Chore: panic is not solved (WIP)

Chore: cleanup logs
Chore: move to own file
2026-03-07 09:40:58 +03:00 · 2026-03-07 09:08:01 +03:00 · 2026-03-07 08:52:10 +03:00 · 2026-03-07 08:35:44 +03:00 · 2026-03-06 19:57:44 +03:00 · 2026-03-06 19:03:26 +03:00
23 changed files with 1848 additions and 235 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/126
+++ b/126
@@ -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 fetch-kokoro-voices install-espeak

 run: setconfig
 	go build -tags extra -o gf-lt && ./gf-lt
@@ -30,6 +30,108 @@ 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
+
+fetch-kokoro-onnx:
+	mkdir -p onnx/kokoro && curl -o onnx/kokoro/config.json -L https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX/resolve/main/config.json && curl -o onnx/kokoro/tokenizer.json -L https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX/resolve/main/tokenizer.json && curl -o onnx/kokoro/model_quantized.onnx -L https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX/resolve/main/onnx/model_quantized.onnx && curl -o onnx/kokoro/voices.bin -L https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin
+
+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

@@ -95,3 +197,25 @@ docker-logs-whisper: ## View logs from Whisper STT service only
 docker-logs-kokoro: ## View logs from Kokoro TTS service only
 	@echo "Displaying logs from Kokoro TTS service..."
 	docker-compose -f batteries/docker-compose.yml logs -f kokoro-tts
+
+# Kokoro ONNX TTS Setup
+install-espeak: ## Install espeak-ng for phoneme tokenization
+	@echo "=== Installing espeak-ng ===" && \
+	if command -v espeak-ng >/dev/null 2>&1; then \
+		echo "espeak-ng is already installed:" && \
+		espeak-ng --version && \
+		exit 0; \
+	fi && \
+	echo "Installing espeak-ng..." && \
+	sudo apt-get update && \
+	sudo apt-get install -y espeak-ng espeak && \
+	echo "espeak-ng installed successfully!" && \
+	espeak-ng --version
+
+fetch-kokoro-voices: ## Download Kokoro voice files (PyTorch format)
+	@echo "=== Downloading Kokoro voices ===" && \
+	mkdir -p onnx/kokoro/voices && \
+	echo "Downloading af_bella voice..." && \
+	curl -L -o onnx/kokoro/voices/af_bella.pt https://raw.githubusercontent.com/hexgrad/kokoro/main/kokoro/voices/af_heart.pt && \
+	echo "Voice file downloaded to onnx/kokoro/voices/" && \
+	ls -lh onnx/kokoro/voices/
--- a/bot.go
+++ b/bot.go
@@ -418,7 +418,9 @@ func fetchLCPModelsWithStatus() (*models.LCPModels, error) {
 	if err := json.NewDecoder(resp.Body).Decode(data); err != nil {
 		return nil, err
 	}
+	localModelsMu.Lock()
 	localModelsData = data
+	localModelsMu.Unlock()
 	return data, nil
 }

@@ -1393,12 +1395,16 @@ 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()
+	for statusLineWidget == nil {
+		time.Sleep(time.Millisecond * 100)
+	}
 	// set already loaded model in llama.cpp
 	if strings.Contains(cfg.CurrentAPI, "localhost") || strings.Contains(cfg.CurrentAPI, "127.0.0.1") {
 		localModelsMu.Lock()
@@ -1493,14 +1499,20 @@ func init() {
 	// load cards
 	basicCard.Role = cfg.AssistantRole
 	logLevel.Set(slog.LevelInfo)
-	logger = slog.New(slog.NewTextHandler(logfile, &slog.HandlerOptions{Level: logLevel}))
+	logger = slog.New(slog.NewTextHandler(logfile, &slog.HandlerOptions{Level: logLevel, AddSource: true}))
 	store = storage.NewProviderSQL(cfg.DBPATH, logger)
 	if store == nil {
 		cancel()
 		os.Exit(1)
 		return
 	}
-	ragger = rag.New(logger, store, cfg)
+	ragger, err = rag.New(logger, store, cfg)
+	if err != nil {
+		logger.Error("failed to create RAG", "error", err)
+	}
+	if ragger != nil && ragger.FallbackMessage() != "" && app != nil {
+		showToast("RAG", "ONNX unavailable, using API: "+ragger.FallbackMessage())
+	}
 	// https://github.com/coreydaley/ggerganov-llama.cpp/blob/master/examples/server/README.md
 	// load all chats in memory
 	if _, err := loadHistoryChats(); err != nil {
--- 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
+RAGOverlapWords = 16
 RAGDir = "ragimport"
 # extra tts
 TTS_ENABLED = false
--- a/config/config.go
+++ b/config/config.go
@@ -36,11 +36,14 @@ type Config struct {
 	// embeddings
 	EmbedURL           string `toml:"EmbedURL"`
 	HFToken            string `toml:"HFToken"`
+	EmbedModelPath     string `toml:"EmbedModelPath"`
+	EmbedTokenizerPath string `toml:"EmbedTokenizerPath"`
+	EmbedDims          int    `toml:"EmbedDims"`
 	// rag settings
-	RAGEnabled   bool   `toml:"RAGEnabled"`
 	RAGDir          string `toml:"RAGDir"`
 	RAGBatchSize    int    `toml:"RAGBatchSize"`
 	RAGWordLimit    uint32 `toml:"RAGWordLimit"`
+	RAGOverlapWords uint32 `toml:"RAGOverlapWords"`
 	// deepseek
 	DeepSeekChatAPI       string `toml:"DeepSeekChatAPI"`
 	DeepSeekCompletionAPI string `toml:"DeepSeekCompletionAPI"`
@@ -58,6 +61,10 @@ type Config struct {
 	TTS_SPEED    float32 `toml:"TTS_SPEED"`
 	TTS_PROVIDER string  `toml:"TTS_PROVIDER"`
 	TTS_LANGUAGE string  `toml:"TTS_LANGUAGE"`
+	// Kokoro ONNX TTS
+	KokoroModelPath  string `toml:"KokoroModelPath"`
+	KokoroVoicesPath string `toml:"KokoroVoicesPath"`
+	KokoroVoice      string `toml:"KokoroVoice"`
 	// STT
 	STT_TYPE          string `toml:"STT_TYPE"` // WHISPER_SERVER, WHISPER_BINARY
 	STT_URL           string `toml:"STT_URL"`
--- a/docs/config.md
+++ b/docs/config.md
@@ -71,9 +71,6 @@ This document explains how to set up and configure the application using the `co
 #### EmbedURL (`"http://localhost:8082/v1/embeddings"`)
 - The endpoint for embedding API, used for RAG (Retrieval Augmented Generation) functionality.

-#### RAGEnabled (`false`)
- Enable or disable RAG functionality for enhanced context retrieval.
-
 #### RAGBatchSize (`1`)
 - Number of documents to process in each RAG batch.

--- a/extra/kokoro_onnx.go
+++ b/extra/kokoro_onnx.go
@@ -0,0 +1,421 @@
+//go:build extra
+// +build extra
+
+package extra
+
+import (
+	"bytes"
+	"fmt"
+	"gf-lt/models"
+	"gf-lt/onnx"
+	"log/slog"
+	"os/exec"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/gopxl/beep/v2"
+	"github.com/gopxl/beep/v2/speaker"
+	"github.com/gopxl/beep/v2/wav"
+	"github.com/neurosnap/sentences/english"
+	"github.com/yalue/onnxruntime_go"
+)
+
+// KokoroONNXOrator implements Kokoro TTS using ONNX runtime
+type KokoroONNXOrator struct {
+	logger        *slog.Logger
+	mu            sync.Mutex
+	session       *onnxruntime_go.DynamicAdvancedSession
+	phonemeMap    map[string]int
+	espeakCmd     string
+	voice         string
+	speed         float32
+	styleVector   []float32
+	currentStream *beep.Ctrl
+	currentDone   chan bool
+	textBuffer    strings.Builder
+	interrupt     bool
+	modelLoaded   bool
+	modelPath     string
+	voicesPath    string
+}
+
+// Phoneme to token ID mapping from Kokoro tokenizer.json
+var kokoroPhonemeMap = map[string]int{
+	"$": 0, ";": 1, ":": 2, ",": 3, ".": 4, "!": 5, "?": 6, "—": 9, "…": 10, "\"": 11, "(": 12, ")": 13, "“": 14, "”": 15, " ": 16, "̃": 17, "ˢ": 18, "ˤ": 19, "˦": 20, "˨": 21, "ᾝ": 22, "⭧": 23,
+	"A": 24, "I": 25, "O": 31, "Q": 33, "S": 35, "T": 36, "W": 39, "Y": 41, "ʲ": 42,
+	"a": 43, "b": 44, "c": 45, "d": 46, "e": 47, "f": 48, "h": 50, "i": 51, "j": 52, "k": 53, "l": 54, "m": 55, "n": 56, "o": 57, "p": 58, "q": 59, "r": 60, "s": 61, "t": 62, "u": 63, "v": 64, "w": 65, "x": 66, "y": 67, "z": 68,
+	"ɑ": 69, "ɐ": 70, "ɒ": 71, "æ": 72, "β": 75, "ɔ": 76, "ɕ": 77, "ç": 78, "ɖ": 80, "ð": 81, "˔": 82, "ə": 83, "ɚ": 85, "ɛ": 86, "ɜ": 87, "ɟ": 90, "ɡ": 92, "ɥ": 99, "ɨ": 101, "ɪ": 102, "ɝ": 103, "ɯ": 110, "ɰ": 111, "ŋ": 112, "ɳ": 113, "ɲ": 114, "ɴ": 115, "ø": 116, "ɸ": 118, "θ": 119, "œ": 120, "ɹ": 123, "ɾ": 125, "ɺ": 126, "ʁ": 128, "ɽ": 129, "ʂ": 130, "ʃ": 131, "ʈ": 132, "˧": 133, "ʊ": 135, "ʋ": 136, "ʌ": 138, "ɢ": 139, "ɣ": 140, "χ": 142, "ʎ": 143, "ʒ": 147, "ʔ": 148,
+	"ˈ": 156, "ˌ": 157, "ː": 158, "̰": 162, "̊": 164, "↕": 169, "→": 171, "↗": 172, "↘": 173, "ᶻ": 177,
+}
+
+func (o *KokoroONNXOrator) ensureInitialized(modelPath string) error {
+	if o.modelLoaded {
+		return nil
+	}
+	o.mu.Lock()
+	defer o.mu.Unlock()
+	if o.modelLoaded {
+		return nil
+	}
+	if modelPath == "" {
+		o.logger.Error("modelPath is empty, cannot load ONNX model")
+		return fmt.Errorf("modelPath is empty, set KokoroModelPath in config")
+	}
+	// Initialize ONNX runtime (shared with embedder)
+	if err := onnx.Init(); err != nil {
+		o.logger.Error("ONNX init failed", "error", err)
+		return fmt.Errorf("ONNX init failed: %w", err)
+	}
+	if onnx.HasCUDASupport() {
+		o.logger.Info("ONNX using CUDA")
+	} else {
+		o.logger.Info("ONNX using CPU fallback")
+	}
+	if o.phonemeMap == nil {
+		o.phonemeMap = kokoroPhonemeMap
+	}
+	if o.espeakCmd == "" {
+		o.espeakCmd = "espeak-ng"
+		if _, err := exec.LookPath(o.espeakCmd); err != nil {
+			o.espeakCmd = "espeak"
+			if _, err := exec.LookPath(o.espeakCmd); err != nil {
+				return fmt.Errorf("espeak-ng or espeak not found. Install with: sudo apt-get install espeak-ng")
+			}
+		}
+	}
+	o.logger.Info("using espeak command", "cmd", o.espeakCmd)
+	// Load voice embedding if not already loaded
+	if o.styleVector == nil {
+		voiceName := o.voice
+		if voiceName == "" {
+			voiceName = "af_bella"
+		}
+		if o.voicesPath != "" {
+			styleVec, err := onnx.LoadVoice(o.voicesPath, voiceName)
+			if err != nil {
+				o.logger.Warn("failed to load voice, using zeros", "error", err, "voice", voiceName)
+				o.styleVector = make([]float32, 256)
+			} else {
+				// Shape is (510, 1, 256), we want the last 256 values (or first? let's use mean or just pick one)
+				// Actually, let's average across all 510 to get a single 256-dim vector
+				if len(styleVec) != 510*256 {
+					o.logger.Error("voice embedding has unexpected size", "len", len(styleVec))
+					err = fmt.Errorf("voice embedding has unexpected size", "len", len(styleVec))
+					return err
+				}
+				o.styleVector = make([]float32, 256)
+				for i := 0; i < 256; i++ {
+					var sum float32
+					for j := 0; j < 510; j++ {
+						sum += styleVec[j*256+i]
+					}
+					o.styleVector[i] = sum / 510.0
+				}
+				o.logger.Info("loaded voice embedding", "voice", voiceName)
+			}
+		} else {
+			o.logger.Warn("no voices path configured, using zeros for style")
+			o.styleVector = make([]float32, 256)
+		}
+	}
+	opts, err := onnx.NewSessionOptions()
+	if err != nil {
+		return fmt.Errorf("failed to create session options: %w", err)
+	}
+	defer func() { _ = opts.Destroy() }()
+	if onnx.HasCUDASupport() {
+		o.logger.Info("session options created with CUDA")
+	} else {
+		o.logger.Info("session options created with CPU")
+	}
+	session, err := onnxruntime_go.NewDynamicAdvancedSession(
+		modelPath,
+		[]string{"input_ids", "style", "speed"},
+		[]string{"waveform"},
+		opts,
+	)
+	if err != nil {
+		o.logger.Error("failed to create ONNX session", "error", err)
+		return fmt.Errorf("failed to create ONNX session: %w", err)
+	}
+	o.session = session
+	o.modelLoaded = true
+	o.logger.Info("Kokoro ONNX model loaded successfully", "model", modelPath)
+	return nil
+}
+
+func (o *KokoroONNXOrator) textToPhonemes(text string) (string, error) {
+	cmd := exec.Command(o.espeakCmd, "-x", "-q", text)
+	output, err := cmd.Output()
+	if err != nil {
+		o.logger.Error("espeak failed", "error", err, "cmd", o.espeakCmd, "text", text)
+		return "", fmt.Errorf("espeak failed: %w", err)
+	}
+
+	phonemeStr := strings.TrimSpace(string(output))
+	return phonemeStr, nil
+}
+
+func (o *KokoroONNXOrator) phonemesToTokens(phonemeStr string) ([]int, error) {
+	if phonemeStr == "" {
+		o.logger.Error("empty phoneme string")
+		return nil, fmt.Errorf("empty phoneme string")
+	}
+	// Iterate over each character in the phoneme string
+	tokens := make([]int, 0)
+	for _, ch := range phonemeStr {
+		chStr := string(ch)
+		if tokenID, ok := o.phonemeMap[chStr]; ok {
+			tokens = append(tokens, tokenID)
+		}
+	}
+	if len(tokens) == 0 {
+		o.logger.Error("no phonemes mapped to tokens", "phonemeStr", phonemeStr)
+		return nil, fmt.Errorf("no valid phonemes mapped to tokens")
+	}
+	return tokens, nil
+}
+
+func (o *KokoroONNXOrator) generateAudio(text string) ([]float32, error) {
+	if err := o.ensureInitialized(o.modelPath); err != nil {
+		o.logger.Error("ensureInitialized failed", "error", err)
+		return nil, err
+	}
+	phonemeStr, err := o.textToPhonemes(text)
+	if err != nil {
+		o.logger.Error("phoneme conversion failed", "error", err)
+		return nil, fmt.Errorf("phoneme conversion failed: %w", err)
+	}
+	tokens, err := o.phonemesToTokens(phonemeStr)
+	if err != nil {
+		o.logger.Error("token conversion failed", "error", err)
+		return nil, fmt.Errorf("token conversion failed: %w", err)
+	}
+	if len(tokens) > 510 {
+		return nil, fmt.Errorf("text too long: %d tokens (max 510)", len(tokens))
+	}
+	tokens = append([]int{0}, tokens...)
+	tokens = append(tokens, 0)
+	inputIDs := make([]int64, len(tokens))
+	for i, t := range tokens {
+		inputIDs[i] = int64(t)
+	}
+	inputTensor, err := onnxruntime_go.NewTensor[int64](
+		onnxruntime_go.NewShape(1, int64(len(inputIDs))),
+		inputIDs,
+	)
+	if err != nil {
+		o.logger.Error("failed to create input tensor", "error", err)
+		return nil, fmt.Errorf("failed to create input tensor: %w", err)
+	}
+	defer func() { _ = inputTensor.Destroy() }()
+	styleTensor, err := onnxruntime_go.NewTensor[float32](
+		onnxruntime_go.NewShape(1, 256),
+		o.styleVector,
+	)
+	if err != nil {
+		o.logger.Error("failed to create style tensor", "error", err)
+		return nil, fmt.Errorf("failed to create style tensor: %w", err)
+	}
+	defer func() { _ = styleTensor.Destroy() }()
+	speedTensor, err := onnxruntime_go.NewTensor[float32](
+		onnxruntime_go.NewShape(1),
+		[]float32{o.speed},
+	)
+	if err != nil {
+		o.logger.Error("failed to create speed tensor", "error", err)
+		return nil, fmt.Errorf("failed to create speed tensor: %w", err)
+	}
+	defer func() { _ = speedTensor.Destroy() }()
+	outputTensor, err := onnxruntime_go.NewEmptyTensor[float32](
+		onnxruntime_go.NewShape(1, 512),
+	)
+	if err != nil {
+		o.logger.Error("failed to create output tensor", "error", err)
+		return nil, fmt.Errorf("failed to create output tensor: %w", err)
+	}
+	defer func() { _ = outputTensor.Destroy() }()
+	err = o.session.Run(
+		[]onnxruntime_go.Value{inputTensor, styleTensor, speedTensor},
+		[]onnxruntime_go.Value{outputTensor},
+	)
+	if err != nil {
+		o.logger.Error("ONNX inference failed", "error", err)
+		return nil, fmt.Errorf("ONNX inference failed: %w", err)
+	}
+	audioData := outputTensor.GetData()
+	if len(audioData) == 0 {
+		o.logger.Error("empty audio output from ONNX")
+		return nil, fmt.Errorf("empty audio output")
+	}
+	audio := make([]float32, len(audioData))
+	copy(audio, audioData)
+	return audio, nil
+}
+
+func (o *KokoroONNXOrator) Speak(text string) error {
+	audio, err := o.generateAudio(text)
+	if err != nil {
+		o.logger.Error("audio generation failed", "error", err)
+		return fmt.Errorf("audio generation failed: %w", err)
+	}
+	// Create streamer for encoding
+	encodeStreamer := beep.StreamerFunc(func(samples [][2]float64) (n int, ok bool) {
+		for i := range samples {
+			if i >= len(audio) {
+				return i, false
+			}
+			samples[i][0] = float64(audio[i])
+			samples[i][1] = float64(audio[i])
+		}
+		return len(audio), true
+	})
+	buf := &seekableBuffer{new(bytes.Buffer)}
+	err = wav.Encode(buf, encodeStreamer, beep.Format{
+		SampleRate:  24000,
+		NumChannels: 1,
+		Precision:   2,
+	})
+	if err != nil {
+		o.logger.Error("wav encoding failed", "error", err)
+		return fmt.Errorf("wav encoding failed: %w", err)
+	}
+	decodedStreamer, format, err := wav.Decode(bytes.NewReader(buf.Bytes()))
+	if err != nil {
+		o.logger.Error("wav decode failed", "error", err)
+		return fmt.Errorf("wav decode failed: %w", err)
+	}
+	defer decodedStreamer.Close()
+	if err := speaker.Init(format.SampleRate, format.SampleRate.N(time.Second/10)); err != nil {
+		o.logger.Error("speaker init failed", "error", err)
+		return fmt.Errorf("speaker init failed: %w", err)
+	}
+	o.logger.Info("playing audio", "sampleRate", format.SampleRate, "channels", format.NumChannels)
+	done := make(chan bool)
+	o.mu.Lock()
+	o.currentDone = done
+	o.currentStream = &beep.Ctrl{Streamer: beep.Seq(decodedStreamer, 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 *KokoroONNXOrator) Stop() {
+	speaker.Lock()
+	defer speaker.Unlock()
+	o.mu.Lock()
+	defer o.mu.Unlock()
+	if o.currentStream != nil {
+		o.currentStream.Streamer = nil
+	}
+}
+
+func (o *KokoroONNXOrator) GetLogger() *slog.Logger {
+	return o.logger
+}
+
+func (o *KokoroONNXOrator) stoproutine() {
+	for {
+		<-TTSDoneChan
+		o.Stop()
+		for len(TTSTextChan) > 0 {
+			<-TTSTextChan
+		}
+		o.mu.Lock()
+		o.textBuffer.Reset()
+		if o.currentDone != nil {
+			select {
+			case o.currentDone <- true:
+			default:
+			}
+		}
+		o.interrupt = true
+		o.mu.Unlock()
+	}
+}
+
+func (o *KokoroONNXOrator) 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 buffer", "error", err)
+				o.mu.Unlock()
+				continue
+			}
+			text := o.textBuffer.String()
+			sentences := tokenizer.Tokenize(text)
+			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.Info("KokoroONNX speak", "text", cleanedText)
+				if err := o.Speak(cleanedText); err != nil {
+					o.logger.Error("KokoroONNX tts failed", "text", cleanedText, "error", err)
+				}
+			}
+		case <-TTSFlushChan:
+			if len(TTSTextChan) > 0 {
+				for chunk := range TTSTextChan {
+					o.mu.Lock()
+					_, err := o.textBuffer.WriteString(chunk)
+					o.mu.Unlock()
+					if err != nil {
+						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
+			}
+			sentencesRem := tokenizer.Tokenize(remaining)
+			for _, rs := range sentencesRem {
+				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", "text", rs.Text, "error", err)
+				}
+			}
+		}
+	}
+}
--- a/extra/tts.go
+++ b/extra/tts.go
@@ -32,6 +32,14 @@ var (
 	// endsWithPunctuation = regexp.MustCompile(`[;.!?]$`)
 )

+type seekableBuffer struct {
+	*bytes.Buffer
+}
+
+func (s *seekableBuffer) Seek(offset int64, whence int) (int64, error) {
+	return 0, nil
+}
+
 type Orator interface {
 	Speak(text string) error
 	Stop()
@@ -194,6 +202,18 @@ func NewOrator(log *slog.Logger, cfg *config.Config) Orator {
 		go orator.readroutine()
 		go orator.stoproutine()
 		return orator
+	case "kokoro_onnx":
+		log.Info("Initializing Kokoro ONNX TTS", "modelPath", cfg.KokoroModelPath, "voicesPath", cfg.KokoroVoicesPath, "voice", cfg.KokoroVoice, "speed", cfg.TTS_SPEED)
+		orator := &KokoroONNXOrator{
+			logger:     log,
+			modelPath:  cfg.KokoroModelPath,
+			voicesPath: cfg.KokoroVoicesPath,
+			speed:      cfg.TTS_SPEED,
+			voice:      cfg.KokoroVoice,
+		}
+		go orator.readroutine()
+		go orator.stoproutine()
+		return orator
 	default:
 		language := cfg.TTS_LANGUAGE
 		if language == "" {
--- a/go.mod
+++ b/go.mod
@@ -7,7 +7,6 @@ require (
 	github.com/GrailFinder/google-translate-tts v0.1.3
 	github.com/GrailFinder/searchagent v0.2.0
 	github.com/PuerkitoBio/goquery v1.11.0
-	github.com/deckarep/golang-set/v2 v2.8.0
 	github.com/gdamore/tcell/v2 v2.13.2
 	github.com/glebarez/go-sqlite v1.22.0
 	github.com/gopxl/beep/v2 v2.1.1
@@ -17,14 +16,18 @@ require (
 	github.com/neurosnap/sentences v1.1.2
 	github.com/playwright-community/playwright-go v0.5700.1
 	github.com/rivo/tview v0.42.0
+	github.com/sugarme/tokenizer v0.3.0
+	github.com/yalue/onnxruntime_go v1.27.0
 	github.com/yuin/goldmark v1.4.13
 )

 require (
 	github.com/andybalholm/cascadia v1.3.3 // indirect
+	github.com/deckarep/golang-set/v2 v2.8.0 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/ebitengine/oto/v3 v3.4.0 // indirect
 	github.com/ebitengine/purego v0.9.1 // indirect
+	github.com/emirpasic/gods v1.18.1 // indirect
 	github.com/gdamore/encoding v1.0.1 // indirect
 	github.com/go-jose/go-jose/v3 v3.0.4 // indirect
 	github.com/go-stack/stack v1.8.1 // indirect
@@ -33,10 +36,14 @@ require (
 	github.com/hajimehoshi/oto/v2 v2.3.1 // indirect
 	github.com/lucasb-eyer/go-colorful v1.3.0 // indirect
 	github.com/mattn/go-isatty v0.0.20 // indirect
+	github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db // indirect
 	github.com/ncruces/go-strftime v1.0.0 // indirect
+	github.com/patrickmn/go-cache v2.1.0+incompatible // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
 	github.com/rivo/uniseg v0.4.7 // indirect
+	github.com/schollz/progressbar/v2 v2.15.0 // indirect
+	github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c // indirect
 	golang.org/x/exp v0.0.0-20251209150349-8475f28825e9 // indirect
 	golang.org/x/net v0.48.0 // indirect
 	golang.org/x/sys v0.39.0 // indirect
--- a/go.sum
+++ b/go.sum
@@ -21,6 +21,8 @@ github.com/ebitengine/oto/v3 v3.4.0 h1:br0PgASsEWaoWn38b2Goe7m1GKFYfNgnsjSd5Gg+/
 github.com/ebitengine/oto/v3 v3.4.0/go.mod h1:IOleLVD0m+CMak3mRVwsYY8vTctQgOM0iiL6S7Ar7eI=
 github.com/ebitengine/purego v0.9.1 h1:a/k2f2HQU3Pi399RPW1MOaZyhKJL9w/xFpKAg4q1s0A=
 github.com/ebitengine/purego v0.9.1/go.mod h1:iIjxzd6CiRiOG0UyXP+V1+jWqUXVjPKLAI0mRfJZTmQ=
+github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc=
+github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ=
 github.com/gdamore/encoding v1.0.1 h1:YzKZckdBL6jVt2Gc+5p82qhrGiqMdG/eNs6Wy0u3Uhw=
 github.com/gdamore/encoding v1.0.1/go.mod h1:0Z0cMFinngz9kS1QfMjCP8TY7em3bZYeeklsSDPivEo=
 github.com/gdamore/tcell/v2 v2.13.2 h1:5j4srfF8ow3HICOv/61/sOhQtA25qxEB2XR3Q/Bhx2g=
@@ -61,10 +63,14 @@ github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWE
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/mattn/go-sqlite3 v1.14.22 h1:2gZY6PC6kBnID23Tichd1K+Z0oS6nE/XwU+Vz/5o4kU=
 github.com/mattn/go-sqlite3 v1.14.22/go.mod h1:Uh1q+B4BYcTPb+yiD3kU8Ct7aC0hY9fxUwlHK0RXw+Y=
+github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db h1:62I3jR2EmQ4l5rM/4FEfDWcRD+abF5XlKShorW5LRoQ=
+github.com/mitchellh/colorstring v0.0.0-20190213212951-d06e56a500db/go.mod h1:l0dey0ia/Uv7NcFFVbCLtqEBQbrT4OCwCSKTEv6enCw=
 github.com/ncruces/go-strftime v1.0.0 h1:HMFp8mLCTPp341M/ZnA4qaf7ZlsbTc+miZjCLOFAw7w=
 github.com/ncruces/go-strftime v1.0.0/go.mod h1:Fwc5htZGVVkseilnfgOVb9mKy6w1naJmn9CehxcKcls=
 github.com/neurosnap/sentences v1.1.2 h1:iphYOzx/XckXeBiLIUBkPu2EKMJ+6jDbz/sLJZ7ZoUw=
 github.com/neurosnap/sentences v1.1.2/go.mod h1:/pwU4E9XNL21ygMIkOIllv/SMy2ujHwpf8GQPu1YPbQ=
+github.com/patrickmn/go-cache v2.1.0+incompatible h1:HRMgzkcYKYpi3C8ajMPV8OFXaaRUnok+kx1WdO15EQc=
+github.com/patrickmn/go-cache v2.1.0+incompatible/go.mod h1:3Qf8kWWT7OJRJbdiICTKqZju1ZixQ/KpMGzzAfe6+WQ=
 github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
 github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/playwright-community/playwright-go v0.5700.1 h1:PNFb1byWqrTT720rEO0JL88C6Ju0EmUnR5deFLvtP/U=
@@ -77,10 +83,19 @@ github.com/rivo/tview v0.42.0 h1:b/ftp+RxtDsHSaynXTbJb+/n/BxDEi+W3UfF5jILK6c=
 github.com/rivo/tview v0.42.0/go.mod h1:cSfIYfhpSGCjp3r/ECJb+GKS7cGJnqV8vfjQPwoXyfY=
 github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=
 github.com/rivo/uniseg v0.4.7/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUcx88=
+github.com/schollz/progressbar/v2 v2.15.0 h1:dVzHQ8fHRmtPjD3K10jT3Qgn/+H+92jhPrhmxIJfDz8=
+github.com/schollz/progressbar/v2 v2.15.0/go.mod h1:UdPq3prGkfQ7MOzZKlDRpYKcFqEMczbD7YmbPgpzKMI=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
 github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.10.0 h1:Xv5erBjTwe/5IxqUQTdXv5kgmIvbHo3QQyRwhJsOfJA=
 github.com/stretchr/testify v1.10.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
+github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c h1:pwb4kNSHb4K89ymCaN+5lPH/MwnfSVg4rzGDh4d+iy4=
+github.com/sugarme/regexpset v0.0.0-20200920021344-4d4ec8eaf93c/go.mod h1:2gwkXLWbDGUQWeL3RtpCmcY4mzCtU13kb9UsAg9xMaw=
+github.com/sugarme/tokenizer v0.3.0 h1:FE8DYbNSz/kSbgEo9l/RjgYHkIJYEdskumitFQBE9FE=
+github.com/sugarme/tokenizer v0.3.0/go.mod h1:VJ+DLK5ZEZwzvODOWwY0cw+B1dabTd3nCB5HuFCItCc=
+github.com/yalue/onnxruntime_go v1.27.0 h1:c1YSgDNtpf0WGtxj3YeRIb8VC5LmM1J+Ve3uHdteC1U=
+github.com/yalue/onnxruntime_go v1.27.0/go.mod h1:b4X26A8pekNb1ACJ58wAXgNKeUCGEAQ9dmACut9Sm/4=
 github.com/yuin/goldmark v1.4.13 h1:fVcFKWvrslecOb/tg+Cc05dkeYx540o0FuFt3nUVDoE=
 github.com/yuin/goldmark v1.4.13/go.mod h1:6yULJ656Px+3vBD8DxQVa3kxgyrAnzto9xy5taEt/CY=
 golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=
--- a/helpfuncs.go
+++ b/helpfuncs.go
@@ -521,7 +521,7 @@ 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)
--- a/props_table.go
+++ b/props_table.go
@@ -115,9 +115,6 @@ func makePropsTable(props map[string]float32) *tview.Table {
 		row++
 	}
 	// Add checkboxes
-	addCheckboxRow("RAG use", cfg.RAGEnabled, func(checked bool) {
-		cfg.RAGEnabled = checked
-	})
 	addCheckboxRow("Inject role", injectRole, func(checked bool) {
 		injectRole = checked
 	})
--- a/rag/embedder.go
+++ b/rag/embedder.go
@@ -7,8 +7,16 @@ import (
 	"fmt"
 	"gf-lt/config"
 	"gf-lt/models"
+	"gf-lt/onnx"
 	"log/slog"
 	"net/http"
+	"os"
+	"sync"
+	"time"
+
+	"github.com/sugarme/tokenizer"
+	"github.com/sugarme/tokenizer/pretrained"
+	"github.com/yalue/onnxruntime_go"
 )

 // Embedder defines the interface for embedding text
@@ -27,7 +35,9 @@ type APIEmbedder struct {
 func NewAPIEmbedder(l *slog.Logger, cfg *config.Config) *APIEmbedder {
 	return &APIEmbedder{
 		logger: l,
-		client: &http.Client{},
+		client: &http.Client{
+			Timeout: 30 * time.Second,
+		},
 		cfg: cfg,
 	}
 }
@@ -134,11 +144,230 @@ 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
+}
+
+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)
+	}
+
+	// Initialize ONNX runtime
+	if err := onnx.Init(); err != nil {
+		return nil, fmt.Errorf("ONNX init failed: %w", err)
+	}
+	if onnx.HasCUDASupport() {
+		logger.Info("ONNX CUDA support enabled")
+	} else {
+		logger.Info("ONNX using CPU fallback")
+	}
+
+	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
+	}
+	// ONNX runtime already initialized by onnx.Init() in NewONNXEmbedder
+	if !onnx.IsReady() {
+		return errors.New("ONNX runtime not ready")
+	}
+	// Create session options
+	opts, err := onnx.NewSessionOptions()
+	if err != nil {
+		return fmt.Errorf("failed to create session options: %w", err)
+	}
+	defer func() {
+		_ = opts.Destroy()
+	}()
+
+	if onnx.HasCUDASupport() {
+		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
@@ -1,6 +1,7 @@
 package rag

 import (
+	"context"
 	"errors"
 	"fmt"
 	"gf-lt/config"
@@ -9,16 +10,20 @@ import (
 	"log/slog"
 	"path"
 	"regexp"
+	"runtime"
 	"sort"
 	"strings"
 	"sync"
+	"time"

 	"github.com/neurosnap/sentences/english"
 )

+const ()
+
 var (
 	// Status messages for TUI integration
-	LongJobStatusCh     = make(chan string, 10) // Increased buffer size to prevent blocking
+	LongJobStatusCh     = make(chan string, 100) // Increased buffer size for parallel batch updates
 	FinishedRAGStatus   = "finished loading RAG file; press Enter"
 	LoadedFileRAGStatus = "loaded file"
 	ErrRAGStatus        = "some error occurred; failed to transfer data to vector db"
@@ -30,30 +35,143 @@ type RAG struct {
 	cfg         *config.Config
 	embedder    Embedder
 	storage     *VectorStorage
-	mu       sync.Mutex
+	mu          sync.RWMutex
+	idleMu      sync.Mutex
+	fallbackMsg string
+	idleTimer   *time.Timer
+	idleTimeout time.Duration
 }

-func New(l *slog.Logger, s storage.FullRepo, cfg *config.Config) *RAG {
-	// Initialize with API embedder by default, could be configurable later
-	embedder := NewAPIEmbedder(l, cfg)
+// batchTask represents a single batch to be embedded
+type batchTask struct {
+	batchIndex   int
+	paragraphs   []string
+	filename     string
+	totalBatches int
+}
+
+// batchResult represents the result of embedding a batch
+type batchResult struct {
+	batchIndex int
+	embeddings [][]float32
+	paragraphs []string
+	filename   string
+}
+
+// sendStatusNonBlocking sends a status message without blocking
+func (r *RAG) sendStatusNonBlocking(status string) {
+	select {
+	case LongJobStatusCh <- status:
+	default:
+		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", status)
+	}
+}
+
+func New(l *slog.Logger, s storage.FullRepo, cfg *config.Config) (*RAG, error) {
+	var embedder Embedder
+	var fallbackMsg string
+	if cfg.EmbedModelPath != "" && cfg.EmbedTokenizerPath != "" {
+		emb, err := NewONNXEmbedder(cfg.EmbedModelPath, cfg.EmbedTokenizerPath, cfg.EmbedDims, l)
+		if err != nil {
+			l.Error("failed to create ONNX embedder, falling back to API", "error", err)
+			fallbackMsg = err.Error()
+			embedder = NewAPIEmbedder(l, cfg)
+		} else {
+			embedder = emb
+			l.Info("using ONNX embedder", "model", cfg.EmbedModelPath, "dims", cfg.EmbedDims)
+		}
+	} else {
+		embedder = NewAPIEmbedder(l, cfg)
+		l.Info("using API embedder", "url", cfg.EmbedURL)
+	}
 	rag := &RAG{
 		logger:      l,
 		store:       s,
 		cfg:         cfg,
 		embedder:    embedder,
 		storage:     NewVectorStorage(l, s),
+		fallbackMsg: fallbackMsg,
+		idleTimeout: 30 * time.Second,
 	}

 	// Note: Vector tables are created via database migrations, not at runtime

-	return rag
+	return rag, nil
 }

-func wordCounter(sentence string) int {
-	return len(strings.Split(strings.TrimSpace(sentence), " "))
+func createChunks(sentences []string, wordLimit, overlapWords uint32) []string {
+	if len(sentences) == 0 {
+		return nil
+	}
+	if overlapWords >= wordLimit {
+		overlapWords = wordLimit / 2
+	}
+	var chunks []string
+	i := 0
+	for i < len(sentences) {
+		var chunkWords []string
+		wordCount := 0
+		j := i
+		for j < len(sentences) && wordCount <= int(wordLimit) {
+			sentence := sentences[j]
+			words := strings.Fields(sentence)
+			chunkWords = append(chunkWords, sentence)
+			wordCount += len(words)
+			j++
+			// If this sentence alone exceeds limit, still include it and stop
+			if wordCount > int(wordLimit) {
+				break
+			}
+		}
+		if len(chunkWords) == 0 {
+			break
+		}
+		chunk := strings.Join(chunkWords, " ")
+		chunks = append(chunks, chunk)
+		if j >= len(sentences) {
+			break
+		}
+		// Move i forward by skipping overlap
+		if overlapWords == 0 {
+			i = j
+			continue
+		}
+		// Calculate how many sentences to skip to achieve overlapWords
+		overlapRemaining := int(overlapWords)
+		newI := i
+		for newI < j && overlapRemaining > 0 {
+			words := len(strings.Fields(sentences[newI]))
+			overlapRemaining -= words
+			if overlapRemaining >= 0 {
+				newI++
+			}
+		}
+		if newI == i {
+			newI = j
+		}
+		i = newI
+	}
+	return chunks
+}
+
+func sanitizeFTSQuery(query string) string {
+	// Remove double quotes and other problematic characters for FTS5
+	query = strings.ReplaceAll(query, "\"", " ")
+	query = strings.ReplaceAll(query, "'", " ")
+	query = strings.ReplaceAll(query, ";", " ")
+	query = strings.ReplaceAll(query, "\\", " ")
+	query = strings.TrimSpace(query)
+	if query == "" {
+		return "*" // match all
+	}
+	return query
 }

 func (r *RAG) LoadRAG(fpath string) error {
+	return r.LoadRAGWithContext(context.Background(), fpath)
+}
+
+func (r *RAG) LoadRAGWithContext(ctx context.Context, fpath string) error {
 	r.mu.Lock()
 	defer r.mu.Unlock()
 	fileText, err := ExtractText(fpath)
@@ -61,11 +179,9 @@ func (r *RAG) LoadRAG(fpath string) error {
 		return err
 	}
 	r.logger.Debug("rag: loaded file", "fp", fpath)
-	select {
-	case LongJobStatusCh <- LoadedFileRAGStatus:
-	default:
-		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", LoadedFileRAGStatus)
-	}
+
+	// Send initial status (non-blocking with retry)
+	r.sendStatusNonBlocking(LoadedFileRAGStatus)
 	tokenizer, err := english.NewSentenceTokenizer(nil)
 	if err != nil {
 		return err
@@ -75,31 +191,9 @@ func (r *RAG) LoadRAG(fpath string) error {
 	for i, s := range sentences {
 		sents[i] = s.Text
 	}
-	// Group sentences into paragraphs based on word limit
-	paragraphs := []string{}
-	par := strings.Builder{}
-	for i := 0; i < len(sents); i++ {
-		if strings.TrimSpace(sents[i]) != "" {
-			if par.Len() > 0 {
-				par.WriteString(" ")
-			}
-			par.WriteString(sents[i])
-		}
-		if wordCounter(par.String()) > int(r.cfg.RAGWordLimit) {
-			paragraph := strings.TrimSpace(par.String())
-			if paragraph != "" {
-				paragraphs = append(paragraphs, paragraph)
-			}
-			par.Reset()
-		}
-	}
-	// Handle any remaining content in the paragraph buffer
-	if par.Len() > 0 {
-		paragraph := strings.TrimSpace(par.String())
-		if paragraph != "" {
-			paragraphs = append(paragraphs, paragraph)
-		}
-	}
+
+	// Create chunks with overlap
+	paragraphs := createChunks(sents, r.cfg.RAGWordLimit, r.cfg.RAGOverlapWords)
 	// Adjust batch size if needed
 	if len(paragraphs) < r.cfg.RAGBatchSize && len(paragraphs) > 0 {
 		r.cfg.RAGBatchSize = len(paragraphs)
@@ -107,15 +201,61 @@ func (r *RAG) LoadRAG(fpath string) error {
 	if len(paragraphs) == 0 {
 		return errors.New("no valid paragraphs found in file")
 	}
-	// Process paragraphs in batches synchronously
-	batchCount := 0
-	for i := 0; i < len(paragraphs); i += r.cfg.RAGBatchSize {
-		end := i + r.cfg.RAGBatchSize
+	totalBatches := (len(paragraphs) + r.cfg.RAGBatchSize - 1) / r.cfg.RAGBatchSize
+	r.logger.Debug("starting parallel embedding", "total_batches", totalBatches, "batch_size", r.cfg.RAGBatchSize)
+
+	// Determine concurrency level
+	concurrency := runtime.NumCPU()
+	if concurrency > totalBatches {
+		concurrency = totalBatches
+	}
+	if concurrency < 1 {
+		concurrency = 1
+	}
+	// If using ONNX embedder, limit concurrency to 1 due to mutex serialization
+	var isONNX bool
+	if _, isONNX = r.embedder.(*ONNXEmbedder); isONNX {
+		concurrency = 1
+	}
+	embedderType := "API"
+	if isONNX {
+		embedderType = "ONNX"
+	}
+	r.logger.Debug("parallel embedding setup",
+		"total_batches", totalBatches,
+		"concurrency", concurrency,
+		"embedder", embedderType,
+		"batch_size", r.cfg.RAGBatchSize)
+
+	// Create context with timeout (30 minutes) and cancellation for error handling
+	ctx, cancel := context.WithTimeout(ctx, 30*time.Minute)
+	defer cancel()
+
+	// Channels for task distribution and results
+	taskCh := make(chan batchTask, totalBatches)
+	resultCh := make(chan batchResult, totalBatches)
+	errorCh := make(chan error, totalBatches)
+
+	// Start worker goroutines
+	var wg sync.WaitGroup
+	for w := 0; w < concurrency; w++ {
+		wg.Add(1)
+		go r.embeddingWorker(ctx, w, taskCh, resultCh, errorCh, &wg)
+	}
+
+	// Close task channel after all tasks are sent (by separate goroutine)
+	go func() {
+		// Ensure task channel is closed when this goroutine exits
+		defer close(taskCh)
+		r.logger.Debug("task distributor started", "total_batches", totalBatches)
+		for i := 0; i < totalBatches; i++ {
+			start := i * r.cfg.RAGBatchSize
+			end := start + r.cfg.RAGBatchSize
 			if end > len(paragraphs) {
 				end = len(paragraphs)
 			}
-		batch := paragraphs[i:end]
-		batchCount++
+			batch := paragraphs[start:end]
+
 			// Filter empty paragraphs
 			nonEmptyBatch := make([]string, 0, len(batch))
 			for _, p := range batch {
@@ -123,75 +263,286 @@ func (r *RAG) LoadRAG(fpath string) error {
 					nonEmptyBatch = append(nonEmptyBatch, strings.TrimSpace(p))
 				}
 			}
-		if len(nonEmptyBatch) == 0 {
+
+			task := batchTask{
+				batchIndex:   i,
+				paragraphs:   nonEmptyBatch,
+				filename:     path.Base(fpath),
+				totalBatches: totalBatches,
+			}
+
+			select {
+			case taskCh <- task:
+				r.logger.Debug("task distributor sent batch", "batch", i, "paragraphs", len(nonEmptyBatch))
+			case <-ctx.Done():
+				r.logger.Debug("task distributor cancelled", "batches_sent", i+1, "total_batches", totalBatches)
+				return
+			}
+		}
+		r.logger.Debug("task distributor finished", "batches_sent", totalBatches)
+	}()
+
+	// Wait for workers to finish and close result channel
+	go func() {
+		wg.Wait()
+		close(resultCh)
+	}()
+
+	// Process results in order and write to database
+	nextExpectedBatch := 0
+	resultsBuffer := make(map[int]batchResult)
+	filename := path.Base(fpath)
+	batchesProcessed := 0
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+
+		case err := <-errorCh:
+			// First error from any worker, cancel everything
+			cancel()
+			r.logger.Error("embedding worker failed", "error", err)
+			r.sendStatusNonBlocking(ErrRAGStatus)
+			return fmt.Errorf("embedding failed: %w", err)
+
+		case result, ok := <-resultCh:
+			if !ok {
+				// All results processed
+				resultCh = nil
+				r.logger.Debug("result channel closed", "batches_processed", batchesProcessed, "total_batches", totalBatches)
 				continue
 			}
-		// Embed the batch
-		embeddings, err := r.embedder.EmbedSlice(nonEmptyBatch)
-		if err != nil {
-			r.logger.Error("failed to embed batch", "error", err, "batch", batchCount)
-			select {
-			case LongJobStatusCh <- ErrRAGStatus:
-			default:
-				r.logger.Warn("LongJobStatusCh channel full, dropping message")
-			}
-			return fmt.Errorf("failed to embed batch %d: %w", batchCount, err)
-		}
-		if len(embeddings) != len(nonEmptyBatch) {
-			err := errors.New("embedding count mismatch")
-			r.logger.Error("embedding mismatch", "expected", len(nonEmptyBatch), "got", len(embeddings))
+
+			// Store result in buffer
+			resultsBuffer[result.batchIndex] = result
+
+			// Process buffered results in order
+			for {
+				if res, exists := resultsBuffer[nextExpectedBatch]; exists {
+					// Write this batch to database
+					if err := r.writeBatchToStorage(ctx, res, filename); err != nil {
+						cancel()
 						return err
 					}
-		// Write vectors to storage
-		filename := path.Base(fpath)
-		for j, text := range nonEmptyBatch {
-			vector := models.VectorRow{
-				Embeddings: embeddings[j],
+
+					batchesProcessed++
+					// Send progress update
+					statusMsg := fmt.Sprintf("processed batch %d/%d", batchesProcessed, totalBatches)
+					r.sendStatusNonBlocking(statusMsg)
+
+					delete(resultsBuffer, nextExpectedBatch)
+					nextExpectedBatch++
+				} else {
+					break
+				}
+			}
+
+		default:
+			// No channels ready, check for deadlock conditions
+			if resultCh == nil && nextExpectedBatch < totalBatches {
+				// Missing batch results after result channel closed
+				r.logger.Error("missing batch results",
+					"expected", totalBatches,
+					"received", nextExpectedBatch,
+					"missing", totalBatches-nextExpectedBatch)
+
+				// Wait a short time for any delayed errors, then cancel
+				select {
+				case <-time.After(5 * time.Second):
+					cancel()
+					return fmt.Errorf("missing batch results: expected %d, got %d", totalBatches, nextExpectedBatch)
+				case <-ctx.Done():
+					return ctx.Err()
+				case err := <-errorCh:
+					cancel()
+					r.logger.Error("embedding worker failed after result channel closed", "error", err)
+					r.sendStatusNonBlocking(ErrRAGStatus)
+					return fmt.Errorf("embedding failed: %w", err)
+				}
+			}
+			// If we reach here, no deadlock yet, just busy loop prevention
+			time.Sleep(100 * time.Millisecond)
+		}
+
+		// Check if we're done
+		if resultCh == nil && nextExpectedBatch >= totalBatches {
+			r.logger.Debug("all batches processed successfully", "total", totalBatches)
+			break
+		}
+	}
+	r.logger.Debug("finished writing vectors", "batches", batchesProcessed)
+	r.resetIdleTimer()
+	r.sendStatusNonBlocking(FinishedRAGStatus)
+	return nil
+}
+
+// embeddingWorker processes batch embedding tasks
+func (r *RAG) embeddingWorker(ctx context.Context, workerID int, taskCh <-chan batchTask, resultCh chan<- batchResult, errorCh chan<- error, wg *sync.WaitGroup) {
+	defer wg.Done()
+	r.logger.Debug("embedding worker started", "worker", workerID)
+
+	// Panic recovery to ensure worker doesn't crash silently
+	defer func() {
+		if rec := recover(); rec != nil {
+			r.logger.Error("embedding worker panicked", "worker", workerID, "panic", rec)
+			// Try to send error, but don't block if channel is full
+			select {
+			case errorCh <- fmt.Errorf("worker %d panicked: %v", workerID, rec):
+			default:
+				r.logger.Warn("error channel full, dropping panic error", "worker", workerID)
+			}
+		}
+	}()
+	for task := range taskCh {
+		select {
+		case <-ctx.Done():
+			r.logger.Debug("embedding worker cancelled", "worker", workerID)
+			return
+		default:
+		}
+		r.logger.Debug("worker processing batch", "worker", workerID, "batch", task.batchIndex, "paragraphs", len(task.paragraphs), "total_batches", task.totalBatches)
+
+		// Skip empty batches
+		if len(task.paragraphs) == 0 {
+			select {
+			case resultCh <- batchResult{
+				batchIndex: task.batchIndex,
+				embeddings: nil,
+				paragraphs: nil,
+				filename:   task.filename,
+			}:
+			case <-ctx.Done():
+				r.logger.Debug("embedding worker cancelled while sending empty batch", "worker", workerID)
+				return
+			}
+			r.logger.Debug("worker sent empty batch", "worker", workerID, "batch", task.batchIndex)
+			continue
+		}
+		// Embed with retry for API embedder
+		embeddings, err := r.embedWithRetry(ctx, task.paragraphs, 3)
+		if err != nil {
+			// Try to send error, but don't block indefinitely
+			select {
+			case errorCh <- fmt.Errorf("worker %d batch %d: %w", workerID, task.batchIndex, err):
+			case <-ctx.Done():
+				r.logger.Debug("embedding worker cancelled while sending error", "worker", workerID)
+			}
+			return
+		}
+		// Send result with context awareness
+		select {
+		case resultCh <- batchResult{
+			batchIndex: task.batchIndex,
+			embeddings: embeddings,
+			paragraphs: task.paragraphs,
+			filename:   task.filename,
+		}:
+		case <-ctx.Done():
+			r.logger.Debug("embedding worker cancelled while sending result", "worker", workerID)
+			return
+		}
+		r.logger.Debug("worker completed batch", "worker", workerID, "batch", task.batchIndex, "embeddings", len(embeddings))
+	}
+	r.logger.Debug("embedding worker finished", "worker", workerID)
+}
+
+// embedWithRetry attempts embedding with exponential backoff for API embedder
+func (r *RAG) embedWithRetry(ctx context.Context, paragraphs []string, maxRetries int) ([][]float32, error) {
+	var lastErr error
+	for attempt := 0; attempt < maxRetries; attempt++ {
+		if attempt > 0 {
+			// Exponential backoff
+			backoff := time.Duration(attempt*attempt) * time.Second
+			if backoff > 10*time.Second {
+				backoff = 10 * time.Second
+			}
+			select {
+			case <-time.After(backoff):
+			case <-ctx.Done():
+				return nil, ctx.Err()
+			}
+			r.logger.Debug("retrying embedding", "attempt", attempt, "max_retries", maxRetries)
+		}
+
+		embeddings, err := r.embedder.EmbedSlice(paragraphs)
+		if err == nil {
+			// Validate embedding count
+			if len(embeddings) != len(paragraphs) {
+				return nil, fmt.Errorf("embedding count mismatch: expected %d, got %d", len(paragraphs), len(embeddings))
+			}
+			return embeddings, nil
+		}
+
+		lastErr = err
+		// Only retry for API embedder errors (network/timeout)
+		// For ONNX embedder, fail fast
+		if _, isAPI := r.embedder.(*APIEmbedder); !isAPI {
+			break
+		}
+	}
+	return nil, fmt.Errorf("embedding failed after %d attempts: %w", maxRetries, lastErr)
+}
+
+// writeBatchToStorage writes a single batch of vectors to the database
+func (r *RAG) writeBatchToStorage(ctx context.Context, result batchResult, filename string) error {
+	if len(result.embeddings) == 0 {
+		// Empty batch, skip
+		return nil
+	}
+	// Check context before starting
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	default:
+	}
+
+	// Build all vectors for batch write
+	vectors := make([]*models.VectorRow, 0, len(result.paragraphs))
+	for j, text := range result.paragraphs {
+		vectors = append(vectors, &models.VectorRow{
+			Embeddings: result.embeddings[j],
 			RawText:    text,
-				Slug:       fmt.Sprintf("%s_%d_%d", filename, batchCount, j),
+			Slug:       fmt.Sprintf("%s_%d_%d", filename, result.batchIndex+1, j),
 			FileName:   filename,
+		})
 	}
-			if err := r.storage.WriteVector(&vector); err != nil {
-				r.logger.Error("failed to write vector to DB", "error", err, "slug", vector.Slug)
-				select {
-				case LongJobStatusCh <- ErrRAGStatus:
-				default:
-					r.logger.Warn("LongJobStatusCh channel full, dropping message")
-				}
-				return fmt.Errorf("failed to write vector: %w", err)
-			}
-		}
-		r.logger.Debug("wrote batch to db", "batch", batchCount, "size", len(nonEmptyBatch))
-		// Send progress status
-		statusMsg := fmt.Sprintf("processed batch %d/%d", batchCount, (len(paragraphs)+r.cfg.RAGBatchSize-1)/r.cfg.RAGBatchSize)
-		select {
-		case LongJobStatusCh <- statusMsg:
-		default:
-			r.logger.Warn("LongJobStatusCh channel full, dropping message")
-		}
-	}
-	r.logger.Debug("finished writing vectors", "batches", batchCount)
-	select {
-	case LongJobStatusCh <- FinishedRAGStatus:
-	default:
-		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", FinishedRAGStatus)
+
+	// Write all vectors in a single transaction
+	if err := r.storage.WriteVectors(vectors); err != nil {
+		r.logger.Error("failed to write vectors batch to DB", "error", err, "batch", result.batchIndex+1, "size", len(vectors))
+		r.sendStatusNonBlocking(ErrRAGStatus)
+		return fmt.Errorf("failed to write vectors batch: %w", err)
 	}
+	r.logger.Debug("wrote batch to db", "batch", result.batchIndex+1, "size", len(result.paragraphs))
 	return nil
 }

 func (r *RAG) LineToVector(line string) ([]float32, error) {
+	r.resetIdleTimer()
 	return r.embedder.Embed(line)
 }

-func (r *RAG) SearchEmb(emb *models.EmbeddingResp) ([]models.VectorRow, error) {
-	return r.storage.SearchClosest(emb.Embedding)
+func (r *RAG) searchEmb(emb *models.EmbeddingResp, limit int) ([]models.VectorRow, error) {
+	r.resetIdleTimer()
+	return r.storage.SearchClosest(emb.Embedding, limit)
+}
+
+func (r *RAG) searchKeyword(query string, limit int) ([]models.VectorRow, error) {
+	r.resetIdleTimer()
+	sanitized := sanitizeFTSQuery(query)
+	return r.storage.SearchKeyword(sanitized, limit)
 }

 func (r *RAG) ListLoaded() ([]string, error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
 	return r.storage.ListFiles()
 }

 func (r *RAG) RemoveFile(filename string) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.resetIdleTimer()
 	return r.storage.RemoveEmbByFileName(filename)
 }

@@ -211,11 +562,14 @@ func (r *RAG) RefineQuery(query string) string {
 		return original
 	}
 	query = strings.ToLower(query)
+	words := strings.Fields(query)
+	if len(words) >= 3 {
 		for _, stopWord := range stopWords {
 			wordPattern := `\b` + stopWord + `\b`
 			re := regexp.MustCompile(wordPattern)
 			query = re.ReplaceAllString(query, "")
 		}
+	}
 	query = strings.TrimSpace(query)
 	if len(query) < 5 {
 		return original
@@ -246,7 +600,7 @@ func (r *RAG) extractImportantPhrases(query string) string {
 				break
 			}
 		}
-		if isImportant || len(word) > 3 {
+		if isImportant || len(word) >= 3 {
 			important = append(important, word)
 		}
 	}
@@ -265,6 +619,36 @@ func (r *RAG) GenerateQueryVariations(query string) []string {
 	if len(parts) == 0 {
 		return variations
 	}
+	// Get loaded filenames to filter out filename terms
+	filenames, err := r.storage.ListFiles()
+	if err == nil && len(filenames) > 0 {
+		// Convert to lowercase for case-insensitive matching
+		lowerFilenames := make([]string, len(filenames))
+		for i, f := range filenames {
+			lowerFilenames[i] = strings.ToLower(f)
+		}
+		filteredParts := make([]string, 0, len(parts))
+		for _, part := range parts {
+			partLower := strings.ToLower(part)
+			skip := false
+			for _, fn := range lowerFilenames {
+				if strings.Contains(fn, partLower) || strings.Contains(partLower, fn) {
+					skip = true
+					break
+				}
+			}
+			if !skip {
+				filteredParts = append(filteredParts, part)
+			}
+		}
+		// If filteredParts not empty and different from original, add filtered query
+		if len(filteredParts) > 0 && len(filteredParts) != len(parts) {
+			filteredQuery := strings.Join(filteredParts, " ")
+			if len(filteredQuery) >= 5 {
+				variations = append(variations, filteredQuery)
+			}
+		}
+	}
 	if len(parts) >= 2 {
 		trimmed := strings.Join(parts[:len(parts)-1], " ")
 		if len(trimmed) >= 5 {
@@ -328,9 +712,14 @@ func (r *RAG) RerankResults(results []models.VectorRow, query string) []models.V
 	})
 	unique := make([]models.VectorRow, 0)
 	seen := make(map[string]bool)
+	fileCounts := make(map[string]int)
 	for i := range scored {
 		if !seen[scored[i].row.Slug] {
+			if fileCounts[scored[i].row.FileName] >= 2 {
+				continue
+			}
 			seen[scored[i].row.Slug] = true
+			fileCounts[scored[i].row.FileName]++
 			unique = append(unique, scored[i].row)
 		}
 	}
@@ -341,6 +730,9 @@ func (r *RAG) RerankResults(results []models.VectorRow, query string) []models.V
 }

 func (r *RAG) SynthesizeAnswer(results []models.VectorRow, query string) (string, error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	r.resetIdleTimer()
 	if len(results) == 0 {
 		return "No relevant information found in the vector database.", nil
 	}
@@ -369,7 +761,7 @@ func (r *RAG) SynthesizeAnswer(results []models.VectorRow, query string) (string
 		Embedding: emb,
 		Index:     0,
 	}
-	topResults, err := r.SearchEmb(embResp)
+	topResults, err := r.searchEmb(embResp, 1)
 	if err != nil {
 		r.logger.Error("failed to search for synthesis context", "error", err)
 		return "", err
@@ -396,9 +788,14 @@ func truncateString(s string, maxLen int) string {
 }

 func (r *RAG) Search(query string, limit int) ([]models.VectorRow, error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	r.resetIdleTimer()
 	refined := r.RefineQuery(query)
 	variations := r.GenerateQueryVariations(refined)
-	allResults := make([]models.VectorRow, 0)
+
+	// Collect embedding search results from all variations
+	var embResults []models.VectorRow
 	seen := make(map[string]bool)
 	for _, q := range variations {
 		emb, err := r.LineToVector(q)
@@ -406,29 +803,78 @@ func (r *RAG) Search(query string, limit int) ([]models.VectorRow, error) {
 			r.logger.Error("failed to embed query variation", "error", err, "query", q)
 			continue
 		}
-
 		embResp := &models.EmbeddingResp{
 			Embedding: emb,
 			Index:     0,
 		}
-
-		results, err := r.SearchEmb(embResp)
+		results, err := r.searchEmb(embResp, limit*2) // Get more candidates
 		if err != nil {
 			r.logger.Error("failed to search embeddings", "error", err, "query", q)
 			continue
 		}
-
 		for _, row := range results {
 			if !seen[row.Slug] {
 				seen[row.Slug] = true
-				allResults = append(allResults, row)
+				embResults = append(embResults, row)
 			}
 		}
 	}
-	reranked := r.RerankResults(allResults, query)
-	if len(reranked) > limit {
-		reranked = reranked[:limit]
+	// Sort embedding results by distance (lower is better)
+	sort.Slice(embResults, func(i, j int) bool {
+		return embResults[i].Distance < embResults[j].Distance
+	})
+
+	// Perform keyword search
+	kwResults, err := r.searchKeyword(refined, limit*2)
+	if err != nil {
+		r.logger.Warn("keyword search failed, using only embeddings", "error", err)
+		kwResults = nil
 	}
+	// Sort keyword results by distance (already sorted by BM25 score)
+	// kwResults already sorted by distance (lower is better)
+
+	// Combine using Reciprocal Rank Fusion (RRF)
+	const rrfK = 60
+	type scoredRow struct {
+		row   models.VectorRow
+		score float64
+	}
+	scoreMap := make(map[string]float64)
+	// Add embedding results
+	for rank, row := range embResults {
+		score := 1.0 / (float64(rank) + rrfK)
+		scoreMap[row.Slug] += score
+	}
+	// Add keyword results
+	for rank, row := range kwResults {
+		score := 1.0 / (float64(rank) + rrfK)
+		scoreMap[row.Slug] += score
+		// Ensure row exists in combined results
+		if _, exists := seen[row.Slug]; !exists {
+			embResults = append(embResults, row)
+		}
+	}
+	// Create slice of scored rows
+	scoredRows := make([]scoredRow, 0, len(embResults))
+	for _, row := range embResults {
+		score := scoreMap[row.Slug]
+		scoredRows = append(scoredRows, scoredRow{row: row, score: score})
+	}
+	// Sort by descending RRF score
+	sort.Slice(scoredRows, func(i, j int) bool {
+		return scoredRows[i].score > scoredRows[j].score
+	})
+	// Take top limit
+	if len(scoredRows) > limit {
+		scoredRows = scoredRows[:limit]
+	}
+	// Convert back to VectorRow
+	finalResults := make([]models.VectorRow, len(scoredRows))
+	for i, sr := range scoredRows {
+		finalResults[i] = sr.row
+	}
+	// Apply reranking heuristics
+	reranked := r.RerankResults(finalResults, query)
 	return reranked, nil
 }

@@ -437,16 +883,58 @@ var (
 	ragOnce     sync.Once
 )

+func (r *RAG) FallbackMessage() string {
+	return r.fallbackMsg
+}
+
 func Init(c *config.Config, l *slog.Logger, s storage.FullRepo) error {
+	var err error
 	ragOnce.Do(func() {
 		if c == nil || l == nil || s == nil {
 			return
 		}
-		ragInstance = New(l, s, c)
+		ragInstance, err = New(l, s, c)
 	})
-	return nil
+	return err
 }

 func GetInstance() *RAG {
 	return ragInstance
 }
+
+func (r *RAG) resetIdleTimer() {
+	r.idleMu.Lock()
+	defer r.idleMu.Unlock()
+	if r.idleTimer != nil {
+		r.idleTimer.Stop()
+	}
+	r.idleTimer = time.AfterFunc(r.idleTimeout, func() {
+		r.freeONNXMemory()
+	})
+}
+
+func (r *RAG) freeONNXMemory() {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if onnx, ok := r.embedder.(*ONNXEmbedder); ok {
+		if err := onnx.Destroy(); err != nil {
+			r.logger.Error("failed to free ONNX memory", "error", err)
+		} else {
+			r.logger.Info("freed ONNX VRAM after idle timeout")
+		}
+	}
+}
+
+func (r *RAG) Destroy() {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if r.idleTimer != nil {
+		r.idleTimer.Stop()
+		r.idleTimer = nil
+	}
+	if onnx, ok := r.embedder.(*ONNXEmbedder); ok {
+		if err := onnx.Destroy(); err != nil {
+			r.logger.Error("failed to destroy ONNX embedder", "error", err)
+		}
+	}
+}
--- 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 +261,100 @@ func (vs *VectorStorage) SearchClosest(query []float32) ([]models.VectorRow, err
 	return results, nil
 }

+// GetVectorBySlug retrieves a vector row by its slug
+func (vs *VectorStorage) GetVectorBySlug(slug string) (*models.VectorRow, error) {
+	embeddingSizes := []int{384, 768, 1024, 1536, 2048, 3072, 4096, 5120}
+	for _, size := range embeddingSizes {
+		table := fmt.Sprintf("embeddings_%d", size)
+		query := fmt.Sprintf("SELECT embeddings, slug, raw_text, filename FROM %s WHERE slug = ?", table)
+		row := vs.sqlxDB.QueryRow(query, slug)
+		var (
+			embeddingsBlob                   []byte
+			retrievedSlug, rawText, fileName string
+		)
+		if err := row.Scan(&embeddingsBlob, &retrievedSlug, &rawText, &fileName); err != nil {
+			// No row in this table, continue to next size
+			continue
+		}
+		storedEmbeddings := DeserializeVector(embeddingsBlob)
+		return &models.VectorRow{
+			Embeddings: storedEmbeddings,
+			Slug:       retrievedSlug,
+			RawText:    rawText,
+			FileName:   fileName,
+		}, nil
+	}
+	return nil, fmt.Errorf("vector with slug %s not found", slug)
+}
+
+// SearchKeyword performs full-text search using FTS5
+func (vs *VectorStorage) SearchKeyword(query string, limit int) ([]models.VectorRow, error) {
+	// Use FTS5 bm25 ranking. bm25 returns negative values where more negative is better.
+	// We'll order by bm25 (ascending) and limit.
+	ftsQuery := `SELECT slug, raw_text, filename, bm25(fts_embeddings) as score 
+				 FROM fts_embeddings 
+				 WHERE fts_embeddings MATCH ? 
+				 ORDER BY score 
+				 LIMIT ?`
+
+	// Try original query first
+	rows, err := vs.sqlxDB.Query(ftsQuery, query, limit)
+	if err != nil {
+		return nil, fmt.Errorf("FTS search failed: %w", err)
+	}
+	results, err := vs.scanRows(rows)
+	rows.Close()
+	if err != nil {
+		return nil, err
+	}
+
+	// If no results and query contains multiple terms, try OR fallback
+	if len(results) == 0 && strings.Contains(query, " ") && !strings.Contains(strings.ToUpper(query), " OR ") {
+		// Build OR query: term1 OR term2 OR term3
+		terms := strings.Fields(query)
+		if len(terms) > 1 {
+			orQuery := strings.Join(terms, " OR ")
+			rows, err := vs.sqlxDB.Query(ftsQuery, orQuery, limit)
+			if err != nil {
+				// Return original empty results rather than error
+				return results, nil
+			}
+			orResults, err := vs.scanRows(rows)
+			rows.Close()
+			if err == nil {
+				results = orResults
+			}
+		}
+	}
+	return results, nil
+}
+
+// scanRows converts SQL rows to VectorRow slice
+func (vs *VectorStorage) scanRows(rows *sql.Rows) ([]models.VectorRow, error) {
+	var results []models.VectorRow
+	for rows.Next() {
+		var slug, rawText, fileName string
+		var score float64
+		if err := rows.Scan(&slug, &rawText, &fileName, &score); err != nil {
+			vs.logger.Error("failed to scan FTS row", "error", err)
+			continue
+		}
+		// Convert BM25 score to distance-like metric (lower is better)
+		// BM25 is negative, more negative is better. We'll normalize to positive distance.
+		distance := float32(-score) // Make positive (since score is negative)
+		if distance < 0 {
+			distance = 0
+		}
+		results = append(results, models.VectorRow{
+			Slug:     slug,
+			RawText:  rawText,
+			FileName: fileName,
+			Distance: distance,
+		})
+	}
+	return results, nil
+}
+
 // ListFiles returns a list of all loaded files
 func (vs *VectorStorage) ListFiles() ([]string, error) {
 	fileLists := make([][]string, 0)
@@ -215,6 +399,10 @@ func (vs *VectorStorage) ListFiles() ([]string, error) {
 // RemoveEmbByFileName removes all embeddings associated with a specific filename
 func (vs *VectorStorage) RemoveEmbByFileName(filename string) error {
 	var errors []string
+	// Delete from FTS table first
+	if _, err := vs.sqlxDB.Exec("DELETE FROM fts_embeddings WHERE filename = ?", filename); err != nil {
+		errors = append(errors, err.Error())
+	}
 	embeddingSizes := []int{384, 768, 1024, 1536, 2048, 3072, 4096, 5120}
 	for _, size := range embeddingSizes {
 		table := fmt.Sprintf("embeddings_%d", size)
--- 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]
+		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]
 	}
-			}
-			topResults = topResults[:3]
-		}
-	}
-
 	// 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/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
@@ -36,6 +38,7 @@ var (
 	roleEditWindow     *tview.InputField
 	shellInput         *tview.InputField
 	confirmModal       *tview.Modal
+	toastTimer         *time.Timer
 	confirmPageName    = "confirm"
 	fullscreenMode     bool
 	positionVisible    bool = true
@@ -137,8 +140,8 @@ func setShellMode(enabled bool) {
 	}()
 }

-// showToast displays a temporary message in the top‑right corner.
-// 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 {
@@ -154,6 +157,11 @@ func showToast(title, message string) {
 	}
 	title = sanitize(title, 50)
 	message = sanitize(message, 197)
+	if toastTimer != nil {
+		toastTimer.Stop()
+	}
+	// show blocking notification to not mess up flex
+	if fullscreenMode {
 		notification := tview.NewTextView().
 			SetTextAlign(tview.AlignCenter).
 			SetDynamicColors(true).
@@ -176,14 +184,44 @@ func showToast(title, message string) {
 		// Generate a unique page name (e.g., using timestamp) to allow multiple toasts.
 		pageName := fmt.Sprintf("toast-%d", time.Now().UnixNano())
 		pages.AddPage(pageName, background, true, true)
-	// Auto‑dismiss after 3 seconds.
-	time.AfterFunc(3*time.Second, func() {
+		// 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)
 				}
 			})
 		})
+		return
+	}
+	notificationWidget.SetTitle(title)
+	notificationWidget.SetText(fmt.Sprintf("[yellow]%s[-]", message))
+	go func() {
+		app.QueueUpdateDraw(func() {
+			flex.RemoveItem(bottomFlex)
+			flex.RemoveItem(statusLineWidget)
+			bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+				AddItem(textArea, 0, 1, true).
+				AddItem(notificationWidget, 40, 1, false)
+			flex.AddItem(bottomFlex, 0, 10, true)
+			if positionVisible {
+				flex.AddItem(statusLineWidget, 0, 2, false)
+			}
+		})
+	}()
+	toastTimer = time.AfterFunc(3*time.Second, func() {
+		app.QueueUpdateDraw(func() {
+			flex.RemoveItem(bottomFlex)
+			flex.RemoveItem(statusLineWidget)
+			bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+				AddItem(textArea, 0, 1, true).
+				AddItem(notificationWidget, 0, 0, false)
+			flex.AddItem(bottomFlex, 0, 10, true)
+			if positionVisible {
+				flex.AddItem(statusLineWidget, 0, 2, false)
+			}
+		})
+	})
 }

 func init() {
@@ -235,7 +273,7 @@ func init() {
 			shellHistoryPos = -1
 		}
 		// Handle Tab key for @ file completion
-		if event.Key() == tcell.KeyTab {
+		if event.Key() == tcell.KeyTab && shellMode {
 			currentText := shellInput.GetText()
 			atIndex := strings.LastIndex(currentText, "@")
 			if atIndex >= 0 {
@@ -286,12 +324,26 @@ func init() {
 		SetDynamicColors(true).
 		SetRegions(true).
 		SetChangedFunc(func() {
+			// INFO:
+			// https://github.com/rivo/tview/wiki/Concurrency#event-handlers
+			// although already called by default per tview specs
+			// calling it explicitly makes text streaming to look more smooth
 			app.Draw()
 		})
+	notificationWidget = tview.NewTextView().
+		SetTextAlign(tview.AlignCenter).
+		SetDynamicColors(true).
+		SetRegions(true).
+		SetChangedFunc(func() {
+		})
+	notificationWidget.SetBorder(true).SetTitle("notification")
+	bottomFlex = tview.NewFlex().SetDirection(tview.FlexColumn).
+		AddItem(textArea, 0, 1, true).
+		AddItem(notificationWidget, 0, 0, false)
 	//
 	flex = tview.NewFlex().SetDirection(tview.FlexRow).
 		AddItem(textView, 0, 40, false).
-		AddItem(textArea, 0, 10, true) // Restore original height
+		AddItem(bottomFlex, 0, 10, true)
 	if positionVisible {
 		flex.AddItem(statusLineWidget, 0, 2, false)
 	}
@@ -360,10 +412,14 @@ func init() {
 	// 	y += h / 2
 	// 	return x, y, w, h
 	// })
+	notificationWidget.SetDrawFunc(func(screen tcell.Screen, x, y, w, h int) (int, int, int, int) {
+		y += h / 2
+		return x, y, w, h
+	})
 	// Initially set up flex without search bar
 	flex = tview.NewFlex().SetDirection(tview.FlexRow).
 		AddItem(textView, 0, 40, false).
-		AddItem(textArea, 0, 10, true) // Restore original height
+		AddItem(bottomFlex, 0, 10, true)
 	if positionVisible {
 		flex.AddItem(statusLineWidget, 0, 2, false)
 	}
@@ -1095,7 +1151,7 @@ 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()
Author	SHA1	Message	Date
Grail Finder	2c495253c2	Chore: panic is not solved (WIP)	2026-03-07 09:40:58 +03:00
Grail Finder	118a0a0d55	Chore: cleanup logs	2026-03-07 09:08:01 +03:00
Grail Finder	44633d64c6	Chore: move to own file	2026-03-07 08:52:10 +03:00
Grail Finder	0598e3e86d	Feat: kokoro onnx (WIP)	2026-03-07 08:35:44 +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