField("response_format", "text")
-	if err != nil {
-		stt.logger.Error("fn: StopRecording", "error", err)
-		return "", err
-	}
-	if writer.Close() != nil {
-		stt.logger.Error("fn: StopRecording", "error", err)
-		return "", err
-	}
-	// Send request
-	resp, err := http.Post(stt.ServerURL, writer.FormDataContentType(), body) //nolint:noctx
-	if err != nil {
-		stt.logger.Error("fn: StopRecording", "error", err)
-		return "", err
-	}
-	defer resp.Body.Close()
-	// Read and print response
-	responseTextBytes, err := io.ReadAll(resp.Body)
-	if err != nil {
-		stt.logger.Error("fn: StopRecording", "error", err)
-		return "", err
-	}
-	resptext := strings.TrimRight(string(responseTextBytes), "\n")
-	// in case there are special tokens like [_BEG_]
-	resptext = specialRE.ReplaceAllString(resptext, "")
-	return strings.TrimSpace(strings.ReplaceAll(resptext, "\n ", "\n")), nil
-}
-
 func (stt *WhisperServer) writeWavHeader(w io.Writer, dataSize int) {
 	header := make([]byte, 44)
 	copy(header[0:4], "RIFF")
@@ -147,56 +68,3 @@ func (stt *WhisperServer) writeWavHeader(w io.Writer, dataSize int) {
 func (stt *WhisperServer) IsRecording() bool {
 	return stt.recording
 }
-
-func (stt *WhisperServer) microphoneStream(sampleRate int) error {
-	// Temporarily redirect stderr to suppress ALSA warnings during PortAudio init
-	origStderr, errDup := syscall.Dup(syscall.Stderr)
-	if errDup != nil {
-		return fmt.Errorf("failed to dup stderr: %w", errDup)
-	}
-	nullFD, err := syscall.Open("/dev/null", syscall.O_WRONLY, 0)
-	if err != nil {
-		_ = syscall.Close(origStderr) // Close the dup'd fd if open fails
-		return fmt.Errorf("failed to open /dev/null: %w", err)
-	}
-	// redirect stderr
-	_ = syscall.Dup2(nullFD, syscall.Stderr)
-	// Initialize PortAudio (this is where ALSA warnings occur)
-	defer func() {
-		// Restore stderr
-		_ = syscall.Dup2(origStderr, syscall.Stderr)
-		_ = syscall.Close(origStderr)
-		_ = syscall.Close(nullFD)
-	}()
-	if err := portaudio.Initialize(); err != nil {
-		return fmt.Errorf("portaudio init failed: %w", err)
-	}
-	in := make([]int16, 64)
-	stream, err := portaudio.OpenDefaultStream(1, 0, float64(sampleRate), len(in), in)
-	if err != nil {
-		if paErr := portaudio.Terminate(); paErr != nil {
-			return fmt.Errorf("failed to open microphone: %w; terminate error: %w", err, paErr)
-		}
-		return fmt.Errorf("failed to open microphone: %w", err)
-	}
-	go func(stream *portaudio.Stream) {
-		if err := stream.Start(); err != nil {
-			stt.logger.Error("microphoneStream", "error", err)
-			return
-		}
-		for {
-			if !stt.IsRecording() {
-				return
-			}
-			if err := stream.Read(); err != nil {
-				stt.logger.Error("reading stream", "error", err)
-				return
-			}
-			if err := binary.Write(stt.AudioBuffer, binary.LittleEndian, in); err != nil {
-				stt.logger.Error("writing to buffer", "error", err)
-				return
-			}
-		}
-	}(stream)
-	return nil
-}
--- a/extra/tts.go
+++ b/extra/tts.go
@@ -4,25 +4,13 @@
 package extra

 import (
-	"bytes"
-	"encoding/json"
-	"fmt"
 	"gf-lt/config"
 	"gf-lt/models"
-	"io"
 	"log/slog"
-	"net/http"
 	"os"
 	"strings"
-	"sync"
-	"time"

 	google_translate_tts "github.com/GrailFinder/google-translate-tts"
-	"github.com/GrailFinder/google-translate-tts/handlers"
-	"github.com/gopxl/beep/v2"
-	"github.com/gopxl/beep/v2/mp3"
-	"github.com/gopxl/beep/v2/speaker"
-	"github.com/neurosnap/sentences/english"
 )

 var (
@@ -39,142 +27,6 @@ type Orator interface {
 	GetLogger() *slog.Logger
 }

-// impl https://github.com/remsky/Kokoro-FastAPI
-type KokoroOrator struct {
-	logger        *slog.Logger
-	mu            sync.Mutex
-	URL           string
-	Format        models.AudioFormat
-	Stream        bool
-	Speed         float32
-	Language      string
-	Voice         string
-	currentStream *beep.Ctrl // Added for playback control
-	currentDone   chan bool
-	textBuffer    strings.Builder
-	interrupt     bool
-	// textBuffer bytes.Buffer
-}
-
-// Google Translate TTS implementation
-type GoogleTranslateOrator struct {
-	logger        *slog.Logger
-	mu            sync.Mutex
-	speech        *google_translate_tts.Speech
-	currentStream *beep.Ctrl
-	currentDone   chan bool
-	textBuffer    strings.Builder
-	interrupt     bool
-}
-
-func (o *KokoroOrator) stoproutine() {
-	for {
-		<-TTSDoneChan
-		o.logger.Debug("orator got done signal")
-		o.Stop()
-		// drain the channel
-		for len(TTSTextChan) > 0 {
-			<-TTSTextChan
-		}
-		o.mu.Lock()
-		o.textBuffer.Reset()
-		if o.currentDone != nil {
-			select {
-			case o.currentDone <- true:
-			default:
-				// Channel might be closed, ignore
-			}
-		}
-		o.interrupt = true
-		o.mu.Unlock()
-	}
-}
-
-func (o *KokoroOrator) readroutine() {
-	tokenizer, _ := english.NewSentenceTokenizer(nil)
-	for {
-		select {
-		case chunk := <-TTSTextChan:
-			o.mu.Lock()
-			o.interrupt = false
-			_, err := o.textBuffer.WriteString(chunk)
-			if err != nil {
-				o.logger.Warn("failed to write to stringbuilder", "error", err)
-				o.mu.Unlock()
-				continue
-			}
-			text := o.textBuffer.String()
-			sentences := tokenizer.Tokenize(text)
-			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
-			if len(sentences) <= 1 {
-				o.mu.Unlock()
-				continue
-			}
-			completeSentences := sentences[:len(sentences)-1]
-			remaining := sentences[len(sentences)-1].Text
-			o.textBuffer.Reset()
-			o.textBuffer.WriteString(remaining)
-			o.mu.Unlock()
-
-			for _, sentence := range completeSentences {
-				o.mu.Lock()
-				interrupted := o.interrupt
-				o.mu.Unlock()
-				if interrupted {
-					return
-				}
-				cleanedText := models.CleanText(sentence.Text)
-				if cleanedText == "" {
-					continue
-				}
-				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
-				if err := o.Speak(cleanedText); err != nil {
-					o.logger.Error("tts failed", "sentence", cleanedText, "error", err)
-				}
-			}
-		case <-TTSFlushChan:
-			o.logger.Debug("got flushchan signal start")
-			// lln is done get the whole message out
-			if len(TTSTextChan) > 0 { // otherwise might get stuck
-				for chunk := range TTSTextChan {
-					o.mu.Lock()
-					_, err := o.textBuffer.WriteString(chunk)
-					o.mu.Unlock()
-					if err != nil {
-						o.logger.Warn("failed to write to stringbuilder", "error", err)
-						continue
-					}
-					if len(TTSTextChan) == 0 {
-						break
-					}
-				}
-			}
-			// flush remaining text
-			o.mu.Lock()
-			remaining := o.textBuffer.String()
-			remaining = models.CleanText(remaining)
-			o.textBuffer.Reset()
-			o.mu.Unlock()
-			if remaining == "" {
-				continue
-			}
-			o.logger.Debug("calling Speak with remainder", "rem", remaining)
-			sentencesRem := tokenizer.Tokenize(remaining)
-			for _, rs := range sentencesRem { // to avoid dumping large volume of text
-				o.mu.Lock()
-				interrupt := o.interrupt
-				o.mu.Unlock()
-				if interrupt {
-					break
-				}
-				if err := o.Speak(rs.Text); err != nil {
-					o.logger.Error("tts failed", "sentence", rs, "error", err)
-				}
-			}
-		}
-	}
-}
-
 func NewOrator(log *slog.Logger, cfg *config.Config) Orator {
 	provider := cfg.TTS_PROVIDER
 	if provider == "" {
@@ -204,270 +56,14 @@ func NewOrator(log *slog.Logger, cfg *config.Config) Orator {
 			Language: language,
 			Proxy:    "", // Proxy not supported
 			Speed:    cfg.TTS_SPEED,
-			Handler:  &handlers.Beep{},
 		}
 		orator := &GoogleTranslateOrator{
 			logger: log,
 			speech: speech,
+			Speed:  cfg.TTS_SPEED,
 		}
 		go orator.readroutine()
 		go orator.stoproutine()
 		return orator
 	}
 }
-
-func (o *KokoroOrator) GetLogger() *slog.Logger {
-	return o.logger
-}
-
-func (o *KokoroOrator) requestSound(text string) (io.ReadCloser, error) {
-	if o.URL == "" {
-		return nil, fmt.Errorf("TTS URL is empty")
-	}
-	payload := map[string]interface{}{
-		"input":           text,
-		"voice":           o.Voice,
-		"response_format": o.Format,
-		"download_format": o.Format,
-		"stream":          o.Stream,
-		"speed":           o.Speed,
-		// "return_download_link": true,
-		"lang_code": o.Language,
-	}
-	payloadBytes, err := json.Marshal(payload)
-	if err != nil {
-		return nil, fmt.Errorf("failed to marshal payload: %w", err)
-	}
-	req, err := http.NewRequest("POST", o.URL, bytes.NewBuffer(payloadBytes)) //nolint:noctx
-	if err != nil {
-		return nil, fmt.Errorf("failed to create request: %w", err)
-	}
-	req.Header.Set("accept", "application/json")
-	req.Header.Set("Content-Type", "application/json")
-	resp, err := http.DefaultClient.Do(req)
-	if err != nil {
-		return nil, fmt.Errorf("request failed: %w", err)
-	}
-	if resp.StatusCode != http.StatusOK {
-		defer resp.Body.Close()
-		return nil, fmt.Errorf("unexpected status code: %d", resp.StatusCode)
-	}
-	return resp.Body, nil
-}
-
-func (o *KokoroOrator) Speak(text string) error {
-	o.logger.Debug("fn: Speak is called", "text-len", len(text))
-	body, err := o.requestSound(text)
-	if err != nil {
-		o.logger.Error("request failed", "error", err)
-		return fmt.Errorf("request failed: %w", err)
-	}
-	defer body.Close()
-	// Decode the mp3 audio from response body
-	streamer, format, err := mp3.Decode(body)
-	if err != nil {
-		o.logger.Error("mp3 decode failed", "error", err)
-		return fmt.Errorf("mp3 decode failed: %w", err)
-	}
-	defer streamer.Close()
-	// here it spams with errors that speaker cannot be initialized more than once, but how would we deal with many audio records then?
-	if err := speaker.Init(format.SampleRate, format.SampleRate.N(time.Second/10)); err != nil {
-		o.logger.Debug("failed to init speaker", "error", err)
-	}
-	done := make(chan bool)
-	o.mu.Lock()
-	o.currentDone = done
-	o.currentStream = &beep.Ctrl{Streamer: beep.Seq(streamer, beep.Callback(func() {
-		o.mu.Lock()
-		close(done)
-		o.currentStream = nil
-		o.currentDone = nil
-		o.mu.Unlock()
-	})), Paused: false}
-	o.mu.Unlock()
-	speaker.Play(o.currentStream)
-	<-done
-	return nil
-}
-
-func (o *KokoroOrator) Stop() {
-	// speaker.Clear()
-	o.logger.Debug("attempted to stop orator", "orator", o)
-	speaker.Lock()
-	defer speaker.Unlock()
-	o.mu.Lock()
-	defer o.mu.Unlock()
-	if o.currentStream != nil {
-		// o.currentStream.Paused = true
-		o.currentStream.Streamer = nil
-	}
-}
-
-func (o *GoogleTranslateOrator) stoproutine() {
-	for {
-		<-TTSDoneChan
-		o.logger.Debug("orator got done signal")
-		o.Stop()
-		// drain the channel
-		for len(TTSTextChan) > 0 {
-			<-TTSTextChan
-		}
-		o.mu.Lock()
-		o.textBuffer.Reset()
-		if o.currentDone != nil {
-			select {
-			case o.currentDone <- true:
-			default:
-				// Channel might be closed, ignore
-			}
-		}
-		o.interrupt = true
-		o.mu.Unlock()
-	}
-}
-
-func (o *GoogleTranslateOrator) readroutine() {
-	tokenizer, _ := english.NewSentenceTokenizer(nil)
-	for {
-		select {
-		case chunk := <-TTSTextChan:
-			o.mu.Lock()
-			o.interrupt = false
-			_, err := o.textBuffer.WriteString(chunk)
-			if err != nil {
-				o.logger.Warn("failed to write to stringbuilder", "error", err)
-				o.mu.Unlock()
-				continue
-			}
-			text := o.textBuffer.String()
-			sentences := tokenizer.Tokenize(text)
-			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
-			if len(sentences) <= 1 {
-				o.mu.Unlock()
-				continue
-			}
-			completeSentences := sentences[:len(sentences)-1]
-			remaining := sentences[len(sentences)-1].Text
-			o.textBuffer.Reset()
-			o.textBuffer.WriteString(remaining)
-			o.mu.Unlock()
-
-			for _, sentence := range completeSentences {
-				o.mu.Lock()
-				interrupted := o.interrupt
-				o.mu.Unlock()
-				if interrupted {
-					return
-				}
-				cleanedText := models.CleanText(sentence.Text)
-				if cleanedText == "" {
-					continue
-				}
-				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
-				if err := o.Speak(cleanedText); err != nil {
-					o.logger.Error("tts failed", "sentence", cleanedText, "error", err)
-				}
-			}
-		case <-TTSFlushChan:
-			o.logger.Debug("got flushchan signal start")
-			// lln is done get the whole message out
-			if len(TTSTextChan) > 0 { // otherwise might get stuck
-				for chunk := range TTSTextChan {
-					o.mu.Lock()
-					_, err := o.textBuffer.WriteString(chunk)
-					o.mu.Unlock()
-					if err != nil {
-						o.logger.Warn("failed to write to stringbuilder", "error", err)
-						continue
-					}
-					if len(TTSTextChan) == 0 {
-						break
-					}
-				}
-			}
-			o.mu.Lock()
-			remaining := o.textBuffer.String()
-			remaining = models.CleanText(remaining)
-			o.textBuffer.Reset()
-			o.mu.Unlock()
-			if remaining == "" {
-				continue
-			}
-			o.logger.Debug("calling Speak with remainder", "rem", remaining)
-			sentencesRem := tokenizer.Tokenize(remaining)
-			for _, rs := range sentencesRem { // to avoid dumping large volume of text
-				o.mu.Lock()
-				interrupt := o.interrupt
-				o.mu.Unlock()
-				if interrupt {
-					break
-				}
-				if err := o.Speak(rs.Text); err != nil {
-					o.logger.Error("tts failed", "sentence", rs.Text, "error", err)
-				}
-			}
-		}
-	}
-}
-
-func (o *GoogleTranslateOrator) GetLogger() *slog.Logger {
-	return o.logger
-}
-
-func (o *GoogleTranslateOrator) Speak(text string) error {
-	o.logger.Debug("fn: Speak is called", "text-len", len(text))
-	// Generate MP3 data using google-translate-tts
-	reader, err := o.speech.GenerateSpeech(text)
-	if err != nil {
-		o.logger.Error("generate speech failed", "error", err)
-		return fmt.Errorf("generate speech failed: %w", err)
-	}
-	// Decode the mp3 audio from reader (wrap with NopCloser for io.ReadCloser)
-	streamer, format, err := mp3.Decode(io.NopCloser(reader))
-	if err != nil {
-		o.logger.Error("mp3 decode failed", "error", err)
-		return fmt.Errorf("mp3 decode failed: %w", err)
-	}
-	defer streamer.Close()
-	playbackStreamer := beep.Streamer(streamer)
-	speed := o.speech.Speed
-	if speed <= 0 {
-		speed = 1.0
-	}
-	if speed != 1.0 {
-		playbackStreamer = beep.ResampleRatio(3, float64(speed), streamer)
-	}
-	// Initialize speaker with the format's sample rate
-	if err := speaker.Init(format.SampleRate, format.SampleRate.N(time.Second/10)); err != nil {
-		o.logger.Debug("failed to init speaker", "error", err)
-	}
-	done := make(chan bool)
-	o.mu.Lock()
-	o.currentDone = done
-	o.currentStream = &beep.Ctrl{Streamer: beep.Seq(playbackStreamer, beep.Callback(func() {
-		o.mu.Lock()
-		close(done)
-		o.currentStream = nil
-		o.currentDone = nil
-		o.mu.Unlock()
-	})), Paused: false}
-	o.mu.Unlock()
-	speaker.Play(o.currentStream)
-	<-done // wait for playback to complete
-	return nil
-}
-
-func (o *GoogleTranslateOrator) Stop() {
-	o.logger.Debug("attempted to stop google translate orator")
-	speaker.Lock()
-	defer speaker.Unlock()
-	o.mu.Lock()
-	defer o.mu.Unlock()
-	if o.currentStream != nil {
-		o.currentStream.Streamer = nil
-	}
-	// Also stop the speech handler if possible
-	if o.speech != nil {
-		_ = o.speech.Stop()
-	}
-}
--- a/extra/whisper_binary.go
+++ b/extra/whisper_binary.go
@@ -9,15 +9,13 @@ import (
 	"errors"
 	"fmt"
 	"gf-lt/config"
-	"io"
 	"log/slog"
 	"os"
 	"os/exec"
 	"strings"
 	"sync"
 	"syscall"
-
-	"github.com/gordonklaus/portaudio"
+	"time"
 )

 type WhisperBinary struct {
@@ -25,11 +23,143 @@ type WhisperBinary struct {
 	whisperPath string
 	modelPath   string
 	lang        string
-	ctx         context.Context
-	cancel      context.CancelFunc
-	mu          sync.Mutex
-	recording   bool
-	audioBuffer []int16
+	// Per-recording fields (protected by mu)
+	mu        sync.Mutex
+	recording bool
+	tempFile  string
+	ctx       context.Context
+	cancel    context.CancelFunc
+	cmd       *exec.Cmd
+	cmdMu     sync.Mutex
+}
+
+func (w *WhisperBinary) StartRecording() error {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if w.recording {
+		return errors.New("recording is already in progress")
+	}
+	// Fresh context for this recording
+	ctx, cancel := context.WithCancel(context.Background())
+	w.ctx = ctx
+	w.cancel = cancel
+	// Create temporary file
+	tempFile, err := os.CreateTemp("", "recording_*.wav")
+	if err != nil {
+		cancel()
+		return fmt.Errorf("failed to create temp file: %w", err)
+	}
+	tempFile.Close()
+	w.tempFile = tempFile.Name()
+	// ffmpeg command: capture from default microphone, write WAV
+	args := []string{
+		"-f", "alsa", // or "pulse" if preferred
+		"-i", "default",
+		"-acodec", "pcm_s16le",
+		"-ar", "16000",
+		"-ac", "1",
+		"-y", // overwrite output file
+		w.tempFile,
+	}
+	cmd := exec.CommandContext(w.ctx, "ffmpeg", args...)
+	// Capture stderr for debugging (optional, but useful for diagnosing)
+	stderr, err := cmd.StderrPipe()
+	if err != nil {
+		cancel()
+		os.Remove(w.tempFile)
+		return fmt.Errorf("failed to create stderr pipe: %w", err)
+	}
+	go func() {
+		buf := make([]byte, 1024)
+		for {
+			n, err := stderr.Read(buf)
+			if n > 0 {
+				w.logger.Debug("ffmpeg stderr", "output", string(buf[:n]))
+			}
+			if err != nil {
+				break
+			}
+		}
+	}()
+	w.cmdMu.Lock()
+	w.cmd = cmd
+	w.cmdMu.Unlock()
+	if err := cmd.Start(); err != nil {
+		cancel()
+		os.Remove(w.tempFile)
+		return fmt.Errorf("failed to start ffmpeg: %w", err)
+	}
+	w.recording = true
+	w.logger.Debug("Recording started", "file", w.tempFile)
+	return nil
+}
+
+func (w *WhisperBinary) StopRecording() (string, error) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if !w.recording {
+		return "", errors.New("not currently recording")
+	}
+	w.recording = false
+	// Gracefully stop ffmpeg
+	w.cmdMu.Lock()
+	if w.cmd != nil && w.cmd.Process != nil {
+		w.logger.Debug("Sending SIGTERM to ffmpeg")
+		w.cmd.Process.Signal(syscall.SIGTERM)
+		// Wait for process to exit (up to 2 seconds)
+		done := make(chan error, 1)
+		go func() {
+			done <- w.cmd.Wait()
+		}()
+		select {
+		case <-done:
+			w.logger.Debug("ffmpeg exited after SIGTERM")
+		case <-time.After(2 * time.Second):
+			w.logger.Warn("ffmpeg did not exit, sending SIGKILL")
+			w.cmd.Process.Kill()
+			<-done
+		}
+	}
+	w.cmdMu.Unlock()
+	// Cancel context (already done, but for cleanliness)
+	if w.cancel != nil {
+		w.cancel()
+	}
+	// Validate temp file
+	if w.tempFile == "" {
+		return "", errors.New("no recording file")
+	}
+	defer os.Remove(w.tempFile)
+	info, err := os.Stat(w.tempFile)
+	if err != nil {
+		return "", fmt.Errorf("failed to stat temp file: %w", err)
+	}
+	if info.Size() < 44 { // WAV header is 44 bytes
+		// Log ffmpeg stderr? Already captured in debug logs.
+		return "", fmt.Errorf("recording file too small (%d bytes), possibly no audio captured", info.Size())
+	}
+	// Run whisper.cpp binary
+	cmd := exec.Command(w.whisperPath, "-m", w.modelPath, "-l", w.lang, w.tempFile)
+	var outBuf, errBuf bytes.Buffer
+	cmd.Stdout = &outBuf
+	cmd.Stderr = &errBuf
+	if err := cmd.Run(); err != nil {
+		w.logger.Error("whisper binary failed",
+			"error", err,
+			"stderr", errBuf.String(),
+			"file_size", info.Size())
+		return "", fmt.Errorf("whisper binary failed: %w (stderr: %s)", err, errBuf.String())
+	}
+	result := strings.TrimRight(outBuf.String(), "\n")
+	result = specialRE.ReplaceAllString(result, "")
+	return strings.TrimSpace(strings.ReplaceAll(result, "\n ", "\n")), nil
+}
+
+// IsRecording returns true if a recording is in progress.
+func (w *WhisperBinary) IsRecording() bool {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	return w.recording
 }

 func NewWhisperBinary(logger *slog.Logger, cfg *config.Config) *WhisperBinary {
@@ -44,283 +174,3 @@ func NewWhisperBinary(logger *slog.Logger, cfg *config.Config) *WhisperBinary {
 		cancel:      cancel,
 	}
 }
-
-func (w *WhisperBinary) StartRecording() error {
-	w.mu.Lock()
-	defer w.mu.Unlock()
-	if w.recording {
-		return errors.New("recording is already in progress")
-	}
-	// If context is cancelled, create a new one for the next recording session
-	if w.ctx.Err() != nil {
-		w.logger.Debug("Context cancelled, creating new context")
-		w.ctx, w.cancel = context.WithCancel(context.Background())
-	}
-	// Temporarily redirect stderr to suppress ALSA warnings during PortAudio init
-	origStderr, errDup := syscall.Dup(syscall.Stderr)
-	if errDup != nil {
-		return fmt.Errorf("failed to dup stderr: %w", errDup)
-	}
-	nullFD, err := syscall.Open("/dev/null", syscall.O_WRONLY, 0)
-	if err != nil {
-		_ = syscall.Close(origStderr) // Close the dup'd fd if open fails
-		return fmt.Errorf("failed to open /dev/null: %w", err)
-	}
-	// redirect stderr
-	_ = syscall.Dup2(nullFD, syscall.Stderr)
-	// Initialize PortAudio (this is where ALSA warnings occur)
-	portaudioErr := portaudio.Initialize()
-	defer func() {
-		// Restore stderr
-		_ = syscall.Dup2(origStderr, syscall.Stderr)
-		_ = syscall.Close(origStderr)
-		_ = syscall.Close(nullFD)
-	}()
-	if portaudioErr != nil {
-		return fmt.Errorf("portaudio init failed: %w", portaudioErr)
-	}
-	// Initialize audio buffer
-	w.audioBuffer = make([]int16, 0)
-	in := make([]int16, 1024) // buffer size
-	stream, err := portaudio.OpenDefaultStream(1, 0, 16000.0, len(in), in)
-	if err != nil {
-		if paErr := portaudio.Terminate(); paErr != nil {
-			return fmt.Errorf("failed to open microphone: %w; terminate error: %w", err, paErr)
-		}
-		return fmt.Errorf("failed to open microphone: %w", err)
-	}
-	go w.recordAudio(stream, in)
-	w.recording = true
-	w.logger.Debug("Recording started")
-	return nil
-}
-
-func (w *WhisperBinary) recordAudio(stream *portaudio.Stream, in []int16) {
-	defer func() {
-		w.logger.Debug("recordAudio defer function called")
-		_ = stream.Stop()         // Stop the stream
-		_ = portaudio.Terminate() // ignoring error as we're shutting down
-		w.logger.Debug("recordAudio terminated")
-	}()
-	w.logger.Debug("Starting audio stream")
-	if err := stream.Start(); err != nil {
-		w.logger.Error("Failed to start audio stream", "error", err)
-		return
-	}
-	w.logger.Debug("Audio stream started, entering recording loop")
-	for {
-		select {
-		case <-w.ctx.Done():
-			w.logger.Debug("Context done, exiting recording loop")
-			return
-		default:
-			// Check recording status with minimal lock time
-			w.mu.Lock()
-			recording := w.recording
-			w.mu.Unlock()
-
-			if !recording {
-				w.logger.Debug("Recording flag is false, exiting recording loop")
-				return
-			}
-			if err := stream.Read(); err != nil {
-				w.logger.Error("Error reading from stream", "error", err)
-				return
-			}
-			// Append samples to buffer - only acquire lock when necessary
-			w.mu.Lock()
-			if w.audioBuffer == nil {
-				w.audioBuffer = make([]int16, 0)
-			}
-			// Make a copy of the input buffer to avoid overwriting
-			tempBuffer := make([]int16, len(in))
-			copy(tempBuffer, in)
-			w.audioBuffer = append(w.audioBuffer, tempBuffer...)
-			w.mu.Unlock()
-		}
-	}
-}
-
-func (w *WhisperBinary) StopRecording() (string, error) {
-	w.logger.Debug("StopRecording called")
-	w.mu.Lock()
-	if !w.recording {
-		w.mu.Unlock()
-		return "", errors.New("not currently recording")
-	}
-	w.logger.Debug("Setting recording to false and cancelling context")
-	w.recording = false
-	w.cancel() // This will stop the recording goroutine
-	w.mu.Unlock()
-	// // Small delay to allow the recording goroutine to react to context cancellation
-	// time.Sleep(20 * time.Millisecond)
-	// Save the recorded audio to a temporary file
-	tempFile, err := w.saveAudioToTempFile()
-	if err != nil {
-		w.logger.Error("Error saving audio to temp file", "error", err)
-		return "", fmt.Errorf("failed to save audio to temp file: %w", err)
-	}
-	w.logger.Debug("Saved audio to temp file", "file", tempFile)
-	// Run the whisper binary with a separate context to avoid cancellation during transcription
-	cmd := exec.Command(w.whisperPath, "-m", w.modelPath, "-l", w.lang, tempFile, "2>/dev/null")
-	var outBuf bytes.Buffer
-	cmd.Stdout = &outBuf
-	// Redirect stderr to suppress ALSA warnings and other stderr output
-	cmd.Stderr = io.Discard // Suppress stderr output from whisper binary
-	w.logger.Debug("Running whisper binary command")
-	if err := cmd.Run(); err != nil {
-		// Clean up audio buffer
-		w.mu.Lock()
-		w.audioBuffer = nil
-		w.mu.Unlock()
-		// Since we're suppressing stderr, we'll just log that the command failed
-		w.logger.Error("Error running whisper binary", "error", err)
-		return "", fmt.Errorf("whisper binary failed: %w", err)
-	}
-	result := outBuf.String()
-	w.logger.Debug("Whisper binary completed", "result", result)
-	// Clean up audio buffer
-	w.mu.Lock()
-	w.audioBuffer = nil
-	w.mu.Unlock()
-	// Clean up the temporary file after transcription
-	w.logger.Debug("StopRecording completed")
-	os.Remove(tempFile)
-	result = strings.TrimRight(result, "\n")
-	// in case there are special tokens like [_BEG_]
-	result = specialRE.ReplaceAllString(result, "")
-	return strings.TrimSpace(strings.ReplaceAll(result, "\n ", "\n")), nil
-}
-
-// saveAudioToTempFile saves the recorded audio data to a temporary WAV file
-func (w *WhisperBinary) saveAudioToTempFile() (string, error) {
-	w.logger.Debug("saveAudioToTempFile called")
-	// Create temporary WAV file
-	tempFile, err := os.CreateTemp("", "recording_*.wav")
-	if err != nil {
-		w.logger.Error("Failed to create temp file", "error", err)
-		return "", fmt.Errorf("failed to create temp file: %w", err)
-	}
-	w.logger.Debug("Created temp file", "file", tempFile.Name())
-	defer tempFile.Close()
-
-	// Write WAV header and data
-	w.logger.Debug("About to write WAV file", "file", tempFile.Name())
-	err = w.writeWAVFile(tempFile.Name())
-	if err != nil {
-		w.logger.Error("Error writing WAV file", "error", err)
-		return "", fmt.Errorf("failed to write WAV file: %w", err)
-	}
-	w.logger.Debug("WAV file written successfully", "file", tempFile.Name())
-
-	return tempFile.Name(), nil
-}
-
-// writeWAVFile creates a WAV file from the recorded audio data
-func (w *WhisperBinary) writeWAVFile(filename string) error {
-	w.logger.Debug("writeWAVFile called", "filename", filename)
-	// Open file for writing
-	file, err := os.Create(filename)
-	if err != nil {
-		w.logger.Error("Error creating file", "error", err)
-		return err
-	}
-	defer file.Close()
-
-	w.logger.Debug("About to acquire mutex in writeWAVFile")
-	w.mu.Lock()
-	w.logger.Debug("Locked mutex, copying audio buffer")
-	audioData := make([]int16, len(w.audioBuffer))
-	copy(audioData, w.audioBuffer)
-	w.mu.Unlock()
-	w.logger.Debug("Unlocked mutex", "audio_data_length", len(audioData))
-
-	if len(audioData) == 0 {
-		w.logger.Warn("No audio data to write")
-		return errors.New("no audio data to write")
-	}
-
-	// Calculate data size (number of samples * size of int16)
-	dataSize := len(audioData) * 2 // 2 bytes per int16 sample
-	w.logger.Debug("Calculated data size", "size", dataSize)
-
-	// Write WAV header with the correct data size
-	header := w.createWAVHeader(16000, 1, 16, dataSize)
-	_, err = file.Write(header)
-	if err != nil {
-		w.logger.Error("Error writing WAV header", "error", err)
-		return err
-	}
-	w.logger.Debug("WAV header written successfully")
-
-	// Write audio data
-	w.logger.Debug("About to write audio data samples")
-	for i, sample := range audioData {
-		// Write little-endian 16-bit sample
-		_, err := file.Write([]byte{byte(sample), byte(sample >> 8)})
-		if err != nil {
-			w.logger.Error("Error writing sample", "index", i, "error", err)
-			return err
-		}
-		// Log progress every 10000 samples to avoid too much output
-		if i%10000 == 0 {
-			w.logger.Debug("Written samples", "count", i)
-		}
-	}
-	w.logger.Debug("All audio data written successfully")
-
-	return nil
-}
-
-// createWAVHeader creates a WAV file header
-func (w *WhisperBinary) createWAVHeader(sampleRate, channels, bitsPerSample int, dataSize int) []byte {
-	header := make([]byte, 44)
-	copy(header[0:4], "RIFF")
-	// Total file size will be updated later
-	copy(header[8:12], "WAVE")
-	copy(header[12:16], "fmt ")
-	// fmt chunk size (16 for PCM)
-	header[16] = 16
-	header[17] = 0
-	header[18] = 0
-	header[19] = 0
-	// Audio format (1 = PCM)
-	header[20] = 1
-	header[21] = 0
-	// Number of channels
-	header[22] = byte(channels)
-	header[23] = 0
-	// Sample rate
-	header[24] = byte(sampleRate)
-	header[25] = byte(sampleRate >> 8)
-	header[26] = byte(sampleRate >> 16)
-	header[27] = byte(sampleRate >> 24)
-	// Byte rate
-	byteRate := sampleRate * channels * bitsPerSample / 8
-	header[28] = byte(byteRate)
-	header[29] = byte(byteRate >> 8)
-	header[30] = byte(byteRate >> 16)
-	header[31] = byte(byteRate >> 24)
-	// Block align
-	blockAlign := channels * bitsPerSample / 8
-	header[32] = byte(blockAlign)
-	header[33] = 0
-	// Bits per sample
-	header[34] = byte(bitsPerSample)
-	header[35] = 0
-	// "data" subchunk
-	copy(header[36:40], "data")
-	// Data size
-	header[40] = byte(dataSize)
-	header[41] = byte(dataSize >> 8)
-	header[42] = byte(dataSize >> 16)
-	header[43] = byte(dataSize >> 24)
-
-	return header
-}
-
-func (w *WhisperBinary) IsRecording() bool {
-	w.mu.Lock()
-	defer w.mu.Unlock()
-	return w.recording
-}
--- a/extra/whisper_server.go
+++ b/extra/whisper_server.go
@@ -0,0 +1,156 @@
+//go:build extra
+// +build extra
+
+package extra
+
+import (
+	"bytes"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"mime/multipart"
+	"net/http"
+	"os/exec"
+	"strings"
+	"sync"
+)
+
+type WhisperServer struct {
+	logger      *slog.Logger
+	ServerURL   string
+	SampleRate  int
+	AudioBuffer *bytes.Buffer
+	recording   bool          // protected by mu
+	mu          sync.Mutex    // protects recording & AudioBuffer
+	cmd         *exec.Cmd     // protected by cmdMu
+	stopCh      chan struct{} // protected by cmdMu
+	cmdMu       sync.Mutex    // protects cmd and stopCh
+}
+
+func (stt *WhisperServer) StartRecording() error {
+	stt.mu.Lock()
+	defer stt.mu.Unlock()
+	if stt.recording {
+		return nil
+	}
+	// Build ffmpeg command for microphone capture
+	args := []string{
+		"-f", "alsa",
+		"-i", "default",
+		"-acodec", "pcm_s16le",
+		"-ar", fmt.Sprint(stt.SampleRate),
+		"-ac", "1",
+		"-f", "s16le",
+		"-",
+	}
+	cmd := exec.Command("ffmpeg", args...)
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		return fmt.Errorf("failed to get stdout pipe: %w", err)
+	}
+	stt.cmdMu.Lock()
+	stt.cmd = cmd
+	stt.stopCh = make(chan struct{})
+	stt.cmdMu.Unlock()
+	if err := cmd.Start(); err != nil {
+		return fmt.Errorf("failed to start ffmpeg: %w", err)
+	}
+	stt.recording = true
+	stt.AudioBuffer.Reset()
+	// Read PCM data in goroutine
+	go func() {
+		buf := make([]byte, 4096)
+		for {
+			select {
+			case <-stt.stopCh:
+				return
+			default:
+				n, err := stdout.Read(buf)
+				if n > 0 {
+					stt.mu.Lock()
+					stt.AudioBuffer.Write(buf[:n])
+					stt.mu.Unlock()
+				}
+				if err != nil {
+					if err != io.EOF {
+						stt.logger.Error("recording read error", "error", err)
+					}
+					return
+				}
+			}
+		}
+	}()
+	return nil
+}
+
+func (stt *WhisperServer) StopRecording() (string, error) {
+	stt.mu.Lock()
+	defer stt.mu.Unlock()
+	if !stt.recording {
+		return "", errors.New("not recording")
+	}
+	stt.recording = false
+	// Stop ffmpeg
+	stt.cmdMu.Lock()
+	if stt.cmd != nil && stt.cmd.Process != nil {
+		stt.cmd.Process.Kill()
+		stt.cmd.Wait()
+	}
+	close(stt.stopCh)
+	stt.cmdMu.Unlock()
+	// Rest of StopRecording unchanged (WAV header + HTTP upload)
+	// ...
+	stt.recording = false
+	// wait loop to finish?
+	if stt.AudioBuffer == nil {
+		err := errors.New("unexpected nil AudioBuffer")
+		stt.logger.Error(err.Error())
+		return "", err
+	}
+	// Create WAV header first
+	body := &bytes.Buffer{}
+	writer := multipart.NewWriter(body)
+	// Add audio file part
+	part, err := writer.CreateFormFile("file", "recording.wav")
+	if err != nil {
+		stt.logger.Error("fn: StopRecording", "error", err)
+		return "", err
+	}
+	// Stream directly to multipart writer: header + raw data
+	dataSize := stt.AudioBuffer.Len()
+	stt.writeWavHeader(part, dataSize)
+	if _, err := io.Copy(part, stt.AudioBuffer); err != nil {
+		stt.logger.Error("fn: StopRecording", "error", err)
+		return "", err
+	}
+	// Reset buffer for next recording
+	stt.AudioBuffer.Reset()
+	// Add response format field
+	err = writer.WriteField("response_format", "text")
+	if err != nil {
+		stt.logger.Error("fn: StopRecording", "error", err)
+		return "", err
+	}
+	if writer.Close() != nil {
+		stt.logger.Error("fn: StopRecording", "error", err)
+		return "", err
+	}
+	// Send request
+	resp, err := http.Post(stt.ServerURL, writer.FormDataContentType(), body) //nolint:noctx
+	if err != nil {
+		stt.logger.Error("fn: StopRecording", "error", err)
+		return "", err
+	}
+	defer resp.Body.Close()
+	// Read and print response
+	responseTextBytes, err := io.ReadAll(resp.Body)
+	if err != nil {
+		stt.logger.Error("fn: StopRecording", "error", err)
+		return "", err
+	}
+	resptext := strings.TrimRight(string(responseTextBytes), "\n")
+	// in case there are special tokens like [_BEG_]
+	resptext = specialRE.ReplaceAllString(resptext, "")
+	return strings.TrimSpace(strings.ReplaceAll(resptext, "\n ", "\n")), nil
+}
--- a/go.mod
+++ b/go.mod
@@ -4,39 +4,40 @@ go 1.25.1

 require (
 	github.com/BurntSushi/toml v1.5.0
-	github.com/GrailFinder/google-translate-tts v0.1.3
+	github.com/GrailFinder/google-translate-tts v0.1.4
 	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
-	github.com/gordonklaus/portaudio v0.0.0-20250206071425-98a94950218b
 	github.com/jmoiron/sqlx v1.4.0
 	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
 	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
 	github.com/google/uuid v1.6.0 // indirect
 	github.com/hajimehoshi/go-mp3 v0.3.4 // indirect
-	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/pkg/errors v0.9.1 // indirect
+	github.com/patrickmn/go-cache v2.1.0+incompatible // 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
--- a/go.sum
+++ b/go.sum
@@ -2,8 +2,8 @@ filippo.io/edwards25519 v1.1.0 h1:FNf4tywRC1HmFuKW5xopWpigGjJKiJSV0Cqo0cJWDaA=
 filippo.io/edwards25519 v1.1.0/go.mod h1:BxyFTGdWcka3PhytdK4V28tE5sGfRvvvRV7EaN4VDT4=
 github.com/BurntSushi/toml v1.5.0 h1:W5quZX/G/csjUnuI8SUYlsHs9M38FC7znL0lIO+DvMg=
 github.com/BurntSushi/toml v1.5.0/go.mod h1:ukJfTF/6rtPPRCnwkur4qwRxa8vTRFBF0uk2lLoLwho=
-github.com/GrailFinder/google-translate-tts v0.1.3 h1:Mww9tNzTWjjSh+OCbTPl/+21oMPKcUecXZfU7nTB/lA=
-github.com/GrailFinder/google-translate-tts v0.1.3/go.mod h1:YIOLKR7sObazdUCrSex3u9OVBovU55eYgWa25vsQJ18=
+github.com/GrailFinder/google-translate-tts v0.1.4 h1:NJoPZUGfBrmouQMN19MUcNPNUx4tmf4a8OZRME4E4Mg=
+github.com/GrailFinder/google-translate-tts v0.1.4/go.mod h1:YIOLKR7sObazdUCrSex3u9OVBovU55eYgWa25vsQJ18=
 github.com/GrailFinder/searchagent v0.2.0 h1:U2GVjLh/9xZt0xX9OcYk9Q2fMkyzyTiADPUmUisRdtQ=
 github.com/GrailFinder/searchagent v0.2.0/go.mod h1:d66tn5+22LI8IGJREUsRBT60P0sFdgQgvQRqyvgItrs=
 github.com/PuerkitoBio/goquery v1.11.0 h1:jZ7pwMQXIITcUXNH83LLk+txlaEy6NVOfTuP43xxfqw=
@@ -17,10 +17,8 @@ github.com/deckarep/golang-set/v2 v2.8.0 h1:swm0rlPCmdWn9mESxKOjWk8hXSqoxOp+Zlfu
 github.com/deckarep/golang-set/v2 v2.8.0/go.mod h1:VAky9rY/yGXJOLEDv3OMci+7wtDpOF4IN+y82NBOac4=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
-github.com/ebitengine/oto/v3 v3.4.0 h1:br0PgASsEWaoWn38b2Goe7m1GKFYfNgnsjSd5Gg+/bQ=
-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=
@@ -39,13 +37,8 @@ github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e h1:ijClszYn+mADRFY17k
 github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e/go.mod h1:boTsfXsheKC2y+lKOCMpSfarhxDeIzfZG1jqGcPl3cA=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
-github.com/gopxl/beep/v2 v2.1.1 h1:6FYIYMm2qPAdWkjX+7xwKrViS1x0Po5kDMdRkq8NVbU=
-github.com/gopxl/beep/v2 v2.1.1/go.mod h1:ZAm9TGQ9lvpoiFLd4zf5B1IuyxZhgRACMId1XJbaW0E=
-github.com/gordonklaus/portaudio v0.0.0-20250206071425-98a94950218b h1:WEuQWBxelOGHA6z9lABqaMLMrfwVyMdN3UgRLT+YUPo=
-github.com/gordonklaus/portaudio v0.0.0-20250206071425-98a94950218b/go.mod h1:esZFQEUwqC+l76f2R8bIWSwXMaPbp79PppwZ1eJhFco=
 github.com/hajimehoshi/go-mp3 v0.3.4 h1:NUP7pBYH8OguP4diaTZ9wJbUbk3tC0KlfzsEpWmYj68=
 github.com/hajimehoshi/go-mp3 v0.3.4/go.mod h1:fRtZraRFcWb0pu7ok0LqyFhCUrPeMsGRSVop0eemFmo=
-github.com/hajimehoshi/oto/v2 v2.3.1 h1:qrLKpNus2UfD674oxckKjNJmesp9hMh7u7QCrStB3Rc=
 github.com/hajimehoshi/oto/v2 v2.3.1/go.mod h1:seWLbgHH7AyUMYKfKYT9pg7PhUu9/SisyJvNTT+ASQo=
 github.com/hashicorp/golang-lru/v2 v2.0.7 h1:a+bsQ5rvGLjzHuww6tVxozPZFVghXaHOwFs4luLUK2k=
 github.com/hashicorp/golang-lru/v2 v2.0.7/go.mod h1:QeFd9opnmA6QUJc5vARoKUSoFhyfM2/ZepoAG6RGpeM=
@@ -61,12 +54,14 @@ 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/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
-github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
+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/playwright-community/playwright-go v0.5700.1 h1:PNFb1byWqrTT720rEO0JL88C6Ju0EmUnR5deFLvtP/U=
 github.com/playwright-community/playwright-go v0.5700.1/go.mod h1:MlSn1dZrx8rszbCxY6x3qK89ZesJUYVx21B2JnkoNF0=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
@@ -77,10 +72,19 @@ 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=
--- a/helpfuncs.go
+++ b/helpfuncs.go
@@ -5,7 +5,6 @@ import (
 	"gf-lt/models"
 	"gf-lt/pngmeta"
 	"image"
-	"net/url"
 	"os"
 	"os/exec"
 	"path"
@@ -13,6 +12,7 @@ import (
 	"slices"
 	"strconv"
 	"strings"
+	"sync/atomic"
 	"time"
 	"unicode"

@@ -20,7 +20,8 @@ import (
 )

 // Cached model color - updated by background goroutine
-var cachedModelColor string = "orange"
+// var cachedModelColor string = "orange"
+var cachedModelColor atomic.Value

 // startModelColorUpdater starts a background goroutine that periodically updates
 // the cached model color. Only runs HTTP requests for local llama.cpp APIs.
@@ -39,20 +40,20 @@ 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)
 	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")
 	}
 }

@@ -108,7 +109,7 @@ func refreshChatDisplay() {
 	textView.SetText(displayText)
 	colorText()
 	updateStatusLine()
-	if scrollToEndEnabled {
+	if cfg.AutoScrollEnabled {
 		textView.ScrollToEnd()
 	}
 }
@@ -323,19 +324,17 @@ func strInSlice(s string, sl []string) bool {

 // isLocalLlamacpp checks if the current API is a local llama.cpp instance.
 func isLocalLlamacpp() bool {
-	u, err := url.Parse(cfg.CurrentAPI)
-	if err != nil {
+	if strings.Contains(cfg.CurrentAPI, "openrouter") || strings.Contains(cfg.CurrentAPI, "deepseek") {
 		return false
 	}
-	host := u.Hostname()
-	return host == "localhost" || host == "127.0.0.1" || host == "::1"
+	return true
 }

 // getModelColor returns the cached color tag for the model name.
 // 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 {
@@ -521,7 +520,7 @@ 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)
@@ -542,7 +541,7 @@ func executeCommandAndDisplay(cmdText string) {
 	cmdText = strings.TrimSpace(cmdText)
 	if cmdText == "" {
 		fmt.Fprintf(textView, "\n[red]Error: No command provided[-:-:-]\n")
-		if scrollToEndEnabled {
+		if cfg.AutoScrollEnabled {
 			textView.ScrollToEnd()
 		}
 		colorText()
@@ -574,7 +573,7 @@ func executeCommandAndDisplay(cmdText string) {
 				Content: "$ " + cmdText + "\n\n" + outputContent,
 			}
 			chatBody.Messages = append(chatBody.Messages, combinedMsg)
-			if scrollToEndEnabled {
+			if cfg.AutoScrollEnabled {
 				textView.ScrollToEnd()
 			}
 			colorText()
@@ -589,7 +588,7 @@ func executeCommandAndDisplay(cmdText string) {
 				Content: "$ " + cmdText + "\n\n" + outputContent,
 			}
 			chatBody.Messages = append(chatBody.Messages, combinedMsg)
-			if scrollToEndEnabled {
+			if cfg.AutoScrollEnabled {
 				textView.ScrollToEnd()
 			}
 			colorText()
@@ -637,7 +636,7 @@ func executeCommandAndDisplay(cmdText string) {
 	}
 	chatBody.Messages = append(chatBody.Messages, combinedMsg)
 	// Scroll to end and update colors
-	if scrollToEndEnabled {
+	if cfg.AutoScrollEnabled {
 		textView.ScrollToEnd()
 	}
 	colorText()
@@ -967,3 +966,52 @@ func extractDisplayPath(p, bp string) string {
 	}
 	return p
 }
+
+func getValidKnowToRecipient(msg *models.RoleMsg) (string, bool) {
+	if cfg == nil || !cfg.CharSpecificContextEnabled {
+		return "", false
+	}
+	// case where all roles are in the tag => public message
+	cr := listChatRoles()
+	slices.Sort(cr)
+	slices.Sort(msg.KnownTo)
+	if slices.Equal(cr, msg.KnownTo) {
+		logger.Info("got msg with tag mentioning every role")
+		return "", false
+	}
+	// Check each character in the KnownTo list
+	for _, recipient := range msg.KnownTo {
+		if recipient == msg.Role || recipient == cfg.ToolRole {
+			// weird cases, skip
+			continue
+		}
+		// Skip if this is the user character (user handles their own turn)
+		// If user is in KnownTo, stop processing - it's the user's turn
+		if recipient == cfg.UserRole || recipient == cfg.WriteNextMsgAs {
+			return "", false
+		}
+		return recipient, true
+	}
+	return "", false
+}
+
+// triggerPrivateMessageResponses checks if a message was sent privately to specific characters
+// and triggers those non-user characters to respond
+func triggerPrivateMessageResponses(msg *models.RoleMsg) {
+	recipient, ok := getValidKnowToRecipient(msg)
+	if !ok || recipient == "" {
+		return
+	}
+	// Trigger the recipient character to respond
+	triggerMsg := recipient + ":\n"
+	// Send empty message so LLM continues naturally from the conversation
+	crr := &models.ChatRoundReq{
+		UserMsg: triggerMsg,
+		Role:    recipient,
+		Resume:  true,
+	}
+	fmt.Fprintf(textView, "\n[-:-:b](%d) ", len(chatBody.Messages))
+	fmt.Fprint(textView, roleToIcon(recipient))
+	fmt.Fprint(textView, "[-:-:-]\n")
+	chatRoundChan <- crr
+}
--- a/llm.go
+++ b/llm.go
@@ -62,11 +62,11 @@ type ChunkParser interface {
 func choseChunkParser() {
 	chunkParser = LCPCompletion{}
 	switch cfg.CurrentAPI {
-	case "http://localhost:8080/completion":
+	case "http://localhost:8080/completion", "http://127.0.0.1:8080/completion":
 		chunkParser = LCPCompletion{}
 		logger.Debug("chosen lcpcompletion", "link", cfg.CurrentAPI)
 		return
-	case "http://localhost:8080/v1/chat/completions":
+	case "http://localhost:8080/v1/chat/completions", "http://127.0.0.1:8080/v1/chat/completions":
 		chunkParser = LCPChat{}
 		logger.Debug("chosen lcpchat", "link", cfg.CurrentAPI)
 		return
@@ -87,6 +87,11 @@ func choseChunkParser() {
 		logger.Debug("chosen openrouterchat", "link", cfg.CurrentAPI)
 		return
 	default:
+		logger.Warn("unexpected case, assuming llama.cpp on non default address", "link", cfg.CurrentAPI)
+		if strings.Contains(cfg.CurrentAPI, "chat") {
+			chunkParser = LCPChat{}
+			return
+		}
 		chunkParser = LCPCompletion{}
 	}
 }
--- a/main.go
+++ b/main.go
@@ -1,13 +1,15 @@
 package main

 import (
+	"sync/atomic"
+
 	"github.com/rivo/tview"
 )

 var (
 	boolColors        = map[bool]string{true: "green", false: "red"}
-	botRespMode       = false
-	toolRunningMode   = false
+	botRespMode       atomic.Bool
+	toolRunningMode   atomic.Bool
 	editMode          = false
 	roleEditMode      = false
 	injectRole        = true
--- a/props_table.go
+++ b/props_table.go
@@ -115,18 +115,12 @@ 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
 	})
 	addCheckboxRow("TTS Enabled", cfg.TTS_ENABLED, func(checked bool) {
 		cfg.TTS_ENABLED = checked
 	})
-	addCheckboxRow("Auto clean tool calls from context", cfg.AutoCleanToolCallsFromCtx, func(checked bool) {
-		cfg.AutoCleanToolCallsFromCtx = checked
-	})
 	addCheckboxRow("Enable Mouse", cfg.EnableMouse, func(checked bool) {
 		cfg.EnableMouse = checked
 		// Reconfigure the app's mouse setting
--- a/rag/embedder.go
+++ b/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
@@ -27,8 +34,10 @@ type APIEmbedder struct {
 func NewAPIEmbedder(l *slog.Logger, cfg *config.Config) *APIEmbedder {
 	return &APIEmbedder{
 		logger: l,
-		client: &http.Client{},
-		cfg:    cfg,
+		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
-//
-// For now, we'll focus on the API implementation which is already working in the current system,
-// and can be extended later when we have ONNX runtime integration
+type ONNXEmbedder struct {
+	session       *onnxruntime_go.DynamicAdvancedSession
+	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
+}
--- a/rag/rag.go
+++ b/rag/rag.go
--- a/rag/rag_integration_test.go
+++ b/rag/rag_integration_test.go
@@ -0,0 +1,409 @@
+package rag
+
+import (
+	"fmt"
+	"gf-lt/config"
+	"gf-lt/models"
+	"gf-lt/storage"
+	"log/slog"
+	"testing"
+
+	_ "github.com/glebarez/go-sqlite"
+	"github.com/jmoiron/sqlx"
+)
+
+// mockEmbedder returns zero vectors of a fixed dimension.
+type mockEmbedder struct {
+	dim int
+}
+
+func (m *mockEmbedder) Embed(text string) ([]float32, error) {
+	vec := make([]float32, m.dim)
+	return vec, nil
+}
+
+func (m *mockEmbedder) EmbedSlice(texts []string) ([][]float32, error) {
+	vecs := make([][]float32, len(texts))
+	for i := range vecs {
+		vecs[i] = make([]float32, m.dim)
+	}
+	return vecs, nil
+}
+
+// dummyStore implements storage.FullRepo with a minimal set of methods.
+// Only DB() is used by VectorStorage; other methods return empty values.
+type dummyStore struct {
+	db *sqlx.DB
+}
+
+func (d dummyStore) DB() *sqlx.DB { return d.db }
+
+// ChatHistory methods
+func (d dummyStore) ListChats() ([]models.Chat, error)                     { return nil, nil }
+func (d dummyStore) GetChatByID(id uint32) (*models.Chat, error)           { return nil, nil }
+func (d dummyStore) GetChatByChar(char string) ([]models.Chat, error)      { return nil, nil }
+func (d dummyStore) GetLastChat() (*models.Chat, error)                    { return nil, nil }
+func (d dummyStore) GetLastChatByAgent(agent string) (*models.Chat, error) { return nil, nil }
+func (d dummyStore) UpsertChat(chat *models.Chat) (*models.Chat, error)    { return chat, nil }
+func (d dummyStore) RemoveChat(id uint32) error                            { return nil }
+func (d dummyStore) ChatGetMaxID() (uint32, error)                         { return 0, nil }
+
+// Memories methods
+func (d dummyStore) Memorise(m *models.Memory) (*models.Memory, error) { return m, nil }
+func (d dummyStore) Recall(agent, topic string) (string, error)        { return "", nil }
+func (d dummyStore) RecallTopics(agent string) ([]string, error)       { return nil, nil }
+
+// VectorRepo methods (not used but required by interface)
+func (d dummyStore) WriteVector(row *models.VectorRow) error { return nil }
+func (d dummyStore) SearchClosest(q []float32, limit int) ([]models.VectorRow, error) {
+	return nil, nil
+}
+func (d dummyStore) ListFiles() ([]string, error)              { return nil, nil }
+func (d dummyStore) RemoveEmbByFileName(filename string) error { return nil }
+
+var _ storage.FullRepo = dummyStore{}
+
+// setupTestRAG creates an in‑memory SQLite database, creates the necessary tables,
+// inserts the provided chunks, and returns a RAG instance with a mock embedder.
+func setupTestRAG(t *testing.T, chunks []*models.VectorRow) (*RAG, error) {
+	t.Helper()
+	db, err := sqlx.Open("sqlite", ":memory:")
+	if err != nil {
+		return nil, fmt.Errorf("open in‑memory db: %w", err)
+	}
+	// Create the required tables (embeddings_768 and fts_embeddings).
+	// Use the same schema as production.
+	_, err = db.Exec(`
+		CREATE TABLE embeddings_768 (
+			id INTEGER PRIMARY KEY AUTOINCREMENT,
+			embeddings BLOB NOT NULL,
+			slug TEXT NOT NULL,
+			raw_text TEXT NOT NULL,
+			filename TEXT NOT NULL DEFAULT ''
+		);
+	`)
+	if err != nil {
+		return nil, fmt.Errorf("create embeddings table: %w", err)
+	}
+	_, err = db.Exec(`
+		CREATE VIRTUAL TABLE fts_embeddings USING fts5(
+			slug UNINDEXED,
+			raw_text,
+			filename UNINDEXED,
+			embedding_size UNINDEXED,
+			tokenize='porter unicode61'
+		);
+	`)
+	if err != nil {
+		return nil, fmt.Errorf("create FTS table: %w", err)
+	}
+	// Create a logger that discards output.
+	logger := slog.New(slog.NewTextHandler(nil, &slog.HandlerOptions{Level: slog.LevelError}))
+	store := dummyStore{db: db}
+	// Create config with embedding dimension 768.
+	cfg := &config.Config{
+		EmbedDims:       768,
+		RAGWordLimit:    250,
+		RAGOverlapWords: 25,
+		RAGBatchSize:    1,
+	}
+	// Create a RAG instance using New, which will create an embedder based on config.
+	// We'll override the embedder afterwards via reflection.
+	rag, err := New(logger, store, cfg)
+	if err != nil {
+		return nil, fmt.Errorf("create RAG: %w", err)
+	}
+	// Replace the embedder with our mock.
+	rag.SetEmbedderForTesting(&mockEmbedder{dim: cfg.EmbedDims})
+	// Insert the provided chunks using the storage directly.
+	if len(chunks) > 0 {
+		// Ensure each chunk has embeddings of correct dimension (zero vector).
+		for _, chunk := range chunks {
+			if len(chunk.Embeddings) != cfg.EmbedDims {
+				chunk.Embeddings = make([]float32, cfg.EmbedDims)
+			}
+		}
+		err = rag.storage.WriteVectors(chunks)
+		if err != nil {
+			return nil, fmt.Errorf("write test chunks: %w", err)
+		}
+	}
+	return rag, nil
+}
+
+// createTestChunks returns a slice of VectorRow representing the target chunk
+// (kjv_bible.epub_1786_0), several bald‑related noise chunks, and unrelated chunks.
+func createTestChunks() []*models.VectorRow {
+	// Target chunk: 2 Kings 2:23‑24 containing "bald head" and "two she bears".
+	targetRaw := `And he said, Ye shall not send. 
+
+
+2:17 And when they urged him till he was ashamed, he said, Send.  They sent
+therefore fifty men; and they sought three days, but found him not. 
+
+
+2:18 And when they came again to him, (for he tarried at Jericho,) he said unto
+them, Did I not say unto you, Go not?  2:19 And the men of the city said unto
+Elisha, Behold, I pray thee, the situation of this city is pleasant, as my lord
+seeth: but the water is naught, and the ground barren. 
+
+
+2:20 And he said, Bring me a new cruse, and put salt therein.  And they brought
+it to him. 
+
+
+2:21 And he went forth unto the spring of the waters, and cast the salt in
+there, and said, Thus saith the LORD, I have healed these waters; there shall
+not be from thence any more death or barren land. 
+
+
+2:22 So the waters were healed unto this day, according to the saying of Elisha
+which he spake. 
+
+
+2:23 And he went up from thence unto Bethel: and as he was going up by the way,
+there came forth little children out of the city, and mocked him, and said unto
+him, Go up, thou bald head; go up, thou bald head. 
+
+
+2:24 And he turned back, and looked on them, and cursed them in the name of the
+LORD.  And there came forth two she bears out of the wood, and tare forty and
+two children of them.`
+	// Noise chunk 1: Leviticus containing "bald locust"
+	noise1Raw := `11:12 Whatsoever hath no fins nor scales in the waters, that shall be an
+abomination unto you. 
+
+
+11:13 And these are they which ye shall have in abomination among the fowls;
+they shall not be eaten, they are an abomination: the eagle, and the ossifrage,
+and the ospray, 11:14 And the vulture, and the kite after his kind; 11:15 Every
+raven after his kind; 11:16 And the owl, and the night hawk, and the cuckow,
+and the hawk after his kind, 11:17 And the little owl, and the cormorant, and
+the great owl, 11:18 And the swan, and the pelican, and the gier eagle, 11:19
+And the stork, the heron after her kind, and the lapwing, and the bat. 
+
+
+11:20 All fowls that creep, going upon all four, shall be an abomination unto
+you. 
+
+
+11:21 Yet these may ye eat of every flying creeping thing that goeth upon all
+four, which have legs above their feet, to leap withal upon the earth; 11:22
+Even these of them ye may eat; the locust after his kind, and the bald locust
+after his kind, and the beetle after his kind, and the grasshopper after his
+kind. 
+
+
+11:23 But all other flying creeping things, which have four feet, shall be an
+abomination unto you. 
+
+
+11:24 And for these ye shall be unclean: whosoever toucheth the carcase of them
+shall be unclean until the even.`
+	// Noise chunk 2: Leviticus containing "bald"
+	noise2Raw := `11:13 And these are they which ye shall have in abomination among the fowls;
+they shall not be eaten, they are an abomination: the eagle, and the ossifrage,
+and the ospray, 11:14 And the vulture, and the kite after his kind; 11:15 Every
+raven after his kind; 11:16 And the owl, and the night hawk, and the cuckow,
+and the hawk after his kind, 11:17 And the little owl, and the cormorant, and
+the great owl, 11:18 And the swan, and the pelican, and the gier eagle, 11:19
+And the stork, the heron after her kind, and the lapwing, and the bat. 
+
+
+11:20 All fowls that creep, going upon all four, shall be an abomination unto
+you. 
+
+
+11:21 Yet these may ye eat of every flying creeping thing that goeth upon all
+four, which have legs above their feet, to leap withal upon the earth; 11:22
+Even these of them ye may eat; the locust after his kind, and the bald locust
+after his kind, and the beetle after his kind, and the grasshopper after his
+kind. 
+
+
+11:23 But all other flying creeping things, which have four feet, shall be an
+abomination unto you. 
+
+
+11:24 And for these ye shall be unclean: whosoever toucheth the carcase of them
+shall be unclean until the even.`
+	// Additional Leviticus noise chunks (simulating 28 bald-related chunks)
+	// Using variations of the same text with different slugs
+	leviticusSlugs := []string{
+		"kjv_bible.epub_564_0",
+		"kjv_bible.epub_565_0",
+		"kjv_bible.epub_579_0",
+		"kjv_bible.epub_580_0",
+		"kjv_bible.epub_581_0",
+		"kjv_bible.epub_582_0",
+		"kjv_bible.epub_583_0",
+		"kjv_bible.epub_584_0",
+		"kjv_bible.epub_585_0",
+		"kjv_bible.epub_586_0",
+		"kjv_bible.epub_587_0",
+		"kjv_bible.epub_588_0",
+		"kjv_bible.epub_589_0",
+		"kjv_bible.epub_590_0",
+	}
+	leviticusTexts := []string{
+		noise1Raw,
+		noise2Raw,
+		`13:40 And the man whose hair is fallen off his head, he is bald; yet is he
+clean. 
+
+
+13:41 And he that hath his hair fallen off from the part of his head toward his
+face, he is forehead bald; yet is he clean.`,
+		`13:42 And if there be in the bald head, or bald forehead, a white reddish sore;
+it is a leprosy sprung up in his bald head, or his bald forehead.`,
+		`13:43 Then the priest shall look upon it: and, behold, if the rising of the
+sore be white reddish in his bald head, or in his bald forehead, as the leprosy
+appearedh in the skin of the flesh;`,
+		`13:44 He is a leprous man, he is unclean: the priest shall pronounce him utterly
+unclean; his plague is in his head.`,
+		`13:45 And the leper in whom the plague is, his clothes shall be rent, and his
+head bare, and he shall put a covering upon his upper lip, and shall cry,
+Unclean, unclean.`,
+		`13:46 All the days wherein the plague shall be in him he shall be defiled; he
+is unclean: he shall dwell alone; without the camp shall his habitation be.`,
+		`13:47 The garment also that the plague of leprosy is in, whether it be a woollen
+garment, or a linen garment;`,
+		`13:48 Whether it be in the warp, or woof; of linen, or of woollen; whether in a
+skin, or in any thing made of skin;`,
+		`13:49 And if the plague be greenish or reddish in the garment, or in the skin,
+either in the warp, or in the woof, or in any thing of skin; it is a plague of
+leprosy, and shall be shewed unto the priest:`,
+		`13:50 And the priest shall look upon the plague, and shut up it that hath the
+plague seven days:`,
+		`13:51 And he shall look on the plague on the seventh day: if the plague be spread
+in the garment, either in the warp, or in the woof, or in a skin, or in any work
+that is made of skin; the plague is a fretting leprosy; it is unclean.`,
+		`13:52 He shall therefore burn that garment, whether warp or woof, in woollen or
+in linen, or any thing of skin, wherein the plague is: for it is a fretting
+leprosy; it shall be burnt in the fire.`,
+	}
+	// Unrelated chunk 1: ghost_7.txt_777_0
+	unrelated1Raw := `Doesn’t he have any pride as a hunter?!  
+
+I didn’t see what other choice I had.  I would just have to grovel and be ready to flee at any given moment.  
+The Hidden Curse clan house was in the central region of the imperial capital.  It was a high-class area with extraordinary property values that hosted the residences of people like Lord Gladis.  This district was near the Imperial Castle, though “near” was a 
+relative term as it was still a few kilometers away.  
+
+The clan house was made of brick and conformed to an older style of architecture.`
+	// Unrelated chunk 2: ghost_7.txt_778_0
+	unrelated2Raw := `I would just have to grovel and be ready to flee at any given moment.  
+The Hidden Curse clan house was in the central region of the imperial capital.  It was a high-class area with extraordinary property values that hosted the residences of people like Lord Gladis.  This district was near the Imperial Castle, though “near” was a 
+relative term as it was still a few kilometers away.  
+
+The clan house was made of brick and conformed to an older style of architecture.  Nearly everyone knew about this mansion and its clock tower.  It stood tall over the neighboring mansions and rumor had it that you could see the whole capital from the top.  It 
+spoke to this clan’s renown and history that they were able to get away with building something that dwarfed the mansions of the nobility.`
+
+	chunks := []*models.VectorRow{
+		{
+			Slug:       "kjv_bible.epub_1786_0",
+			RawText:    targetRaw,
+			FileName:   "kjv_bible.epub",
+			Embeddings: nil, // will be filled with zero vector later
+		},
+	}
+	// Add Leviticus noise chunks
+	for i, slug := range leviticusSlugs {
+		text := leviticusTexts[i%len(leviticusTexts)]
+		chunks = append(chunks, &models.VectorRow{
+			Slug:       slug,
+			RawText:    text,
+			FileName:   "kjv_bible.epub",
+			Embeddings: nil,
+		})
+	}
+	// Add unrelated chunks
+	chunks = append(chunks,
+		&models.VectorRow{
+			Slug:       "ghost_7.txt_777_0",
+			RawText:    unrelated1Raw,
+			FileName:   "ghost_7.txt",
+			Embeddings: nil,
+		},
+		&models.VectorRow{
+			Slug:       "ghost_7.txt_778_0",
+			RawText:    unrelated2Raw,
+			FileName:   "ghost_7.txt",
+			Embeddings: nil,
+		},
+	)
+	return chunks
+}
+func assertTargetInTopN(t *testing.T, results []models.VectorRow, topN int) bool {
+	t.Helper()
+	for i, row := range results {
+		if i >= topN {
+			break
+		}
+		if row.Slug == "kjv_bible.epub_1786_0" {
+			return true
+		}
+	}
+	return false
+}
+
+func TestBiblicalQuery(t *testing.T) {
+	chunks := createTestChunks()
+	rag, err := setupTestRAG(t, chunks)
+	if err != nil {
+		t.Fatalf("setup failed: %v", err)
+	}
+	query := "bald prophet and two she bears"
+	results, err := rag.Search(query, 10)
+	if err != nil {
+		t.Fatalf("search failed: %v", err)
+	}
+	// The target chunk should be in the top results.
+	if !assertTargetInTopN(t, results, 5) {
+		t.Errorf("target chunk not found in top 5 results for query %q", query)
+		t.Logf("results slugs: %v", func() []string {
+			slugs := make([]string, len(results))
+			for i, r := range results {
+				slugs[i] = r.Slug
+			}
+			return slugs
+		}())
+	}
+}
+
+func TestQueryVariations(t *testing.T) {
+	chunks := createTestChunks()
+	rag, err := setupTestRAG(t, chunks)
+	if err != nil {
+		t.Fatalf("setup failed: %v", err)
+	}
+	tests := []struct {
+		name  string
+		query string
+		topN  int
+	}{
+		{"she bears", "she bears", 5},
+		{"bald head", "bald head", 5},
+		{"two she bears out of the wood", "two she bears out of the wood", 5},
+		{"bald prophet", "bald prophet", 10},
+		{"go up thou bald head", "\"go up thou bald head\"", 5},
+		{"two she bears", "\"two she bears\"", 5},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			results, err := rag.Search(tt.query, 10)
+			if err != nil {
+				t.Fatalf("search failed: %v", err)
+			}
+			if !assertTargetInTopN(t, results, tt.topN) {
+				t.Errorf("target chunk not found in top %d results for query %q", tt.topN, tt.query)
+				t.Logf("results slugs: %v", func() []string {
+					slugs := make([]string, len(results))
+					for i, r := range results {
+						slugs[i] = r.Slug
+					}
+					return slugs
+				}())
+			}
+		})
+	}
+}
--- a/rag/rag_real_test.go
+++ b/rag/rag_real_test.go
@@ -0,0 +1,131 @@
+package rag
+
+import (
+	"gf-lt/config"
+	"gf-lt/storage"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestRealBiblicalQuery(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping real embedder test in short mode")
+	}
+	// Check if the embedder model exists
+	modelPath := filepath.Join("..", "onnx", "embedgemma", "model_q4.onnx")
+	if _, err := os.Stat(modelPath); os.IsNotExist(err) {
+		t.Skipf("embedder model not found at %s; skipping real embedder test", modelPath)
+	}
+	tokenizerPath := filepath.Join("..", "onnx", "embedgemma", "tokenizer.json")
+	dbPath := filepath.Join("..", "gflt.db")
+	if _, err := os.Stat(dbPath); os.IsNotExist(err) {
+		t.Skipf("database not found at %s; skipping real embedder test", dbPath)
+	}
+	cfg := &config.Config{
+		EmbedModelPath:     modelPath,
+		EmbedTokenizerPath: tokenizerPath,
+		EmbedDims:          768,
+		RAGWordLimit:       250,
+		RAGOverlapWords:    25,
+		RAGBatchSize:       1,
+	}
+	logger := slog.New(slog.NewTextHandler(nil, &slog.HandlerOptions{Level: slog.LevelError}))
+	store := storage.NewProviderSQL(dbPath, logger)
+	if store == nil {
+		t.Fatal("failed to create storage provider")
+	}
+	rag, err := New(logger, store, cfg)
+	if err != nil {
+		t.Fatalf("failed to create RAG instance: %v", err)
+	}
+	t.Cleanup(func() { rag.Destroy() })
+
+	query := "bald prophet and two she bears"
+	results, err := rag.Search(query, 30)
+	if err != nil {
+		t.Fatalf("search failed: %v", err)
+	}
+	found := false
+	for i, row := range results {
+		if row.Slug == "kjv_bible.epub_1786_0" {
+			found = true
+			t.Logf("target chunk found at rank %d", i+1)
+			break
+		}
+	}
+	if !found {
+		t.Errorf("target chunk not found in search results for query %q", query)
+		t.Logf("results slugs:")
+		for i, r := range results {
+			t.Logf("%d: %s", i+1, r.Slug)
+		}
+	}
+}
+
+func TestRealQueryVariations(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping real embedder test in short mode")
+	}
+	modelPath := filepath.Join("..", "onnx", "embedgemma", "model_q4.onnx")
+	if _, err := os.Stat(modelPath); os.IsNotExist(err) {
+		t.Skipf("embedder model not found at %s; skipping real embedder test", modelPath)
+	}
+	tokenizerPath := filepath.Join("..", "onnx", "embedgemma", "tokenizer.json")
+	dbPath := filepath.Join("..", "gflt.db")
+	if _, err := os.Stat(dbPath); os.IsNotExist(err) {
+		t.Skipf("database not found at %s; skipping real embedder test", dbPath)
+	}
+	cfg := &config.Config{
+		EmbedModelPath:     modelPath,
+		EmbedTokenizerPath: tokenizerPath,
+		EmbedDims:          768,
+		RAGWordLimit:       250,
+		RAGOverlapWords:    25,
+		RAGBatchSize:       1,
+	}
+	logger := slog.New(slog.NewTextHandler(nil, &slog.HandlerOptions{Level: slog.LevelError}))
+	store := storage.NewProviderSQL(dbPath, logger)
+	if store == nil {
+		t.Fatal("failed to create storage provider")
+	}
+	rag, err := New(logger, store, cfg)
+	if err != nil {
+		t.Fatalf("failed to create RAG instance: %v", err)
+	}
+	t.Cleanup(func() { rag.Destroy() })
+
+	tests := []struct {
+		name  string
+		query string
+	}{
+		{"she bears", "she bears"},
+		{"bald head", "bald head"},
+		{"two she bears out of the wood", "two she bears out of the wood"},
+		{"bald prophet", "bald prophet"},
+		{"go up thou bald head", "\"go up thou bald head\""},
+		{"two she bears", "\"two she bears\""},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			results, err := rag.Search(tt.query, 10)
+			if err != nil {
+				t.Fatalf("search failed: %v", err)
+			}
+			found := false
+			for _, row := range results {
+				if row.Slug == "kjv_bible.epub_1786_0" {
+					found = true
+					break
+				}
+			}
+			if !found {
+				t.Errorf("target chunk not found for query %q", tt.query)
+				for i, r := range results {
+					t.Logf("%d: %s", i+1, r.Slug)
+				}
+			}
+		})
+	}
+}
--- a/rag/rag_test.go
+++ b/rag/rag_test.go
@@ -0,0 +1,155 @@
+package rag
+
+import (
+	"testing"
+)
+
+func TestDetectPhrases(t *testing.T) {
+	tests := []struct {
+		query  string
+		expect []string
+	}{
+		{
+			query:  "bald prophet and two she bears",
+			expect: []string{"bald prophet", "two she", "two she bears", "she bears"},
+		},
+		{
+			query:  "she bears",
+			expect: []string{"she bears"},
+		},
+		{
+			query:  "the quick brown fox",
+			expect: []string{"quick brown", "quick brown fox", "brown fox"},
+		},
+		{
+			query:  "in the house", // stop words
+			expect: []string{},     // "in" and "the" are stop words
+		},
+		{
+			query:  "a", // short
+			expect: []string{},
+		},
+	}
+
+	for _, tt := range tests {
+		got := detectPhrases(tt.query)
+		if len(got) != len(tt.expect) {
+			t.Errorf("detectPhrases(%q) = %v, want %v", tt.query, got, tt.expect)
+			continue
+		}
+		for i := range got {
+			if got[i] != tt.expect[i] {
+				t.Errorf("detectPhrases(%q) = %v, want %v", tt.query, got, tt.expect)
+				break
+			}
+		}
+	}
+}
+
+func TestCountPhraseMatches(t *testing.T) {
+	tests := []struct {
+		text   string
+		query  string
+		expect int
+	}{
+		{
+			text:   "two she bears came out of the wood",
+			query:  "she bears",
+			expect: 1,
+		},
+		{
+			text:   "bald head and she bears",
+			query:  "bald prophet and two she bears",
+			expect: 1, // only "she bears" matches
+		},
+		{
+			text:   "no match here",
+			query:  "she bears",
+			expect: 0,
+		},
+		{
+			text:   "she bears and bald prophet",
+			query:  "bald prophet she bears",
+			expect: 2, // "she bears" and "bald prophet"
+		},
+	}
+
+	for _, tt := range tests {
+		got := countPhraseMatches(tt.text, tt.query)
+		if got != tt.expect {
+			t.Errorf("countPhraseMatches(%q, %q) = %d, want %d", tt.text, tt.query, got, tt.expect)
+		}
+	}
+}
+
+func TestAreSlugsAdjacent(t *testing.T) {
+	tests := []struct {
+		slug1  string
+		slug2  string
+		expect bool
+	}{
+		{
+			slug1:  "kjv_bible.epub_1786_0",
+			slug2:  "kjv_bible.epub_1787_0",
+			expect: true,
+		},
+		{
+			slug1:  "kjv_bible.epub_1787_0",
+			slug2:  "kjv_bible.epub_1786_0",
+			expect: true,
+		},
+		{
+			slug1:  "kjv_bible.epub_1786_0",
+			slug2:  "kjv_bible.epub_1788_0",
+			expect: false,
+		},
+		{
+			slug1:  "otherfile.txt_1_0",
+			slug2:  "kjv_bible.epub_1786_0",
+			expect: false,
+		},
+		{
+			slug1:  "file_1_0",
+			slug2:  "file_1_1",
+			expect: true,
+		},
+		{
+			slug1:  "file_1_0",
+			slug2:  "file_2_0", // different batch
+			expect: true,       // sequential batches with same chunk index are adjacent
+		},
+	}
+
+	for _, tt := range tests {
+		got := areSlugsAdjacent(tt.slug1, tt.slug2)
+		if got != tt.expect {
+			t.Errorf("areSlugsAdjacent(%q, %q) = %v, want %v", tt.slug1, tt.slug2, got, tt.expect)
+		}
+	}
+}
+
+func TestParseSlugIndices(t *testing.T) {
+	tests := []struct {
+		slug      string
+		wantBatch int
+		wantChunk int
+		wantOk    bool
+	}{
+		{"kjv_bible.epub_1786_0", 1786, 0, true},
+		{"file_1_5", 1, 5, true},
+		{"no_underscore", 0, 0, false},
+		{"file_abc_def", 0, 0, false},
+		{"file_123_456_extra", 456, 0, false}, // regex matches last two numbers
+	}
+
+	for _, tt := range tests {
+		batch, chunk, ok := parseSlugIndices(tt.slug)
+		if ok != tt.wantOk {
+			t.Errorf("parseSlugIndices(%q) ok = %v, want %v", tt.slug, ok, tt.wantOk)
+			continue
+		}
+		if ok && (batch != tt.wantBatch || chunk != tt.wantChunk) {
+			t.Errorf("parseSlugIndices(%q) = (%d, %d), want (%d, %d)", tt.slug, batch, chunk, tt.wantBatch, tt.wantChunk)
+		}
+	}
+}
--- a/rag/storage.go
+++ b/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) > 3 {
-			topResults = topResults[:3] // Keep only closest 3
+		if len(topResults) > limit {
+			topResults = topResults[:limit]
 		}
 	}
-
-	// Convert back to VectorRow slice
 	results := make([]models.VectorRow, 0, len(topResults))
 	for _, result := range topResults {
 		result.vector.Distance = result.distance
@@ -171,6 +261,98 @@ 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. Keep as negative.
+		distance := float32(score) // Keep negative, more negative is better
+		// No clamping needed; negative distances are fine
+		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 +397,10 @@ 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)
--- a/storage/migrations/003_add_fts.down.sql
+++ b/storage/migrations/003_add_fts.down.sql
@@ -0,0 +1,2 @@
+-- Drop FTS5 virtual table
+DROP TABLE IF EXISTS fts_embeddings;
--- a/storage/migrations/003_add_fts.up.sql
+++ b/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.
--- a/storage/migrations/004_populate_fts.down.sql
+++ b/storage/migrations/004_populate_fts.down.sql
@@ -0,0 +1,2 @@
+-- Clear FTS table (optional)
+DELETE FROM fts_embeddings;
--- a/storage/migrations/004_populate_fts.up.sql
+++ b/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;
--- a/storage/storage.go
+++ b/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)
--- a/storage/vector.go
+++ b/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,
 		}
-
-		// 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]
-					}
-				}
-			}
-			topResults = topResults[:3]
-		}
+		allResults = append(allResults, result)
+	}
+	// 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]
 	}
-
 	// Convert back to VectorRow slice
-	results := make([]models.VectorRow, len(topResults))
-	for i, result := range topResults {
+	results := make([]models.VectorRow, len(allResults))
+	for i, result := range allResults {
 		result.vector.Distance = result.distance
 		results[i] = result.vector
 	}
--- a/sysprompts/cluedo.json
+++ b/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?"
-}
--- a/tables.go
+++ b/tables.go
@@ -243,11 +243,9 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 	for _, f := range loadedFiles {
 		loadedSet[f] = true
 	}
-
 	// Build merged list: files from ragdir + orphaned files from DB
 	ragFiles := make([]ragFileInfo, 0, len(fileList)+len(loadedFiles))
 	seen := make(map[string]bool)
-
 	// Add files from ragdir
 	for _, f := range fileList {
 		ragFiles = append(ragFiles, ragFileInfo{
@@ -258,7 +256,6 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 		})
 		seen[f] = true
 	}
-
 	// Add orphaned files (in DB but not in ragdir)
 	for _, f := range loadedFiles {
 		if !seen[f] {
@@ -275,7 +272,7 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 	fileTable := tview.NewTable().
 		SetBorders(true)
 	longStatusView := tview.NewTextView()
-	longStatusView.SetText("status text")
+	longStatusView.SetText("press x to exit")
 	longStatusView.SetBorder(true).SetTitle("status")
 	longStatusView.SetChangedFunc(func() {
 		app.Draw()
@@ -376,7 +373,6 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 	}
 	errCh := make(chan error, 1) // why?
 	go func() {
-		defer pages.RemovePage(RAGPage)
 		for {
 			select {
 			case err := <-errCh:
@@ -417,7 +413,6 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 			}
 			return
 		}
-		// defer pages.RemovePage(RAGPage)
 		tc := fileTable.GetCell(row, column)
 		tc.SetTextColor(tcell.ColorRed)
 		fileTable.SetSelectable(false, false)
@@ -430,7 +425,6 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 		f := ragFiles[row-1]
 		// Handle "-" case (orphaned file with no delete option)
 		if tc.Text == "-" {
-			pages.RemovePage(RAGPage)
 			return
 		}
 		switch tc.Text {
@@ -441,14 +435,14 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 				if err := ragger.LoadRAG(fpath); err != nil {
 					logger.Error("failed to embed file", "chat", fpath, "error", err)
 					showToast("RAG", "failed to embed file; error: "+err.Error())
-					app.QueueUpdate(func() {
-						pages.RemovePage(RAGPage)
-					})
 					return
 				}
 				showToast("RAG", "file loaded successfully")
 				app.QueueUpdate(func() {
 					pages.RemovePage(RAGPage)
+					loadedFiles, _ := ragger.ListLoaded()
+					chatRAGTable := makeRAGTable(fileList, loadedFiles)
+					pages.AddPage(RAGPage, chatRAGTable, true, true)
 				})
 			}()
 			return
@@ -458,14 +452,14 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 				if err := ragger.RemoveFile(f.name); err != nil {
 					logger.Error("failed to unload file from RAG", "filename", f.name, "error", err)
 					showToast("RAG", "failed to unload file; error: "+err.Error())
-					app.QueueUpdate(func() {
-						pages.RemovePage(RAGPage)
-					})
 					return
 				}
 				showToast("RAG", "file unloaded successfully")
 				app.QueueUpdate(func() {
 					pages.RemovePage(RAGPage)
+					loadedFiles, _ := ragger.ListLoaded()
+					chatRAGTable := makeRAGTable(fileList, loadedFiles)
+					pages.AddPage(RAGPage, chatRAGTable, true, true)
 				})
 			}()
 			return
@@ -476,6 +470,21 @@ func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
 				return
 			}
 			showToast("chat deleted", fpath+" was deleted")
+			go func() {
+				app.QueueUpdate(func() {
+					pages.RemovePage(RAGPage)
+					newFileList, _ := os.ReadDir(cfg.RAGDir)
+					loadedFiles, _ := ragger.ListLoaded()
+					var newFiles []string
+					for _, f := range newFileList {
+						if !f.IsDir() {
+							newFiles = append(newFiles, f.Name())
+						}
+					}
+					chatRAGTable := makeRAGTable(newFiles, loadedFiles)
+					pages.AddPage(RAGPage, chatRAGTable, true, true)
+				})
+			}()
 			return
 		default:
 			pages.RemovePage(RAGPage)
--- a/tools.go
+++ b/tools.go
@@ -207,7 +207,7 @@ var (
 	modelHasVision       bool
 )

-func init() {
+func initTools() {
 	sysMap[basicCard.ID] = basicCard
 	roleToID["assistant"] = basicCard.ID
 	sa, err := searcher.NewWebSurfer(searcher.SearcherTypeScraper, "")
@@ -360,13 +360,13 @@ func ragsearch(args map[string]string) []byte {
 	}
 	limitS, ok := args["limit"]
 	if !ok || limitS == "" {
-		limitS = "3"
+		limitS = "10"
 	}
 	limit, err := strconv.Atoi(limitS)
 	if err != nil || limit == 0 {
 		logger.Warn("ragsearch limit; passed bad value; setting to default (3)",
 			"limit_arg", limitS, "error", err)
-		limit = 3
+		limit = 10
 	}
 	ragInstance := rag.GetInstance()
 	if ragInstance == nil {
@@ -2273,56 +2273,3 @@ var baseTools = []models.Tool{
 		},
 	},
 }
-
-func init() {
-	if windowToolsAvailable {
-		baseTools = append(baseTools,
-			models.Tool{
-				Type: "function",
-				Function: models.ToolFunc{
-					Name:        "list_windows",
-					Description: "List all visible windows with their IDs and names. Returns a map of window ID to window name.",
-					Parameters: models.ToolFuncParams{
-						Type:       "object",
-						Required:   []string{},
-						Properties: map[string]models.ToolArgProps{},
-					},
-				},
-			},
-			models.Tool{
-				Type: "function",
-				Function: models.ToolFunc{
-					Name:        "capture_window",
-					Description: "Capture a screenshot of a specific window and save it to /tmp. Requires window parameter (window ID or name substring).",
-					Parameters: models.ToolFuncParams{
-						Type:     "object",
-						Required: []string{"window"},
-						Properties: map[string]models.ToolArgProps{
-							"window": models.ToolArgProps{
-								Type:        "string",
-								Description: "window ID or window name (partial match)",
-							},
-						},
-					},
-				},
-			},
-			models.Tool{
-				Type: "function",
-				Function: models.ToolFunc{
-					Name:        "capture_window_and_view",
-					Description: "Capture a screenshot of a specific window, save it to /tmp, and return the image for viewing. Requires window parameter (window ID or name substring).",
-					Parameters: models.ToolFuncParams{
-						Type:     "object",
-						Required: []string{"window"},
-						Properties: map[string]models.ToolArgProps{
-							"window": models.ToolArgProps{
-								Type:        "string",
-								Description: "window ID or window name (partial match)",
-							},
-						},
-					},
-				},
-			},
-		)
-	}
-}
--- a/tui.go
+++ b/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
@@ -36,10 +38,10 @@ var (
 	roleEditWindow     *tview.InputField
 	shellInput         *tview.InputField
 	confirmModal       *tview.Modal
+	toastTimer         *time.Timer
 	confirmPageName    = "confirm"
 	fullscreenMode     bool
 	positionVisible    bool = true
-	scrollToEndEnabled bool = true
 	// pages
 	historyPage    = "historyPage"
 	agentPage      = "agentPage"
@@ -48,7 +50,6 @@ var (
 	helpPage       = "helpPage"
 	renamePage     = "renamePage"
 	RAGPage        = "RAGPage"
-	RAGLoadedPage  = "RAGLoadedPage"
 	propsPage      = "propsPage"
 	codeBlockPage  = "codeBlockPage"
 	imgPage        = "imgPage"
@@ -137,8 +138,8 @@ func setShellMode(enabled bool) {
 	}()
 }

-// showToast displays a temporary message in the top‑right corner.
-// It auto‑hides after 3 seconds and disappears when clicked.
+// showToast displays a temporary notification in the bottom-right corner.
+// 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 {
@@ -154,39 +155,74 @@ func showToast(title, message string) {
 	}
 	title = sanitize(title, 50)
 	message = sanitize(message, 197)
-	notification := tview.NewTextView().
-		SetTextAlign(tview.AlignCenter).
-		SetDynamicColors(true).
-		SetRegions(true).
-		SetText(fmt.Sprintf("[yellow]%s[-]\n", message)).
-		SetChangedFunc(func() {
-			app.Draw()
+	if toastTimer != nil {
+		toastTimer.Stop()
+	}
+	// show blocking notification to not mess up flex
+	if fullscreenMode {
+		notification := tview.NewTextView().
+			SetTextAlign(tview.AlignCenter).
+			SetDynamicColors(true).
+			SetRegions(true).
+			SetText(fmt.Sprintf("[yellow]%s[-]\n", message)).
+			SetChangedFunc(func() {
+				app.Draw()
+			})
+		notification.SetTitleAlign(tview.AlignLeft).
+			SetBorder(true).
+			SetTitle(title)
+		// Wrap it in a full‑screen Flex to position it in the top‑right corner.
+		// Outer Flex (row) pushes content to the top; inner Flex (column) pushes to the right.
+		background := tview.NewFlex().SetDirection(tview.FlexRow).
+			AddItem(nil, 0, 1, false). // top spacer
+			AddItem(tview.NewFlex().SetDirection(tview.FlexColumn).
+				AddItem(nil, 0, 1, false).          // left spacer
+				AddItem(notification, 40, 1, true), // notification width 40
+				5, 1, false) // notification height 5
+		// 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 2 seconds, since blocking is more annoying
+		time.AfterFunc(2*time.Second, func() {
+			app.QueueUpdateDraw(func() {
+				if pages.HasPage(pageName) {
+					pages.RemovePage(pageName)
+				}
+			})
 		})
-	notification.SetTitleAlign(tview.AlignLeft).
-		SetBorder(true).
-		SetTitle(title)
-	// Wrap it in a full‑screen Flex to position it in the top‑right corner.
-	// Outer Flex (row) pushes content to the top; inner Flex (column) pushes to the right.
-	background := tview.NewFlex().SetDirection(tview.FlexRow).
-		AddItem(nil, 0, 1, false). // top spacer
-		AddItem(tview.NewFlex().SetDirection(tview.FlexColumn).
-			AddItem(nil, 0, 1, false).          // left spacer
-			AddItem(notification, 40, 1, true), // notification width 40
-			5, 1, false) // notification height 5
-	// 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.
-	time.AfterFunc(3*time.Second, func() {
+		return
+	}
+	notificationWidget.SetTitle(title)
+	notificationWidget.SetText(fmt.Sprintf("[yellow]%s[-]", message))
+	go func() {
 		app.QueueUpdateDraw(func() {
-			if pages.HasPage(pageName) {
-				pages.RemovePage(pageName)
+			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() {
+func initTUI() {
 	// Start background goroutine to update model color cache
 	startModelColorUpdater()
 	tview.Styles = colorschemes["default"]
@@ -235,7 +271,7 @@ 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 +322,26 @@ 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)
 	}
@@ -360,10 +410,14 @@ 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)
 	}
@@ -578,7 +632,7 @@ func init() {
 	updateStatusLine()
 	textView.SetText(chatToText(chatBody.Messages, cfg.ShowSys))
 	colorText()
-	if scrollToEndEnabled {
+	if cfg.AutoScrollEnabled {
 		textView.ScrollToEnd()
 	}
 	// init sysmap
@@ -607,9 +661,9 @@ func init() {
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == '2' && event.Modifiers()&tcell.ModAlt != 0 {
 			// toggle auto-scrolling
-			scrollToEndEnabled = !scrollToEndEnabled
+			cfg.AutoScrollEnabled = !cfg.AutoScrollEnabled
 			status := "disabled"
-			if scrollToEndEnabled {
+			if cfg.AutoScrollEnabled {
 				status = "enabled"
 			}
 			showToast("autoscroll", "Auto-scrolling "+status)
@@ -676,7 +730,7 @@ func init() {
 			updateStatusLine()
 			return nil
 		}
-		if event.Key() == tcell.KeyF2 && !botRespMode {
+		if event.Key() == tcell.KeyF2 && !botRespMode.Load() {
 			// regen last msg
 			if len(chatBody.Messages) == 0 {
 				showToast("info", "no messages to regenerate")
@@ -693,7 +747,7 @@ func init() {
 			chatRoundChan <- &models.ChatRoundReq{Role: cfg.UserRole, Regen: true}
 			return nil
 		}
-		if event.Key() == tcell.KeyF3 && !botRespMode {
+		if event.Key() == tcell.KeyF3 && !botRespMode.Load() {
 			// delete last msg
 			// check textarea text; if it ends with bot icon delete only icon:
 			text := textView.GetText(true)
@@ -749,9 +803,9 @@ func init() {
 			return nil
 		}
 		if event.Key() == tcell.KeyF6 {
-			interruptResp = true
-			botRespMode = false
-			toolRunningMode = false
+			interruptResp.Store(true)
+			botRespMode.Store(false)
+			toolRunningMode.Store(false)
 			return nil
 		}
 		if event.Key() == tcell.KeyF7 {
@@ -1046,7 +1100,7 @@ func init() {
 			return nil
 		}
 		// cannot send msg in editMode or botRespMode
-		if event.Key() == tcell.KeyEscape && !editMode && !botRespMode {
+		if event.Key() == tcell.KeyEscape && !editMode && !botRespMode.Load() {
 			if shellMode {
 				cmdText := shellInput.GetText()
 				if cmdText != "" {
@@ -1083,7 +1137,7 @@ func init() {
 				fmt.Fprintf(textView, "%s[-:-:b](%d) <%s>: [-:-:-]\n%s\n",
 					nl, len(chatBody.Messages), persona, msgText)
 				textArea.SetText("", true)
-				if scrollToEndEnabled {
+				if cfg.AutoScrollEnabled {
 					textView.ScrollToEnd()
 				}
 				colorText()
@@ -1095,7 +1149,7 @@ 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()
@@ -1112,9 +1166,10 @@ func init() {
 			app.SetFocus(focusSwitcher[currentF])
 			return nil
 		}
-		if isASCII(string(event.Rune())) && !botRespMode {
+		if isASCII(string(event.Rune())) && !botRespMode.Load() {
 			return event
 		}
 		return event
 	})
+	go updateModelLists()
 }
Author	SHA1	Message	Date
Grail Finder	a1b5f9cdc5	Enha: rag tuning and tests	2026-03-08 16:12:32 +03:00
Grail Finder	e74ff8c03f	Enha (rag): semantic hybrid search	2026-03-08 13:27:09 +03:00
Grail Finder	b6e802c12e	Enha (rag): bigger default batch	2026-03-08 11:38:56 +03:00
Grail Finder	c0d5db29a5	Chore (rag): x to exit label	2026-03-08 10:28:30 +03:00
Grail Finder	6ed96c9bd3	Fix (ctrl+w): avoid msg duplication	2026-03-08 09:42:07 +03:00
Grail Finder	b5f0eabeea	Fix (rag): do not hang on delete	2026-03-08 09:00:24 +03:00
Grail Finder	e0201886f8	Enha (rag): keep page open until user closes it	2026-03-08 08:50:50 +03:00
Grail Finder	5b175c12a6	Chore: update readme	2026-03-08 07:29:04 +03:00
Grail Finder	c200c9328c	Enha: botresp, toolresp to atomic	2026-03-08 07:13:27 +03:00
Grail Finder	23cb8f2578	Chore: remove AutoCleanToolCallsFromCtx, atomic model color	2026-03-08 06:45:51 +03:00
Grail Finder	4f0bce50c5	Chore: one init for clear call order	2026-03-07 19:11:13 +03:00
Grail Finder	bf655a1087	Enha: llama.cpp on non localhost	2026-03-07 18:42:12 +03:00
Grail Finder	c8f00198d6	Dep (stt): use ffmpeg instead of portaudio	2026-03-07 18:13:11 +03:00
Grail Finder	c5a24b2a3f	Enha: google-tts replay speed	2026-03-07 16:37:09 +03:00
Grail Finder	0f0c43f327	Dep: remove beep/portaudio dependancy	2026-03-07 16:24:39 +03:00
Grail Finder	0e55e44f62	Enha (kokoro): use ffplay instead of beep (portaudio)	2026-03-07 15:41:39 +03:00
Grail Finder	014e297ae3	Chore: linter complaints	2026-03-06 19:57:44 +03:00
Grail Finder	5f273681df	Chore: remove plan doc	2026-03-06 19:03:26 +03:00
Grail Finder	17b68bc21f	Enha (rag): async writes	2026-03-06 18:58:23 +03:00
Grail Finder	edfd43c52a	Doc: update	2026-03-06 13:45:12 +03:00
Grail Finder	62ec55505c	Enha (rag): query each doc	2026-03-06 13:17:49 +03:00
Grail Finder	f9866bcf5a	Feat (rag): hybrid search attempt	2026-03-06 11:20:50 +03:00
Grail Finder	822cc48834	Fix: avoid panic if statuslinewidget not loaded yet	2026-03-06 10:37:08 +03:00
Grail Finder	4ef0a21511	Enha (onnx): unload model if noop for 30s	2026-03-06 09:32:45 +03:00
Grail Finder	d2caebdb4f	Enha (onnx): use gpu	2026-03-06 09:11:25 +03:00
Grail Finder	e1f2a8cd7b	Chore: remove unused RagEnabled var	2026-03-06 07:46:15 +03:00
Grail Finder	efc92d884c	Chore: onnx library lookup	2026-03-05 20:02:46 +03:00
Grail Finder	ac8c8bb055	Enha: onnx config vars	2026-03-05 19:20:21 +03:00
Grail Finder	c2c107c786	Dep: make-fetch onnx embed gemma	2026-03-05 16:05:03 +03:00
Grail Finder	c2757653a3	Fix: buildable	2026-03-05 14:49:59 +03:00
Grail Finder	4bd6883966	WIP	2026-03-05 14:38:26 +03:00
Grail Finder	7c56e27dbe	Dep: trying sugarme tokenizer	2026-03-05 14:27:19 +03:00
Grail Finder	fbc955ca37	Enha: local onnx	2026-03-05 14:13:58 +03:00
Grail Finder	c65c11bcfb	Fix: shellmode tab completion	2026-03-05 11:36:35 +03:00
Grail Finder	04f1fd464b	Chore: remove cluedo sysprompt	2026-03-05 11:17:01 +03:00
Grail Finder	6e9c453ee0	Enha: explicit app.Draw per textView update for smooth streaming	2026-03-05 10:35:17 +03:00
Grail Finder	645b7351a8	Fix: add different kind of notifiction for fullscreen mode	2026-03-05 09:09:13 +03:00
Grail Finder	57088565bd	Fix (notification): being closed by prev notification early	2026-03-05 08:51:04 +03:00
Grail Finder	4b6769e531	Fix (notification): non-blocking way to notify	2026-03-05 08:43:50 +03:00