Fix (rag): epub load

move cleaning image attachment to the end of chatRound
fmt cleanup

Makefile

@@ -1,4 +1,4 @@
-.PHONY: setconfig run lint setup-whisper build-whisper download-whisper-model docker-up docker-down docker-logs noextra-run
+.PHONY: setconfig run lint setup-whisper build-whisper download-whisper-model docker-up docker-down docker-logs noextra-run installdelve checkdelve
 
 run: setconfig
 	go build -tags extra -o gf-lt && ./gf-lt
@@ -15,6 +15,12 @@ noextra-run: setconfig
 setconfig:
 	find config.toml &>/dev/null || cp config.example.toml config.toml
 
+installdelve:
+	go install github.com/go-delve/delve/cmd/dlv@latest
+
+checkdelve:
+	which dlv &>/dev/null || installdelve
+
 lint: ## Run linters. Use make install-linters first.
 	golangci-lint run -c .golangci.yml ./...
 

README.md

@@ -8,7 +8,7 @@ made with use of [tview](https://github.com/rivo/tview)
 - tts/stt (run make commands to get deps);
 - image input;
 - function calls (function calls are implemented natively, to avoid calling outside sources);
-- [character specific context (unique feature)](char-specific-context.md)
+- [character specific context (unique feature)](docs/char-specific-context.md)
 
 #### how it looks
 ![how it looks](assets/ex01.png)

bot.go

@@ -46,6 +46,7 @@ var (
 	ragger          *rag.RAG
 	chunkParser     ChunkParser
 	lastToolCall    *models.FuncCall
+	lastRespStats   *models.ResponseStats
 	//nolint:unused // TTS_ENABLED conditionally uses this
 	orator          Orator
 	asr             STT
@@ -410,14 +411,21 @@ func fetchLCPModelsWithLoadStatus() ([]string, error) {
 		return nil, err
 	}
 	result := make([]string, 0, len(models.Data))
-	for _, m := range models.Data {
+	li := 0 // loaded index
+	for i, m := range models.Data {
 		modelName := m.ID
 		if m.Status.Value == "loaded" {
-			modelName = modelName + " (loaded)"
+			modelName = "(loaded) " + modelName
+			li = i
 		}
 		result = append(result, modelName)
 	}
-	return result, nil
+	if li == 0 {
+		return result, nil // no loaded models
+	}
+	loadedModel := result[li]
+	result = append(result[:li], result[li+1:]...)
+	return slices.Concat([]string{loadedModel}, result), nil
 }
 
 // fetchLCPModelsWithStatus returns the full LCPModels struct including status information.
@@ -484,30 +492,28 @@ func monitorModelLoad(modelID string) {
 // extractDetailedErrorFromBytes extracts detailed error information from response body bytes
 func extractDetailedErrorFromBytes(body []byte, statusCode int) string {
 	// Try to parse as JSON to extract detailed error information
-	var errorResponse map[string]interface{}
+	var errorResponse map[string]any
 	if err := json.Unmarshal(body, &errorResponse); err == nil {
 		// Check if it's an error response with detailed information
 		if errorData, ok := errorResponse["error"]; ok {
-			if errorMap, ok := errorData.(map[string]interface{}); ok {
+			if errorMap, ok := errorData.(map[string]any); ok {
 				var errorMsg string
 				if msg, ok := errorMap["message"]; ok {
 					errorMsg = fmt.Sprintf("%v", msg)
 				}
-
 				var details []string
 				if code, ok := errorMap["code"]; ok {
 					details = append(details, fmt.Sprintf("Code: %v", code))
 				}
-
 				if metadata, ok := errorMap["metadata"]; ok {
 					// Handle metadata which might contain raw error details
-					if metadataMap, ok := metadata.(map[string]interface{}); ok {
+					if metadataMap, ok := metadata.(map[string]any); ok {
 						if raw, ok := metadataMap["raw"]; ok {
 							// Parse the raw error string if it's JSON
-							var rawError map[string]interface{}
+							var rawError map[string]any
 							if rawStr, ok := raw.(string); ok && json.Unmarshal([]byte(rawStr), &rawError) == nil {
 								if rawErrorData, ok := rawError["error"]; ok {
-									if rawErrorMap, ok := rawErrorData.(map[string]interface{}); ok {
+									if rawErrorMap, ok := rawErrorData.(map[string]any); ok {
 										if rawMsg, ok := rawErrorMap["message"]; ok {
 											return fmt.Sprintf("API Error: %s", rawMsg)
 										}
@@ -518,20 +524,30 @@ func extractDetailedErrorFromBytes(body []byte, statusCode int) string {
 					}
 					details = append(details, fmt.Sprintf("Metadata: %v", metadata))
 				}
-
 				if len(details) > 0 {
 					return fmt.Sprintf("API Error: %s (%s)", errorMsg, strings.Join(details, ", "))
 				}
-
 				return "API Error: " + errorMsg
 			}
 		}
 	}
-
 	// If not a structured error response, return the raw body with status
 	return fmt.Sprintf("HTTP Status: %d, Response Body: %s", statusCode, string(body))
 }
 
+func finalizeRespStats(tokenCount int, startTime time.Time) {
+	duration := time.Since(startTime).Seconds()
+	var tps float64
+	if duration > 0 {
+		tps = float64(tokenCount) / duration
+	}
+	lastRespStats = &models.ResponseStats{
+		Tokens:       tokenCount,
+		Duration:     duration,
+		TokensPerSec: tps,
+	}
+}
+
 // sendMsgToLLM expects streaming resp
 func sendMsgToLLM(body io.Reader) {
 	choseChunkParser()
@@ -560,7 +576,6 @@ func sendMsgToLLM(body io.Reader) {
 		streamDone <- true
 		return
 	}
-
 	// Check if the initial response is an error before starting to stream
 	if resp.StatusCode >= 400 {
 		// Read the response body to get detailed error information
@@ -575,7 +590,6 @@ func sendMsgToLLM(body io.Reader) {
 			streamDone <- true
 			return
 		}
-
 		// Parse the error response for detailed information
 		detailedError := extractDetailedErrorFromBytes(bodyBytes, resp.StatusCode)
 		logger.Error("API returned error status", "status_code", resp.StatusCode, "detailed_error", detailedError)
@@ -586,12 +600,17 @@ func sendMsgToLLM(body io.Reader) {
 		streamDone <- true
 		return
 	}
-
+	//
 	defer resp.Body.Close()
 	reader := bufio.NewReader(resp.Body)
 	counter := uint32(0)
+	tokenCount := 0
+	startTime := time.Now()
 	hasReasoning := false
 	reasoningSent := false
+	defer func() {
+		finalizeRespStats(tokenCount, startTime)
+	}()
 	for {
 		var (
 			answerText string
@@ -667,11 +686,13 @@ func sendMsgToLLM(body io.Reader) {
 			// Close the thinking block if we were streaming reasoning and haven't closed it yet
 			if hasReasoning && !reasoningSent {
 				chunkChan <- "</think>"
+				tokenCount++
 			}
 			if chunk.Chunk != "" {
 				logger.Warn("text inside of finish llmchunk", "chunk", chunk, "counter", counter)
 				answerText = strings.ReplaceAll(chunk.Chunk, "\n\n", "\n")
 				chunkChan <- answerText
+				tokenCount++
 			}
 			streamDone <- true
 			break
@@ -684,22 +705,23 @@ func sendMsgToLLM(body io.Reader) {
 			if !hasReasoning {
 				// First reasoning chunk - send opening tag
 				chunkChan <- "<think>"
+				tokenCount++
 				hasReasoning = true
 			}
 			// Stream reasoning content immediately
 			answerText = strings.ReplaceAll(chunk.Reasoning, "\n\n", "\n")
 			if answerText != "" {
 				chunkChan <- answerText
+				tokenCount++
 			}
 		}
-
 		// When we get content and have been streaming reasoning, close the thinking block
 		if chunk.Chunk != "" && hasReasoning && !reasoningSent {
 			// Close the thinking block before sending actual content
 			chunkChan <- "</think>"
+			tokenCount++
 			reasoningSent = true
 		}
-
 		// bot sends way too many \n
 		answerText = strings.ReplaceAll(chunk.Chunk, "\n\n", "\n")
 		// Accumulate text to check for stop strings that might span across chunks
@@ -709,9 +731,11 @@ func sendMsgToLLM(body io.Reader) {
 			slices.Contains(stopStrings, answerText) {
 			logger.Debug("stop string detected on client side for completion endpoint", "stop_string", answerText)
 			streamDone <- true
+			break
 		}
 		if answerText != "" {
 			chunkChan <- answerText
+			tokenCount++
 		}
 		openAIToolChan <- chunk.ToolChunk
 		if chunk.FuncName != "" {
@@ -743,12 +767,10 @@ func chatRagUse(qText string) (string, error) {
 		questions[i] = q.Text
 		logger.Debug("RAG question extracted", "index", i, "question", q.Text)
 	}
-
 	if len(questions) == 0 {
 		logger.Warn("No questions extracted from query text", "query", qText)
 		return "No related results from RAG vector storage.", nil
 	}
-
 	respVecs := []models.VectorRow{}
 	for i, q := range questions {
 		logger.Debug("Processing RAG question", "index", i, "question", q)
@@ -758,7 +780,6 @@ func chatRagUse(qText string) (string, error) {
 			continue
 		}
 		logger.Debug("Got embeddings for question", "index", i, "question_len", len(q), "embedding_len", len(emb))
-
 		// Create EmbeddingResp struct for the search
 		embeddingResp := &models.EmbeddingResp{
 			Embedding: emb,
@@ -772,7 +793,6 @@ func chatRagUse(qText string) (string, error) {
 		logger.Debug("RAG search returned vectors", "index", i, "question", q, "vector_count", len(vecs))
 		respVecs = append(respVecs, vecs...)
 	}
-
 	// get raw text
 	resps := []string{}
 	logger.Debug("RAG query final results", "total_vecs_found", len(respVecs))
@@ -780,12 +800,10 @@ func chatRagUse(qText string) (string, error) {
 		resps = append(resps, rv.RawText)
 		logger.Debug("RAG result", "slug", rv.Slug, "filename", rv.FileName, "raw_text_len", len(rv.RawText))
 	}
-
 	if len(resps) == 0 {
 		logger.Info("No RAG results found for query", "original_query", qText, "question_count", len(questions))
 		return "No related results from RAG vector storage.", nil
 	}
-
 	result := strings.Join(resps, "\n")
 	logger.Debug("RAG query completed", "result_len", len(result), "response_count", len(resps))
 	return result, nil
@@ -808,14 +826,45 @@ func chatWatcher(ctx context.Context) {
 	}
 }
 
+// inpired by https://github.com/rivo/tview/issues/225
+func showSpinner() {
+	spinners := []string{"⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"}
+	var i int
+	botPersona := cfg.AssistantRole
+	if cfg.WriteNextMsgAsCompletionAgent != "" {
+		botPersona = cfg.WriteNextMsgAsCompletionAgent
+	}
+	for botRespMode || toolRunningMode {
+		time.Sleep(100 * time.Millisecond)
+		spin := i % len(spinners)
+		app.QueueUpdateDraw(func() {
+			if toolRunningMode {
+				textArea.SetTitle(spinners[spin] + " tool")
+			} else if botRespMode {
+				textArea.SetTitle(spinners[spin] + " " + botPersona)
+			} else {
+				textArea.SetTitle(spinners[spin] + " input")
+			}
+		})
+		i++
+	}
+	app.QueueUpdateDraw(func() {
+		textArea.SetTitle("input")
+	})
+}
+
 func chatRound(r *models.ChatRoundReq) error {
 	botRespMode = true
+	go showSpinner()
 	updateStatusLine()
 	botPersona := cfg.AssistantRole
 	if cfg.WriteNextMsgAsCompletionAgent != "" {
 		botPersona = cfg.WriteNextMsgAsCompletionAgent
 	}
-	defer func() { botRespMode = false }()
+	defer func() {
+		botRespMode = false
+		ClearImageAttachment()
+	}()
 	// check that there is a model set to use if is not local
 	choseChunkParser()
 	reader, err := chunkParser.FormMsg(r.UserMsg, r.Role, r.Resume)
@@ -834,13 +883,14 @@ func chatRound(r *models.ChatRoundReq) error {
 		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{
 			Role: botPersona, Content: "",
 		})
-		fmt.Fprintf(textView, "\n[-:-:b](%d) ", msgIdx)
+		nl := "\n\n"
-		fmt.Fprint(textView, roleToIcon(botPersona))
+		prevText := textView.GetText(true)
-		fmt.Fprint(textView, "[-:-:-]\n")
+		if strings.HasSuffix(prevText, nl) {
-		if cfg.ThinkUse && !strings.Contains(cfg.CurrentAPI, "v1") {
+			nl = ""
-			// fmt.Fprint(textView, "<think>")
+		} else if strings.HasSuffix(prevText, "\n") {
-			chunkChan <- "<think>"
+			nl = "\n"
 		}
+		fmt.Fprintf(textView, "%s[-:-:b](%d) %s[-:-:-]\n", nl, msgIdx, roleToIcon(botPersona))
 	} else {
 		msgIdx = len(chatBody.Messages) - 1
 	}
@@ -914,7 +964,6 @@ out:
 				textView.ScrollToEnd()
 			}
 		case <-streamDone:
-			// drain any remaining chunks from chunkChan before exiting
 			for len(chunkChan) > 0 {
 				chunk := <-chunkChan
 				fmt.Fprint(textView, chunk)
@@ -923,31 +972,40 @@ out:
 					textView.ScrollToEnd()
 				}
 				if cfg.TTS_ENABLED {
-					// Send chunk to audio stream handler
 					TTSTextChan <- chunk
 				}
 			}
 			if cfg.TTS_ENABLED {
-				// msg is done; flush it down
 				TTSFlushChan <- true
 			}
 			break out
 		}
 	}
+	var msgStats *models.ResponseStats
+	if lastRespStats != nil {
+		msgStats = &models.ResponseStats{
+			Tokens:       lastRespStats.Tokens,
+			Duration:     lastRespStats.Duration,
+			TokensPerSec: lastRespStats.TokensPerSec,
+		}
+		lastRespStats = nil
+	}
 	botRespMode = false
-	// numbers in chatbody and displayed must be the same
 	if r.Resume {
 		chatBody.Messages[len(chatBody.Messages)-1].Content += respText.String()
-		// lastM.Content = lastM.Content + respText.String()
-		// Process the updated message to check for known_to tags in resumed response
 		updatedMsg := chatBody.Messages[len(chatBody.Messages)-1]
 		processedMsg := processMessageTag(&updatedMsg)
 		chatBody.Messages[len(chatBody.Messages)-1] = *processedMsg
+		if msgStats != nil && chatBody.Messages[len(chatBody.Messages)-1].Role != cfg.ToolRole {
+			chatBody.Messages[len(chatBody.Messages)-1].Stats = msgStats
+		}
 	} else {
-		// Message was already added at the start, just process it for known_to tags
 		chatBody.Messages[msgIdx].Content = respText.String()
 		processedMsg := processMessageTag(&chatBody.Messages[msgIdx])
 		chatBody.Messages[msgIdx] = *processedMsg
+		if msgStats != nil && chatBody.Messages[msgIdx].Role != cfg.ToolRole {
+			chatBody.Messages[msgIdx].Stats = msgStats
+		}
 		stopTTSIfNotForUser(&chatBody.Messages[msgIdx])
 	}
 	cleanChatBody()
@@ -1169,7 +1227,11 @@ func findCall(msg, toolCall string) bool {
 		chatRoundChan <- crr
 		return true
 	}
+	// Show tool call progress indicator before execution
+	fmt.Fprintf(textView, "\n[yellow::i][tool: %s...][-:-:-]", fc.Name)
+	toolRunningMode = true
 	resp := callToolWithAgent(fc.Name, fc.Args)
+	toolRunningMode = false
 	toolMsg := string(resp) // Remove the "tool response: " prefix and %+v formatting
 	logger.Info("llm used a tool call", "tool_name", fc.Name, "too_args", fc.Args, "id", fc.ID, "tool_resp", toolMsg)
 	fmt.Fprintf(textView, "%s[-:-:b](%d) <%s>: [-:-:-]\n%s\n",
@@ -1208,7 +1270,6 @@ func chatToTextSlice(messages []models.RoleMsg, showSys bool) []string {
 func chatToText(messages []models.RoleMsg, showSys bool) string {
 	s := chatToTextSlice(messages, showSys)
 	text := strings.Join(s, "\n")
-
 	// Collapse thinking blocks if enabled
 	if thinkingCollapsed {
 		text = thinkRE.ReplaceAllStringFunc(text, func(match string) string {
@@ -1232,7 +1293,6 @@ func chatToText(messages []models.RoleMsg, showSys bool) string {
 			}
 		}
 	}
-
 	return text
 }
 
@@ -1369,7 +1429,7 @@ func init() {
 	var err error
 	cfg, err = config.LoadConfig("config.toml")
 	if err != nil {
-		fmt.Println("failed to load config.toml")
+		fmt.Println("failed to load config.toml", err)
 		cancel()
 		os.Exit(1)
 		return

config.example.toml

@@ -10,7 +10,10 @@ DeepSeekModel = "deepseek-reasoner"
 OpenRouterCompletionAPI = "https://openrouter.ai/api/v1/completions"
 OpenRouterChatAPI = "https://openrouter.ai/api/v1/chat/completions"
 # OpenRouterToken = ""
+# embeddings
 EmbedURL = "http://localhost:8082/v1/embeddings"
+HFToken = ""
+#
 ShowSys = true
 LogFile = "log.txt"
 UserRole = "user"
@@ -21,6 +24,7 @@ ChunkLimit = 100000
 AutoScrollEnabled = true
 AutoCleanToolCallsFromCtx = false
 # rag settings
+RAGEnabled = false
 RAGBatchSize = 1
 RAGWordLimit = 80
 RAGWorkers = 2
@@ -41,9 +45,8 @@ STT_LANG = "en"  # Language for speech recognition (for WHISPER_BINARY mode)
 STT_SR = 16000  # Sample rate for audio recording
 #
 DBPATH = "gflt.db"
-FilePickerDir = "." # Directory where file picker should start
+FilePickerDir = "." # Directory for file picker start and coding assistant file operations (relative paths resolved against this)
 FilePickerExts = "png,jpg,jpeg,gif,webp" # Comma-separated list of allowed file extensions for file picker
-CodingDir = "." # Default directory for coding assistant file operations (relative paths resolved against this)
 EnableMouse = false # Enable mouse support in the UI
 # character specific context
 CharSpecificContextEnabled = true

config/config.go

@@ -18,7 +18,6 @@ type Config struct {
 	UserRole                      string `toml:"UserRole"`
 	ToolRole                      string `toml:"ToolRole"`
 	ToolUse                       bool   `toml:"ToolUse"`
-	ThinkUse                      bool   `toml:"ThinkUse"`
 	StripThinkingFromAPI          bool   `toml:"StripThinkingFromAPI"`
 	ReasoningEffort               string `toml:"ReasoningEffort"`
 	AssistantRole                 string `toml:"AssistantRole"`
@@ -32,15 +31,14 @@ type Config struct {
 	DBPATH                        string `toml:"DBPATH"`
 	FilePickerDir                 string `toml:"FilePickerDir"`
 	FilePickerExts                string `toml:"FilePickerExts"`
-	CodingDir                     string `toml:"CodingDir"`
 	ImagePreview                  bool   `toml:"ImagePreview"`
 	EnableMouse                   bool   `toml:"EnableMouse"`
 	// embeddings
-	RAGEnabled bool   `toml:"RAGEnabled"`
+	EmbedURL string `toml:"EmbedURL"`
-	EmbedURL   string `toml:"EmbedURL"`
+	HFToken  string `toml:"HFToken"`
-	HFToken    string `toml:"HFToken"`
-	RAGDir     string `toml:"RAGDir"`
 	// rag settings
+	RAGEnabled   bool   `toml:"RAGEnabled"`
+	RAGDir       string `toml:"RAGDir"`
 	RAGWorkers   uint32 `toml:"RAGWorkers"`
 	RAGBatchSize int    `toml:"RAGBatchSize"`
 	RAGWordLimit uint32 `toml:"RAGWordLimit"`
@@ -126,6 +124,9 @@ func LoadConfig(fn string) (*Config, error) {
 	if config.CompletionAPI != "" {
 		config.ApiLinks = append(config.ApiLinks, config.CompletionAPI)
 	}
+	if config.RAGDir == "" {
+		config.RAGDir = "ragimport"
+	}
 	// if any value is empty fill with default
 	return config, nil
 }

docs/config.md

@@ -140,17 +140,24 @@ This document explains how to set up and configure the application using the `co
 - Path to the SQLite database file used for storing conversation history and other data.
 
 #### FilePickerDir (`"."`)
-- Directory where the file (image) picker should start when selecting files.
+- Directory where the file picker starts and where relative paths in coding assistant file tools (file_read, file_write, etc.) are resolved against. Use absolute paths (starting with `/`) to bypass this.
-
-#### FilePickerExts (`"png,jpg,jpeg,gif,webp"`)
-- Comma-separated list of allowed file extensions for the file picker.
-
-#### CodingDir (`"."`)
-- Default directory for coding assistant file operations. Relative paths in file tools (file_read, file_write, etc.) are resolved against this directory. Use absolute paths (starting with `/`) to bypass this.
 
 #### EnableMouse (`false`)
 - Enable or disable mouse support in the UI. When set to `true`, allows clicking buttons and interacting with UI elements using the mouse, but prevents the terminal from handling mouse events normally (such as selecting and copying text). When set to `false`, enables default terminal behavior allowing you to select and copy text, but disables mouse interaction with UI elements.
 
+### Character-Specific Context Settings (/completion only)
+
+[character specific context page for more info](./char-specific-context.md)
+
+#### CharSpecificContextEnabled (`true`)
+- Enable or disable character-specific context functionality.
+
+#### CharSpecificContextTag (`"@"`)
+- The tag prefix used to reference character-specific context in messages.
+
+#### AutoTurn (`true`)
+- Enable or disable automatic turn detection/switching.
+
 ### Additional Features
 
 Those could be switched in program, but also bould be setup in config.
@@ -158,8 +165,11 @@ Those could be switched in program, but also bould be setup in config.
 #### ToolUse
 - Enable or disable explanation of tools to llm, so it could use them.
 
-#### ThinkUse
+### StripThinkingFromAPI (`true`)
-- Enable or disable insertion of <think> token at the beggining of llm resp.
+- Strip thinking blocks from messages before sending to LLM. Keeps them in chat history for local viewing but reduces token usage in API calls.
+
+#### ReasoningEffort (`"medium"`)
+- OpenRouter reasoning configuration (only applies to OpenRouter chat API). Valid values: `xhigh`, `high`, `medium`, `low`, `minimal`, `none`. Empty or `none` disables reasoning.
 
 ## Environment Variables
 

extra/tts.go

@@ -92,8 +92,6 @@ func (o *KokoroOrator) stoproutine() {
 
 func (o *KokoroOrator) readroutine() {
 	tokenizer, _ := english.NewSentenceTokenizer(nil)
-	// var sentenceBuf bytes.Buffer
-	// var remainder strings.Builder
 	for {
 		select {
 		case chunk := <-TTSTextChan:
@@ -106,24 +104,28 @@ func (o *KokoroOrator) readroutine() {
 				continue
 			}
 			text := o.textBuffer.String()
-			o.mu.Unlock()
 			sentences := tokenizer.Tokenize(text)
 			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
-			for i, sentence := range sentences {
+			if len(sentences) <= 1 {
-				if i == len(sentences)-1 { // last sentence
+				o.mu.Unlock()
-					o.mu.Lock()
+				continue
-					o.textBuffer.Reset()
+			}
-					_, err := o.textBuffer.WriteString(sentence.Text)
+			completeSentences := sentences[:len(sentences)-1]
-					o.mu.Unlock()
+			remaining := sentences[len(sentences)-1].Text
-					if err != nil {
+			o.textBuffer.Reset()
-						o.logger.Warn("failed to write to stringbuilder", "error", err)
+			o.textBuffer.WriteString(remaining)
-						continue
+			o.mu.Unlock()
-					}
+
-					continue // if only one (often incomplete) sentence; wait for next chunk
+			for _, sentence := range completeSentences {
+				o.mu.Lock()
+				interrupted := o.interrupt
+				o.mu.Unlock()
+				if interrupted {
+					return
 				}
 				cleanedText := models.CleanText(sentence.Text)
 				if cleanedText == "" {
-					continue // Skip empty text after cleaning
+					continue
 				}
 				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
 				if err := o.Speak(cleanedText); err != nil {
@@ -338,24 +340,28 @@ func (o *GoogleTranslateOrator) readroutine() {
 				continue
 			}
 			text := o.textBuffer.String()
-			o.mu.Unlock()
 			sentences := tokenizer.Tokenize(text)
 			o.logger.Debug("adding chunk", "chunk", chunk, "text", text, "sen-len", len(sentences))
-			for i, sentence := range sentences {
+			if len(sentences) <= 1 {
-				if i == len(sentences)-1 { // last sentence
+				o.mu.Unlock()
-					o.mu.Lock()
+				continue
-					o.textBuffer.Reset()
+			}
-					_, err := o.textBuffer.WriteString(sentence.Text)
+			completeSentences := sentences[:len(sentences)-1]
-					o.mu.Unlock()
+			remaining := sentences[len(sentences)-1].Text
-					if err != nil {
+			o.textBuffer.Reset()
-						o.logger.Warn("failed to write to stringbuilder", "error", err)
+			o.textBuffer.WriteString(remaining)
-						continue
+			o.mu.Unlock()
-					}
+
-					continue // if only one (often incomplete) sentence; wait for next chunk
+			for _, sentence := range completeSentences {
+				o.mu.Lock()
+				interrupted := o.interrupt
+				o.mu.Unlock()
+				if interrupted {
+					return
 				}
 				cleanedText := models.CleanText(sentence.Text)
 				if cleanedText == "" {
-					continue // Skip empty text after cleaning
+					continue
 				}
 				o.logger.Debug("calling Speak with sentence", "sent", cleanedText)
 				if err := o.Speak(cleanedText); err != nil {

go.mod

@@ -6,17 +6,19 @@ require (
 	github.com/BurntSushi/toml v1.5.0
 	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/gdamore/tcell/v2 v2.13.2
 	github.com/glebarez/go-sqlite v1.22.0
 	github.com/gopxl/beep/v2 v2.1.1
 	github.com/gordonklaus/portaudio v0.0.0-20250206071425-98a94950218b
 	github.com/jmoiron/sqlx v1.4.0
+	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
 	github.com/neurosnap/sentences v1.1.2
 	github.com/rivo/tview v0.42.0
+	github.com/yuin/goldmark v1.4.13
 )
 
 require (
-	github.com/PuerkitoBio/goquery v1.11.0 // indirect
 	github.com/andybalholm/cascadia v1.3.3 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/ebitengine/oto/v3 v3.4.0 // indirect

go.sum

@@ -43,6 +43,8 @@ github.com/hashicorp/golang-lru/v2 v2.0.7 h1:a+bsQ5rvGLjzHuww6tVxozPZFVghXaHOwFs
 github.com/hashicorp/golang-lru/v2 v2.0.7/go.mod h1:QeFd9opnmA6QUJc5vARoKUSoFhyfM2/ZepoAG6RGpeM=
 github.com/jmoiron/sqlx v1.4.0 h1:1PLqN7S1UYp5t4SrVVnt4nUVNemrDAtxlulVe+Qgm3o=
 github.com/jmoiron/sqlx v1.4.0/go.mod h1:ZrZ7UsYB/weZdl2Bxg6jCRO9c3YHl8r3ahlKmRT4JLY=
+github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728 h1:QwWKgMY28TAXaDl+ExRDqGQltzXqN/xypdKP86niVn8=
+github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728/go.mod h1:1fEHWurg7pvf5SG6XNE5Q8UZmOwex51Mkx3SLhrW5B4=
 github.com/lib/pq v1.10.9 h1:YXG7RB+JIjhP29X+OtkiDnYaXQwpS4JEWq7dtCCRUEw=
 github.com/lib/pq v1.10.9/go.mod h1:AlVN5x4E4T544tWzH6hKfbfQvm3HdbOxrmggDNAPY9o=
 github.com/lucasb-eyer/go-colorful v1.3.0 h1:2/yBRLdWBZKrf7gB40FoiKfAWYQ0lqNcbuQwVHXptag=
@@ -67,6 +69,7 @@ github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=
 github.com/rivo/uniseg v0.4.7/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUcx88=
 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/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=
 golang.org/x/crypto v0.0.0-20210921155107-089bfa567519/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc=

helpfuncs.go

@@ -108,14 +108,17 @@ func refreshChatDisplay() {
 	}
 }
 
+// stopTTSIfNotForUser: character specific context, not meant fot the human to hear
 func stopTTSIfNotForUser(msg *models.RoleMsg) {
+	if strings.Contains(cfg.CurrentAPI, "/chat") || !cfg.CharSpecificContextEnabled {
+		return
+	}
 	viewingAs := cfg.UserRole
 	if cfg.WriteNextMsgAs != "" {
 		viewingAs = cfg.WriteNextMsgAs
 	}
 	// stop tts if msg is not for user
-	if cfg.CharSpecificContextEnabled &&
+	if !slices.Contains(msg.KnownTo, viewingAs) && cfg.TTS_ENABLED {
-		!slices.Contains(msg.KnownTo, viewingAs) && cfg.TTS_ENABLED {
 		TTSDoneChan <- true
 	}
 }
@@ -354,13 +357,17 @@ func makeStatusLine() string {
 	}
 	// Get model color based on load status for local llama.cpp models
 	modelColor := getModelColor()
-	statusLine := fmt.Sprintf(statusLineTempl, boolColors[botRespMode], botRespMode, activeChatName,
+	statusLine := fmt.Sprintf(statusLineTempl, boolColors[botRespMode], activeChatName,
-		boolColors[cfg.ToolUse], cfg.ToolUse, modelColor, chatBody.Model, boolColors[cfg.SkipLLMResp],
+		boolColors[cfg.ToolUse], modelColor, chatBody.Model, boolColors[cfg.SkipLLMResp],
-		cfg.SkipLLMResp, cfg.CurrentAPI, boolColors[isRecording], isRecording, persona,
+		cfg.CurrentAPI, persona, botPersona)
-		botPersona)
+	if cfg.STT_ENABLED {
+		recordingS := fmt.Sprintf(" | [%s:-:b]voice recording[-:-:-] (ctrl+r)",
+			boolColors[isRecording])
+		statusLine += recordingS
+	}
 	// completion endpoint
 	if !strings.Contains(cfg.CurrentAPI, "chat") {
-		roleInject := fmt.Sprintf(" | role injection [%s:-:b]%v[-:-:-] (alt+7)", boolColors[injectRole], injectRole)
+		roleInject := fmt.Sprintf(" | [%s:-:b]role injection[-:-:-] (alt+7)", boolColors[injectRole])
 		statusLine += roleInject
 	}
 	return statusLine + imageInfo + shellModeInfo

llm.go

@@ -11,7 +11,6 @@ import (
 
 var imageAttachmentPath string // Global variable to track image attachment for next message
 var lastImg string             // for ctrl+j
-var RAGMsg = "Retrieved context for user's query:\n"
 
 // containsToolSysMsg checks if the toolSysMsg already exists in the chat body
 func containsToolSysMsg() bool {
@@ -142,22 +141,6 @@ func (lcp LCPCompletion) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		newMsg = *processMessageTag(&newMsg)
 		chatBody.Messages = append(chatBody.Messages, newMsg)
 	}
-	// if rag - add as system message to avoid conflicts with tool usage
-	if !resume && cfg.RAGEnabled {
-		um := chatBody.Messages[len(chatBody.Messages)-1].Content
-		logger.Debug("RAG is enabled, preparing RAG context", "user_message", um)
-		ragResp, err := chatRagUse(um)
-		if err != nil {
-			logger.Error("failed to form a rag msg", "error", err)
-			return nil, err
-		}
-		logger.Debug("RAG response received", "response_len", len(ragResp),
-			"response_preview", ragResp[:min(len(ragResp), 100)])
-		// Use system role for RAG context to avoid conflicts with tool usage
-		ragMsg := models.RoleMsg{Role: "system", Content: RAGMsg + ragResp}
-		chatBody.Messages = append(chatBody.Messages, ragMsg)
-		logger.Debug("RAG message added to chat body", "message_count", len(chatBody.Messages))
-	}
 	// sending description of the tools and how to use them
 	if cfg.ToolUse && !resume && role == cfg.UserRole && !containsToolSysMsg() {
 		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
@@ -184,9 +167,6 @@ func (lcp LCPCompletion) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		botMsgStart := "\n" + botPersona + ":\n"
 		prompt += botMsgStart
 	}
-	if cfg.ThinkUse && !cfg.ToolUse {
-		prompt += "<think>"
-	}
 	logger.Debug("checking prompt for /completion", "tool_use", cfg.ToolUse,
 		"msg", msg, "resume", resume, "prompt", prompt, "multimodal_data_count", len(multimodalData))
 	payload := models.NewLCPReq(prompt, chatBody.Model, multimodalData,
@@ -304,23 +284,6 @@ func (op LCPChat) FormMsg(msg, role string, resume bool) (io.Reader, error) {
 		logger.Debug("LCPChat FormMsg: added message to chatBody", "role", newMsg.Role,
 			"content_len", len(newMsg.Content), "message_count_after_add", len(chatBody.Messages))
 	}
-	// if rag - add as system message to avoid conflicts with tool usage
-	if !resume && cfg.RAGEnabled {
-		um := chatBody.Messages[len(chatBody.Messages)-1].Content
-		logger.Debug("LCPChat: RAG is enabled, preparing RAG context", "user_message", um)
-		ragResp, err := chatRagUse(um)
-		if err != nil {
-			logger.Error("LCPChat: failed to form a rag msg", "error", err)
-			return nil, err
-		}
-		logger.Debug("LCPChat: RAG response received",
-			"response_len", len(ragResp), "response_preview", ragResp[:min(len(ragResp), 100)])
-		// Use system role for RAG context to avoid conflicts with tool usage
-		ragMsg := models.RoleMsg{Role: "system", Content: RAGMsg + ragResp}
-		chatBody.Messages = append(chatBody.Messages, ragMsg)
-		logger.Debug("LCPChat: RAG message added to chat body", "role", ragMsg.Role,
-			"rag_content_len", len(ragMsg.Content), "message_count_after_rag", len(chatBody.Messages))
-	}
 	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
 	// openai /v1/chat does not support custom roles; needs to be user, assistant, system
 	// Add persona suffix to the last user message to indicate who the assistant should reply as
@@ -392,22 +355,6 @@ func (ds DeepSeekerCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 		newMsg = *processMessageTag(&newMsg)
 		chatBody.Messages = append(chatBody.Messages, newMsg)
 	}
-	// if rag - add as system message to avoid conflicts with tool usage
-	if !resume && cfg.RAGEnabled {
-		um := chatBody.Messages[len(chatBody.Messages)-1].Content
-		logger.Debug("DeepSeekerCompletion: RAG is enabled, preparing RAG context", "user_message", um)
-		ragResp, err := chatRagUse(um)
-		if err != nil {
-			logger.Error("DeepSeekerCompletion: failed to form a rag msg", "error", err)
-			return nil, err
-		}
-		logger.Debug("DeepSeekerCompletion: RAG response received",
-			"response_len", len(ragResp), "response_preview", ragResp[:min(len(ragResp), 100)])
-		// Use system role for RAG context to avoid conflicts with tool usage
-		ragMsg := models.RoleMsg{Role: "system", Content: RAGMsg + ragResp}
-		chatBody.Messages = append(chatBody.Messages, ragMsg)
-		logger.Debug("DeepSeekerCompletion: RAG message added to chat body", "message_count", len(chatBody.Messages))
-	}
 	// sending description of the tools and how to use them
 	if cfg.ToolUse && !resume && role == cfg.UserRole && !containsToolSysMsg() {
 		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
@@ -423,9 +370,6 @@ func (ds DeepSeekerCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 		botMsgStart := "\n" + botPersona + ":\n"
 		prompt += botMsgStart
 	}
-	if cfg.ThinkUse && !cfg.ToolUse {
-		prompt += "<think>"
-	}
 	logger.Debug("checking prompt for /completion", "tool_use", cfg.ToolUse,
 		"msg", msg, "resume", resume, "prompt", prompt)
 	payload := models.NewDSCompletionReq(prompt, chatBody.Model,
@@ -480,22 +424,6 @@ func (ds DeepSeekerChat) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		newMsg = *processMessageTag(&newMsg)
 		chatBody.Messages = append(chatBody.Messages, newMsg)
 	}
-	// if rag - add as system message to avoid conflicts with tool usage
-	if !resume && cfg.RAGEnabled {
-		um := chatBody.Messages[len(chatBody.Messages)-1].Content
-		logger.Debug("RAG is enabled, preparing RAG context", "user_message", um)
-		ragResp, err := chatRagUse(um)
-		if err != nil {
-			logger.Error("failed to form a rag msg", "error", err)
-			return nil, err
-		}
-		logger.Debug("RAG response received", "response_len", len(ragResp),
-			"response_preview", ragResp[:min(len(ragResp), 100)])
-		// Use system role for RAG context to avoid conflicts with tool usage
-		ragMsg := models.RoleMsg{Role: "system", Content: RAGMsg + ragResp}
-		chatBody.Messages = append(chatBody.Messages, ragMsg)
-		logger.Debug("RAG message added to chat body", "message_count", len(chatBody.Messages))
-	}
 	// Create copy of chat body with standardized user role
 	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
 	// Add persona suffix to the last user message to indicate who the assistant should reply as
@@ -558,22 +486,6 @@ func (or OpenRouterCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 		newMsg = *processMessageTag(&newMsg)
 		chatBody.Messages = append(chatBody.Messages, newMsg)
 	}
-	// if rag - add as system message to avoid conflicts with tool usage
-	if !resume && cfg.RAGEnabled {
-		um := chatBody.Messages[len(chatBody.Messages)-1].Content
-		logger.Debug("RAG is enabled, preparing RAG context", "user_message", um)
-		ragResp, err := chatRagUse(um)
-		if err != nil {
-			logger.Error("failed to form a rag msg", "error", err)
-			return nil, err
-		}
-		logger.Debug("RAG response received", "response_len",
-			len(ragResp), "response_preview", ragResp[:min(len(ragResp), 100)])
-		// Use system role for RAG context to avoid conflicts with tool usage
-		ragMsg := models.RoleMsg{Role: "system", Content: RAGMsg + ragResp}
-		chatBody.Messages = append(chatBody.Messages, ragMsg)
-		logger.Debug("RAG message added to chat body", "message_count", len(chatBody.Messages))
-	}
 	// sending description of the tools and how to use them
 	if cfg.ToolUse && !resume && role == cfg.UserRole && !containsToolSysMsg() {
 		chatBody.Messages = append(chatBody.Messages, models.RoleMsg{Role: cfg.ToolRole, Content: toolSysMsg})
@@ -589,9 +501,6 @@ func (or OpenRouterCompletion) FormMsg(msg, role string, resume bool) (io.Reader
 		botMsgStart := "\n" + botPersona + ":\n"
 		prompt += botMsgStart
 	}
-	if cfg.ThinkUse && !cfg.ToolUse {
-		prompt += "<think>"
-	}
 	stopSlice := chatBody.MakeStopSliceExcluding("", listChatRoles())
 	logger.Debug("checking prompt for /completion", "tool_use", cfg.ToolUse,
 		"msg", msg, "resume", resume, "prompt", prompt, "stop_strings", stopSlice)
@@ -679,22 +588,6 @@ func (or OpenRouterChat) FormMsg(msg, role string, resume bool) (io.Reader, erro
 		newMsg = *processMessageTag(&newMsg)
 		chatBody.Messages = append(chatBody.Messages, newMsg)
 	}
-	// if rag - add as system message to avoid conflicts with tool usage
-	if !resume && cfg.RAGEnabled {
-		um := chatBody.Messages[len(chatBody.Messages)-1].Content
-		logger.Debug("RAG is enabled, preparing RAG context", "user_message", um)
-		ragResp, err := chatRagUse(um)
-		if err != nil {
-			logger.Error("failed to form a rag msg", "error", err)
-			return nil, err
-		}
-		logger.Debug("RAG response received", "response_len", len(ragResp),
-			"response_preview", ragResp[:min(len(ragResp), 100)])
-		// Use system role for RAG context to avoid conflicts with tool usage
-		ragMsg := models.RoleMsg{Role: "system", Content: RAGMsg + ragResp}
-		chatBody.Messages = append(chatBody.Messages, ragMsg)
-		logger.Debug("RAG message added to chat body", "message_count", len(chatBody.Messages))
-	}
 	// Create copy of chat body with standardized user role
 	filteredMessages, _ := filterMessagesForCurrentCharacter(chatBody.Messages)
 	// Add persona suffix to the last user message to indicate who the assistant should reply as

main.go

@@ -7,13 +7,14 @@ import (
 var (
 	boolColors        = map[bool]string{true: "green", false: "red"}
 	botRespMode       = false
+	toolRunningMode   = false
 	editMode          = false
 	roleEditMode      = false
 	injectRole        = true
 	selectedIndex     = int(-1)
 	shellMode         = false
 	thinkingCollapsed = false
-	statusLineTempl   = "help (F12) | llm turn: [%s:-:b]%v[-:-:-] (F6) | chat: [orange:-:b]%s[-:-:-] (F1) |tool-use: [%s:-:b]%v[-:-:-] (ctrl+k) | model: [%s:-:b]%s[-:-:-] (ctrl+l) | skip LLM resp: [%s:-:b]%v[-:-:-] (F10)\nAPI: [orange:-:b]%s[-:-:-] (ctrl+v) | voice recording: [%s:-:b]%v[-:-:-] (ctrl+r) | writing as: [orange:-:b]%s[-:-:-] (ctrl+q) | bot will write as [orange:-:b]%s[-:-:-] (ctrl+x)"
+	statusLineTempl   = "help (F12) | [%s:-:b]llm writes[-:-:-] (F6 to interrupt) | chat: [orange:-:b]%s[-:-:-] (F1) | [%s:-:b]tool use[-:-:-] (ctrl+k) | model: [%s:-:b]%s[-:-:-] (ctrl+l) | [%s:-:b]skip LLM resp[-:-:-] (F10)\nAPI: [orange:-:b]%s[-:-:-] (ctrl+v) | writing as: [orange:-:b]%s[-:-:-] (ctrl+q) | bot will write as [orange:-:b]%s[-:-:-] (ctrl+x)"
 	focusSwitcher     = map[tview.Primitive]tview.Primitive{}
 )
 

models/models.go

@@ -105,12 +105,13 @@ type ImageContentPart struct {
 
 // RoleMsg represents a message with content that can be either a simple string or structured content parts
 type RoleMsg struct {
-	Role            string   `json:"role"`
+	Role            string         `json:"role"`
-	Content         string   `json:"-"`
+	Content         string         `json:"-"`
-	ContentParts    []any    `json:"-"`
+	ContentParts    []any          `json:"-"`
-	ToolCallID      string   `json:"tool_call_id,omitempty"` // For tool response messages
+	ToolCallID      string         `json:"tool_call_id,omitempty"` // For tool response messages
-	KnownTo         []string `json:"known_to,omitempty"`
+	KnownTo         []string       `json:"known_to,omitempty"`
-	hasContentParts bool     // Flag to indicate which content type to marshal
+	Stats           *ResponseStats `json:"stats"`
+	hasContentParts bool           // Flag to indicate which content type to marshal
 }
 
 // MarshalJSON implements custom JSON marshaling for RoleMsg
@@ -183,13 +184,11 @@ func (m *RoleMsg) UnmarshalJSON(data []byte) error {
 }
 
 func (m *RoleMsg) ToText(i int) string {
-	// Convert content to string representation
 	var contentStr string
 	var imageIndicators []string
 	if !m.hasContentParts {
 		contentStr = m.Content
 	} else {
-		// For structured content, collect text parts and image indicators
 		var textParts []string
 		for _, part := range m.ContentParts {
 			switch p := part.(type) {
@@ -198,7 +197,6 @@ func (m *RoleMsg) ToText(i int) string {
 					textParts = append(textParts, p.Text)
 				}
 			case ImageContentPart:
-				// Collect image indicator
 				displayPath := p.Path
 				if displayPath == "" {
 					displayPath = "image"
@@ -216,7 +214,6 @@ func (m *RoleMsg) ToText(i int) string {
 							}
 						}
 					case "image_url":
-						// Handle unmarshaled image content
 						var displayPath string
 						if pathVal, pathExists := p["path"]; pathExists {
 							if pathStr, isStr := pathVal.(string); isStr && pathStr != "" {
@@ -233,23 +230,19 @@ func (m *RoleMsg) ToText(i int) string {
 		}
 		contentStr = strings.Join(textParts, " ") + " "
 	}
-	// check if already has role annotation (/completion makes them)
-	// in that case remove it, and then add to icon
-	// since icon and content are separated by \n
 	contentStr, _ = strings.CutPrefix(contentStr, m.Role+":")
-	// if !strings.HasPrefix(contentStr, m.Role+":") {
 	icon := fmt.Sprintf("(%d) <%s>: ", i, m.Role)
-	// }
-	// Build final message with image indicators before text
 	var finalContent strings.Builder
 	if len(imageIndicators) > 0 {
-		// Add each image indicator on its own line
 		for _, indicator := range imageIndicators {
 			finalContent.WriteString(indicator)
 			finalContent.WriteString("\n")
 		}
 	}
 	finalContent.WriteString(contentStr)
+	if m.Stats != nil {
+		fmt.Fprintf(&finalContent, "\n[gray::i][%d tok, %.1fs, %.1f t/s][-:-:-]", m.Stats.Tokens, m.Stats.Duration, m.Stats.TokensPerSec)
+	}
 	textMsg := fmt.Sprintf("[-:-:b]%s[-:-:-]\n%s\n", icon, finalContent.String())
 	return strings.ReplaceAll(textMsg, "\n\n", "\n")
 }
@@ -331,6 +324,7 @@ func (m *RoleMsg) Copy() RoleMsg {
 		ContentParts:    m.ContentParts,
 		ToolCallID:      m.ToolCallID,
 		KnownTo:         m.KnownTo,
+		Stats:           m.Stats,
 		hasContentParts: m.hasContentParts,
 	}
 }
@@ -643,6 +637,12 @@ func (lcp *LCPModels) ListModels() []string {
 	return resp
 }
 
+type ResponseStats struct {
+	Tokens       int
+	Duration     float64
+	TokensPerSec float64
+}
+
 type ChatRoundReq struct {
 	UserMsg string
 	Role    string

popups.go

@@ -51,7 +51,7 @@ func showModelSelectionPopup() {
 	// Find the current model index to set as selected
 	currentModelIndex := -1
 	for i, model := range modelList {
-		if model == chatBody.Model {
+		if strings.TrimPrefix(model, "(loaded) ") == chatBody.Model {
 			currentModelIndex = i
 		}
 		modelListWidget.AddItem(model, "", 0, nil)
@@ -61,23 +61,23 @@ func showModelSelectionPopup() {
 		modelListWidget.SetCurrentItem(currentModelIndex)
 	}
 	modelListWidget.SetSelectedFunc(func(index int, mainText string, secondaryText string, shortcut rune) {
-		// Strip "(loaded)" suffix if present for local llama.cpp models
+		modelName := strings.TrimPrefix(mainText, "(loaded) ")
-		modelName := strings.TrimSuffix(mainText, " (loaded)")
-		// Update the model in both chatBody and config
 		chatBody.Model = modelName
 		cfg.CurrentModel = chatBody.Model
-		// Remove the popup page
 		pages.RemovePage("modelSelectionPopup")
-		// Update the status line to reflect the change
+		app.SetFocus(textArea)
+		updateCachedModelColor()
 		updateStatusLine()
 	})
 	modelListWidget.SetInputCapture(func(event *tcell.EventKey) *tcell.EventKey {
 		if event.Key() == tcell.KeyEscape {
 			pages.RemovePage("modelSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == 'x' {
 			pages.RemovePage("modelSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		return event
@@ -162,19 +162,21 @@ func showAPILinkSelectionPopup() {
 			chatBody.Model = newModelList[0]
 			cfg.CurrentModel = chatBody.Model
 		}
-		// Remove the popup page
 		pages.RemovePage("apiLinkSelectionPopup")
-		// Update the parser and status line to reflect the change
+		app.SetFocus(textArea)
 		choseChunkParser()
+		updateCachedModelColor()
 		updateStatusLine()
 	})
 	apiListWidget.SetInputCapture(func(event *tcell.EventKey) *tcell.EventKey {
 		if event.Key() == tcell.KeyEscape {
 			pages.RemovePage("apiLinkSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == 'x' {
 			pages.RemovePage("apiLinkSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		return event
@@ -234,6 +236,7 @@ func showUserRoleSelectionPopup() {
 		textView.SetText(chatToText(filtered, cfg.ShowSys))
 		// Remove the popup page
 		pages.RemovePage("userRoleSelectionPopup")
+		app.SetFocus(textArea)
 		// Update the status line to reflect the change
 		updateStatusLine()
 		colorText()
@@ -241,10 +244,12 @@ func showUserRoleSelectionPopup() {
 	roleListWidget.SetInputCapture(func(event *tcell.EventKey) *tcell.EventKey {
 		if event.Key() == tcell.KeyEscape {
 			pages.RemovePage("userRoleSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == 'x' {
 			pages.RemovePage("userRoleSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		return event
@@ -307,16 +312,19 @@ func showBotRoleSelectionPopup() {
 		cfg.WriteNextMsgAsCompletionAgent = mainText
 		// Remove the popup page
 		pages.RemovePage("botRoleSelectionPopup")
+		app.SetFocus(textArea)
 		// Update the status line to reflect the change
 		updateStatusLine()
 	})
 	roleListWidget.SetInputCapture(func(event *tcell.EventKey) *tcell.EventKey {
 		if event.Key() == tcell.KeyEscape {
 			pages.RemovePage("botRoleSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == 'x' {
 			pages.RemovePage("botRoleSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		return event
@@ -336,7 +344,7 @@ func showBotRoleSelectionPopup() {
 }
 
 func showFileCompletionPopup(filter string) {
-	baseDir := cfg.CodingDir
+	baseDir := cfg.FilePickerDir
 	if baseDir == "" {
 		baseDir = "."
 	}
@@ -368,14 +376,17 @@ func showFileCompletionPopup(filter string) {
 			textArea.SetText(before+mainText, true)
 		}
 		pages.RemovePage("fileCompletionPopup")
+		app.SetFocus(textArea)
 	})
 	widget.SetInputCapture(func(event *tcell.EventKey) *tcell.EventKey {
 		if event.Key() == tcell.KeyEscape {
 			pages.RemovePage("fileCompletionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == 'x' {
 			pages.RemovePage("fileCompletionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		return event
@@ -488,14 +499,17 @@ func showColorschemeSelectionPopup() {
 		}
 		// Remove the popup page
 		pages.RemovePage("colorschemeSelectionPopup")
+		app.SetFocus(textArea)
 	})
 	schemeListWidget.SetInputCapture(func(event *tcell.EventKey) *tcell.EventKey {
 		if event.Key() == tcell.KeyEscape {
 			pages.RemovePage("colorschemeSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == 'x' {
 			pages.RemovePage("colorschemeSelectionPopup")
+			app.SetFocus(textArea)
 			return nil
 		}
 		return event

props_table.go

@@ -115,9 +115,6 @@ func makePropsTable(props map[string]float32) *tview.Table {
 		row++
 	}
 	// Add checkboxes
-	addCheckboxRow("Insert <think> tag (/completion only)", cfg.ThinkUse, func(checked bool) {
-		cfg.ThinkUse = checked
-	})
 	addCheckboxRow("RAG use", cfg.RAGEnabled, func(checked bool) {
 		cfg.RAGEnabled = checked
 	})

rag/extractors.go

@@ -0,0 +1,184 @@
+package rag
+
+import (
+	"archive/zip"
+	"bytes"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path"
+	"strings"
+
+	"github.com/PuerkitoBio/goquery"
+	"github.com/ledongthuc/pdf"
+	"github.com/yuin/goldmark"
+	"github.com/yuin/goldmark/extension"
+	"github.com/yuin/goldmark/parser"
+	"github.com/yuin/goldmark/renderer/html"
+)
+
+func ExtractText(fpath string) (string, error) {
+	ext := strings.ToLower(path.Ext(fpath))
+	switch ext {
+	case ".txt":
+		return extractTextFromFile(fpath)
+	case ".md", ".markdown":
+		return extractTextFromMarkdown(fpath)
+	case ".html", ".htm":
+		return extractTextFromHtmlFile(fpath)
+	case ".epub":
+		return extractTextFromEpub(fpath)
+	case ".pdf":
+		return extractTextFromPdf(fpath)
+	default:
+		return "", fmt.Errorf("unsupported file format: %s", ext)
+	}
+}
+
+func extractTextFromFile(fpath string) (string, error) {
+	data, err := os.ReadFile(fpath)
+	if err != nil {
+		return "", err
+	}
+	return string(data), nil
+}
+
+func extractTextFromHtmlFile(fpath string) (string, error) {
+	data, err := os.ReadFile(fpath)
+	if err != nil {
+		return "", err
+	}
+	return extractTextFromHtmlContent(data)
+}
+
+// non utf-8 encoding?
+func extractTextFromHtmlContent(data []byte) (string, error) {
+	doc, err := goquery.NewDocumentFromReader(bytes.NewReader(data))
+	if err != nil {
+		return "", err
+	}
+	// Remove script and style tags
+	doc.Find("script, style, noscript").Each(func(i int, s *goquery.Selection) {
+		s.Remove()
+	})
+	// Get text and clean it
+	text := doc.Text()
+	// Collapse all whitespace (newlines, tabs, multiple spaces) into single spaces
+	cleaned := strings.Join(strings.Fields(text), " ")
+	return cleaned, nil
+}
+
+func extractTextFromMarkdown(fpath string) (string, error) {
+	data, err := os.ReadFile(fpath)
+	if err != nil {
+		return "", err
+	}
+	// Convert markdown to HTML
+	md := goldmark.New(
+		goldmark.WithExtensions(extension.GFM),
+		goldmark.WithParserOptions(parser.WithAutoHeadingID()),
+		goldmark.WithRendererOptions(html.WithUnsafe()), // allow raw HTML if needed
+	)
+	var buf bytes.Buffer
+	if err := md.Convert(data, &buf); err != nil {
+		return "", err
+	}
+	// Now extract text from the resulting HTML (using goquery or similar)
+	return extractTextFromHtmlContent(buf.Bytes())
+}
+
+func extractTextFromEpub(fpath string) (string, error) {
+	r, err := zip.OpenReader(fpath)
+	if err != nil {
+		return "", fmt.Errorf("failed to open epub: %w", err)
+	}
+	defer r.Close()
+
+	var sb strings.Builder
+
+	for _, f := range r.File {
+		ext := strings.ToLower(path.Ext(f.Name))
+		if ext != ".xhtml" && ext != ".html" && ext != ".htm" && ext != ".xml" {
+			continue
+		}
+
+		// Skip manifest, toc, ncx files - they don't contain book content
+		nameLower := strings.ToLower(f.Name)
+		if strings.Contains(nameLower, "toc") || strings.Contains(nameLower, "nav") ||
+			strings.Contains(nameLower, "manifest") || strings.Contains(nameLower, ".opf") ||
+			strings.HasSuffix(nameLower, ".ncx") {
+			continue
+		}
+
+		rc, err := f.Open()
+		if err != nil {
+			continue
+		}
+
+		if sb.Len() > 0 {
+			sb.WriteString("\n\n")
+		}
+		sb.WriteString(f.Name)
+		sb.WriteString("\n")
+
+		buf, readErr := io.ReadAll(rc)
+		rc.Close()
+		if readErr == nil {
+			sb.WriteString(stripHTML(string(buf)))
+		}
+	}
+
+	if sb.Len() == 0 {
+		return "", errors.New("no content extracted from epub")
+	}
+	return sb.String(), nil
+}
+
+func stripHTML(html string) string {
+	var sb strings.Builder
+	inTag := false
+	for _, r := range html {
+		switch r {
+		case '<':
+			inTag = true
+		case '>':
+			inTag = false
+		default:
+			if !inTag {
+				sb.WriteRune(r)
+			}
+		}
+	}
+	return sb.String()
+}
+
+func extractTextFromPdf(fpath string) (string, error) {
+	_, err := exec.LookPath("pdftotext")
+	if err == nil {
+		out, err := exec.Command("pdftotext", "-layout", fpath, "-").Output()
+		if err == nil && len(out) > 0 {
+			return string(out), nil
+		}
+	}
+	return extractTextFromPdfPureGo(fpath)
+}
+
+func extractTextFromPdfPureGo(fpath string) (string, error) {
+	df, r, err := pdf.Open(fpath)
+	if err != nil {
+		return "", fmt.Errorf("failed to open pdf: %w", err)
+	}
+	defer df.Close()
+	textReader, err := r.GetPlainText()
+	if err != nil {
+		return "", fmt.Errorf("failed to extract text from pdf: %w", err)
+	}
+	var buf bytes.Buffer
+	_, err = io.Copy(&buf, textReader)
+	if err != nil {
+		return "", fmt.Errorf("failed to read pdf text: %w", err)
+	}
+	return buf.String(), nil
+}

rag/rag.go

@@ -7,8 +7,9 @@ import (
 	"gf-lt/models"
 	"gf-lt/storage"
 	"log/slog"
-	"os"
 	"path"
+	"regexp"
+	"sort"
 	"strings"
 	"sync"
 
@@ -23,13 +24,13 @@ var (
 	ErrRAGStatus        = "some error occurred; failed to transfer data to vector db"
 )
 
-
 type RAG struct {
 	logger   *slog.Logger
 	store    storage.FullRepo
 	cfg      *config.Config
 	embedder Embedder
 	storage  *VectorStorage
+	mu       sync.Mutex
 }
 
 func New(l *slog.Logger, s storage.FullRepo, cfg *config.Config) *RAG {
@@ -54,7 +55,9 @@ func wordCounter(sentence string) int {
 }
 
 func (r *RAG) LoadRAG(fpath string) error {
-	data, err := os.ReadFile(fpath)
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	fileText, err := ExtractText(fpath)
 	if err != nil {
 		return err
 	}
@@ -63,10 +66,7 @@ func (r *RAG) LoadRAG(fpath string) error {
 	case LongJobStatusCh <- LoadedFileRAGStatus:
 	default:
 		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", LoadedFileRAGStatus)
-		// Channel is full or closed, ignore the message to prevent panic
 	}
-
-	fileText := string(data)
 	tokenizer, err := english.NewSentenceTokenizer(nil)
 	if err != nil {
 		return err
@@ -76,19 +76,16 @@ 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++ {
-		// Only add sentences that aren't empty
 		if strings.TrimSpace(sents[i]) != "" {
 			if par.Len() > 0 {
-				par.WriteString(" ") // Add space between sentences
+				par.WriteString(" ")
 			}
 			par.WriteString(sents[i])
 		}
-
 		if wordCounter(par.String()) > int(r.cfg.RAGWordLimit) {
 			paragraph := strings.TrimSpace(par.String())
 			if paragraph != "" {
@@ -97,7 +94,6 @@ func (r *RAG) LoadRAG(fpath string) error {
 			par.Reset()
 		}
 	}
-
 	// Handle any remaining content in the paragraph buffer
 	if par.Len() > 0 {
 		paragraph := strings.TrimSpace(par.String())
@@ -105,215 +101,82 @@ func (r *RAG) LoadRAG(fpath string) error {
 			paragraphs = append(paragraphs, paragraph)
 		}
 	}
-
 	// Adjust batch size if needed
 	if len(paragraphs) < r.cfg.RAGBatchSize && len(paragraphs) > 0 {
 		r.cfg.RAGBatchSize = len(paragraphs)
 	}
-
 	if len(paragraphs) == 0 {
 		return errors.New("no valid paragraphs found in file")
 	}
-
+	// Process paragraphs in batches synchronously
-	var (
+	batchCount := 0
-		maxChSize = 100
+	for i := 0; i < len(paragraphs); i += r.cfg.RAGBatchSize {
-		left      = 0
+		end := i + r.cfg.RAGBatchSize
-		right     = r.cfg.RAGBatchSize
+		if end > len(paragraphs) {
-		batchCh   = make(chan map[int][]string, maxChSize)
+			end = len(paragraphs)
-		vectorCh  = make(chan []models.VectorRow, maxChSize)
-		errCh     = make(chan error, 1)
-		wg        = new(sync.WaitGroup)
-		lock      = new(sync.Mutex)
-	)
-
-	defer close(errCh)
-	defer close(batchCh)
-
-	// Fill input channel with batches
-	ctn := 0
-	totalParagraphs := len(paragraphs)
-	for {
-		if right > totalParagraphs {
-			batchCh <- map[int][]string{left: paragraphs[left:]}
-			break
 		}
-		batchCh <- map[int][]string{left: paragraphs[left:right]}
+		batch := paragraphs[i:end]
-		left, right = right, right+r.cfg.RAGBatchSize
+		batchCount++
-		ctn++
+		// Filter empty paragraphs
-	}
+		nonEmptyBatch := make([]string, 0, len(batch))
-
+		for _, p := range batch {
-	finishedBatchesMsg := fmt.Sprintf("finished batching batches#: %d; paragraphs: %d; sentences: %d\n", ctn+1, len(paragraphs), len(sents))
+			if strings.TrimSpace(p) != "" {
-	r.logger.Debug(finishedBatchesMsg)
+				nonEmptyBatch = append(nonEmptyBatch, strings.TrimSpace(p))
-	select {
+			}
-	case LongJobStatusCh <- finishedBatchesMsg:
+		}
-	default:
+		if len(nonEmptyBatch) == 0 {
-		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", finishedBatchesMsg)
+			continue
-		// Channel is full or closed, ignore the message to prevent panic
+		}
-	}
+		// Embed the batch
-
+		embeddings, err := r.embedder.EmbedSlice(nonEmptyBatch)
-	// Start worker goroutines with WaitGroup
-	wg.Add(int(r.cfg.RAGWorkers))
-	for w := 0; w < int(r.cfg.RAGWorkers); w++ {
-		go func(workerID int) {
-			defer wg.Done()
-			r.batchToVectorAsync(lock, workerID, batchCh, vectorCh, errCh, path.Base(fpath))
-		}(w)
-	}
-
-	// Use a goroutine to close the batchCh when all batches are sent
-	go func() {
-		wg.Wait()
-		close(vectorCh) // Close vectorCh when all workers are done
-	}()
-
-	// Check for errors from workers
-	// Use a non-blocking check for errors
-	select {
-	case err := <-errCh:
 		if err != nil {
-			r.logger.Error("error during RAG processing", "error", err)
+			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))
 			return err
 		}
-	default:
+		// Write vectors to storage
-		// No immediate error, continue
+		filename := path.Base(fpath)
-	}
+		for j, text := range nonEmptyBatch {
-
+			vector := models.VectorRow{
-	// Write vectors to storage - this will block until vectorCh is closed
+				Embeddings: embeddings[j],
-	return r.writeVectors(vectorCh)
+				RawText:    text,
-}
+				Slug:       fmt.Sprintf("%s_%d_%d", filename, batchCount, j),
-
+				FileName:   filename,
-func (r *RAG) writeVectors(vectorCh chan []models.VectorRow) error {
-	for {
-		for batch := range vectorCh {
-			for _, vector := range batch {
-				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 is full or closed, dropping status message", "message", ErrRAGStatus)
-						// Channel is full or closed, ignore the message to prevent panic
-					}
-					return err // Stop the entire RAG operation on DB error
-				}
 			}
-			r.logger.Debug("wrote batch to db", "size", len(batch), "vector_chan_len", len(vectorCh))
+			if err := r.storage.WriteVector(&vector); err != nil {
-			if len(vectorCh) == 0 {
+				r.logger.Error("failed to write vector to DB", "error", err, "slug", vector.Slug)
-				r.logger.Debug("finished writing vectors")
 				select {
-				case LongJobStatusCh <- FinishedRAGStatus:
+				case LongJobStatusCh <- ErrRAGStatus:
 				default:
-					r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", FinishedRAGStatus)
+					r.logger.Warn("LongJobStatusCh channel full, dropping message")
-					// Channel is full or closed, ignore the message to prevent panic
 				}
-				return nil
+				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)
-func (r *RAG) batchToVectorAsync(lock *sync.Mutex, id int, inputCh <-chan map[int][]string,
-	vectorCh chan<- []models.VectorRow, errCh chan error, filename string) {
-	var err error
-
-	defer func() {
-		// For errCh, make sure we only send if there's actually an error and the channel can accept it
-		if err != nil {
-			select {
-			case errCh <- err:
-			default:
-				// errCh might be full or closed, log but don't panic
-				r.logger.Warn("errCh channel full or closed, skipping error propagation", "worker", id, "error", err)
-			}
-		}
-	}()
-
-	for {
-		lock.Lock()
-		if len(inputCh) == 0 {
-			lock.Unlock()
-			return
-		}
-
-		select {
-		case linesMap := <-inputCh:
-			for leftI, lines := range linesMap {
-				if err := r.fetchEmb(lines, errCh, vectorCh, fmt.Sprintf("%s_%d", filename, leftI), filename); err != nil {
-					r.logger.Error("error fetching embeddings", "error", err, "worker", id)
-					lock.Unlock()
-					return
-				}
-			}
-			lock.Unlock()
-		case err = <-errCh:
-			r.logger.Error("got an error from error channel", "error", err)
-			lock.Unlock()
-			return
-		default:
-			lock.Unlock()
-		}
-
-		r.logger.Debug("processed batch", "batches#", len(inputCh), "worker#", id)
-		statusMsg := fmt.Sprintf("converted to vector; batches: %d, worker#: %d", len(inputCh), id)
 		select {
 		case LongJobStatusCh <- statusMsg:
 		default:
-			r.logger.Warn("LongJobStatusCh channel full or closed, dropping status message", "message", statusMsg)
+			r.logger.Warn("LongJobStatusCh channel full, dropping message")
-			// Channel is full or closed, ignore the message to prevent panic
 		}
 	}
-}
+	r.logger.Debug("finished writing vectors", "batches", batchCount)
-
+	select {
-func (r *RAG) fetchEmb(lines []string, errCh chan error, vectorCh chan<- []models.VectorRow, slug, filename string) error {
+	case LongJobStatusCh <- FinishedRAGStatus:
-	// Filter out empty lines before sending to embedder
+	default:
-	nonEmptyLines := make([]string, 0, len(lines))
+		r.logger.Warn("LongJobStatusCh channel is full or closed, dropping status message", "message", FinishedRAGStatus)
-	for _, line := range lines {
-		trimmed := strings.TrimSpace(line)
-		if trimmed != "" {
-			nonEmptyLines = append(nonEmptyLines, trimmed)
-		}
 	}
-
-	// Skip if no non-empty lines
-	if len(nonEmptyLines) == 0 {
-		// Send empty result but don't error
-		vectorCh <- []models.VectorRow{}
-		return nil
-	}
-
-	embeddings, err := r.embedder.EmbedSlice(nonEmptyLines)
-	if err != nil {
-		r.logger.Error("failed to embed lines", "err", err.Error())
-		errCh <- err
-		return err
-	}
-
-	if len(embeddings) == 0 {
-		err := errors.New("no embeddings returned")
-		r.logger.Error("empty embeddings")
-		errCh <- err
-		return err
-	}
-
-	if len(embeddings) != len(nonEmptyLines) {
-		err := errors.New("mismatch between number of lines and embeddings returned")
-		r.logger.Error("embedding mismatch", "err", err.Error())
-		errCh <- err
-		return err
-	}
-
-	// Create a VectorRow for each line in the batch
-	vectors := make([]models.VectorRow, len(nonEmptyLines))
-	for i, line := range nonEmptyLines {
-		vectors[i] = models.VectorRow{
-			Embeddings: embeddings[i],
-			RawText:    line,
-			Slug:       fmt.Sprintf("%s_%d", slug, i),
-			FileName:   filename,
-		}
-	}
-
-	vectorCh <- vectors
 	return nil
 }
 
@@ -332,3 +195,309 @@ func (r *RAG) ListLoaded() ([]string, error) {
 func (r *RAG) RemoveFile(filename string) error {
 	return r.storage.RemoveEmbByFileName(filename)
 }
+
+var (
+	queryRefinementPattern = regexp.MustCompile(`(?i)(based on my (vector db|vector db|vector database|rags?|past (conversations?|chat|messages?))|from my (files?|documents?|data|information|memory)|search (in|my) (vector db|database|rags?)|rag search for)`)
+	importantKeywords      = []string{"project", "architecture", "code", "file", "chat", "conversation", "topic", "summary", "details", "history", "previous", "my", "user", "me"}
+	stopWords              = []string{"the", "a", "an", "and", "or", "but", "in", "on", "at", "to", "for", "of", "with", "by", "from", "up", "down", "left", "right"}
+)
+
+func (r *RAG) RefineQuery(query string) string {
+	original := query
+	query = strings.TrimSpace(query)
+
+	if len(query) == 0 {
+		return original
+	}
+
+	if len(query) <= 3 {
+		return original
+	}
+
+	query = strings.ToLower(query)
+
+	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
+	}
+
+	if queryRefinementPattern.MatchString(original) {
+		cleaned := queryRefinementPattern.ReplaceAllString(original, "")
+		cleaned = strings.TrimSpace(cleaned)
+		if len(cleaned) >= 5 {
+			return cleaned
+		}
+	}
+
+	query = r.extractImportantPhrases(query)
+
+	if len(query) < 5 {
+		return original
+	}
+
+	return query
+}
+
+func (r *RAG) extractImportantPhrases(query string) string {
+	words := strings.Fields(query)
+
+	var important []string
+	for _, word := range words {
+		word = strings.Trim(word, ".,!?;:'\"()[]{}")
+
+		isImportant := false
+		for _, kw := range importantKeywords {
+			if strings.Contains(strings.ToLower(word), kw) {
+				isImportant = true
+				break
+			}
+		}
+
+		if isImportant || len(word) > 3 {
+			important = append(important, word)
+		}
+	}
+
+	if len(important) == 0 {
+		return query
+	}
+
+	return strings.Join(important, " ")
+}
+
+func (r *RAG) GenerateQueryVariations(query string) []string {
+	variations := []string{query}
+
+	if len(query) < 5 {
+		return variations
+	}
+
+	parts := strings.Fields(query)
+	if len(parts) == 0 {
+		return variations
+	}
+
+	if len(parts) >= 2 {
+		trimmed := strings.Join(parts[:len(parts)-1], " ")
+		if len(trimmed) >= 5 {
+			variations = append(variations, trimmed)
+		}
+	}
+
+	if len(parts) >= 2 {
+		trimmed := strings.Join(parts[1:], " ")
+		if len(trimmed) >= 5 {
+			variations = append(variations, trimmed)
+		}
+	}
+
+	if !strings.HasSuffix(query, " explanation") {
+		variations = append(variations, query+" explanation")
+	}
+	if !strings.HasPrefix(query, "what is ") {
+		variations = append(variations, "what is "+query)
+	}
+	if !strings.HasSuffix(query, " details") {
+		variations = append(variations, query+" details")
+	}
+	if !strings.HasSuffix(query, " summary") {
+		variations = append(variations, query+" summary")
+	}
+
+	return variations
+}
+
+func (r *RAG) RerankResults(results []models.VectorRow, query string) []models.VectorRow {
+	type scoredResult struct {
+		row      models.VectorRow
+		distance float32
+	}
+
+	scored := make([]scoredResult, 0, len(results))
+
+	for i := range results {
+		row := results[i]
+
+		score := float32(0)
+
+		rawTextLower := strings.ToLower(row.RawText)
+		queryLower := strings.ToLower(query)
+
+		if strings.Contains(rawTextLower, queryLower) {
+			score += 10
+		}
+
+		queryWords := strings.Fields(queryLower)
+		matchCount := 0
+		for _, word := range queryWords {
+			if len(word) > 2 && strings.Contains(rawTextLower, word) {
+				matchCount++
+			}
+		}
+		if len(queryWords) > 0 {
+			score += float32(matchCount) / float32(len(queryWords)) * 5
+		}
+
+		if row.FileName == "chat" || strings.Contains(strings.ToLower(row.FileName), "conversation") {
+			score += 3
+		}
+
+		distance := row.Distance - score/100
+
+		scored = append(scored, scoredResult{row: row, distance: distance})
+	}
+
+	sort.Slice(scored, func(i, j int) bool {
+		return scored[i].distance < scored[j].distance
+	})
+
+	unique := make([]models.VectorRow, 0)
+	seen := make(map[string]bool)
+
+	for i := range scored {
+		if !seen[scored[i].row.Slug] {
+			seen[scored[i].row.Slug] = true
+			unique = append(unique, scored[i].row)
+		}
+	}
+
+	if len(unique) > 10 {
+		unique = unique[:10]
+	}
+
+	return unique
+}
+
+func (r *RAG) SynthesizeAnswer(results []models.VectorRow, query string) (string, error) {
+	if len(results) == 0 {
+		return "No relevant information found in the vector database.", nil
+	}
+
+	var contextBuilder strings.Builder
+	contextBuilder.WriteString("User Query: ")
+	contextBuilder.WriteString(query)
+	contextBuilder.WriteString("\n\nRetrieved Context:\n")
+
+	for i, row := range results {
+		contextBuilder.WriteString(fmt.Sprintf("[Source %d: %s]\n", i+1, row.FileName))
+		contextBuilder.WriteString(row.RawText)
+		contextBuilder.WriteString("\n\n")
+	}
+
+	contextBuilder.WriteString("Instructions: ")
+	contextBuilder.WriteString("Based on the retrieved context above, provide a concise, coherent answer to the user's query. ")
+	contextBuilder.WriteString("Extract only the most relevant information. ")
+	contextBuilder.WriteString("If no relevant information is found, state that clearly. ")
+	contextBuilder.WriteString("Cite sources by filename when relevant. ")
+	contextBuilder.WriteString("Do not include unnecessary preamble or explanations.")
+
+	synthesisPrompt := contextBuilder.String()
+
+	emb, err := r.LineToVector(synthesisPrompt)
+	if err != nil {
+		r.logger.Error("failed to embed synthesis prompt", "error", err)
+		return "", err
+	}
+
+	embResp := &models.EmbeddingResp{
+		Embedding: emb,
+		Index:     0,
+	}
+
+	topResults, err := r.SearchEmb(embResp)
+	if err != nil {
+		r.logger.Error("failed to search for synthesis context", "error", err)
+		return "", err
+	}
+
+	if len(topResults) > 0 && topResults[0].RawText != synthesisPrompt {
+		return topResults[0].RawText, nil
+	}
+
+	var finalAnswer strings.Builder
+	finalAnswer.WriteString("Based on the retrieved context:\n\n")
+
+	for i, row := range results {
+		if i >= 5 {
+			break
+		}
+		finalAnswer.WriteString(fmt.Sprintf("- From %s: %s\n", row.FileName, truncateString(row.RawText, 200)))
+	}
+
+	return finalAnswer.String(), nil
+}
+
+func truncateString(s string, maxLen int) string {
+	if len(s) <= maxLen {
+		return s
+	}
+	return s[:maxLen] + "..."
+}
+
+func (r *RAG) Search(query string, limit int) ([]models.VectorRow, error) {
+	refined := r.RefineQuery(query)
+	variations := r.GenerateQueryVariations(refined)
+
+	allResults := make([]models.VectorRow, 0)
+	seen := make(map[string]bool)
+
+	for _, q := range variations {
+		emb, err := r.LineToVector(q)
+		if err != nil {
+			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)
+		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)
+			}
+		}
+	}
+
+	reranked := r.RerankResults(allResults, query)
+
+	if len(reranked) > limit {
+		reranked = reranked[:limit]
+	}
+
+	return reranked, nil
+}
+
+var (
+	ragInstance *RAG
+	ragOnce     sync.Once
+)
+
+func Init(c *config.Config, l *slog.Logger, s storage.FullRepo) error {
+	ragOnce.Do(func() {
+		if c == nil || l == nil || s == nil {
+			return
+		}
+		ragInstance = New(l, s, c)
+	})
+	return nil
+}
+
+func GetInstance() *RAG {
+	return ragInstance
+}

sysprompts/coding_assistant.json

@@ -1,5 +1,5 @@
 {
-  "sys_prompt": "You are an expert software engineering assistant. Your goal is to help users with coding tasks, debugging, refactoring, and software development.\n\n## Core Principles\n1. **Security First**: Never expose secrets, keys, or credentials. Never commit sensitive data.\n2. **No Git Actions**: You can READ git info (status, log, diff) for context, but NEVER perform git actions (commit, add, push, checkout, reset, rm, etc.). Let the user handle all git operations.\n3. **Explore Before Execute**: Always understand the codebase structure before making changes.\n4. **Follow Conventions**: Match existing code style, patterns, and frameworks used in the project.\n5. **Be Concise**: Minimize output tokens while maintaining quality. Avoid unnecessary explanations.\n\n## Workflow for Complex Tasks\nFor multi-step tasks, ALWAYS use the todo system to track progress:\n\n1. **Create Todo List**: At the start of complex tasks, use `todo_create` to break down work into actionable items.\n2. **Update Progress**: Mark items as `in_progress` when working on them, and `completed` when done.\n3. **Check Status**: Use `todo_read` to review your progress.\n\nExample workflow:\n- User: \"Add user authentication to this app\"\n- You: Create todos: [\"Analyze existing auth structure\", \"Check frameworks in use\", \"Implement auth middleware\", \"Add login endpoints\", \"Test implementation\"]\n\n## Task Execution Flow\n\n### Phase 1: Exploration (Always First)\n- Use `file_list` to understand directory structure (path defaults to CodingDir if not specified)\n- Use `file_read` to examine relevant files (paths are relative to CodingDir unless starting with `/`)\n- Use `execute_command` with `grep`/`find` to search for patterns\n- Check `README` or documentation files\n- Identify: frameworks, conventions, testing approach\n- **Git reads allowed**: You may use `git status`, `git log`, `git diff` for context, but only to inform your work\n- **Path handling**: Relative paths are resolved against CodingDir (configurable via Alt+O). Use absolute paths (starting with `/`) to bypass CodingDir.\n\n### Phase 2: Planning\n- For complex tasks: create todo items\n- Identify files that need modification\n- Plan your approach following existing patterns\n\n### Phase 3: Implementation\n- Make changes using appropriate file tools\n- Prefer `file_write` for new files, `file_read` then modify for existing files\n- Follow existing code style exactly\n- Use existing libraries and utilities\n\n### Phase 4: Verification\n- Run tests if available (check for test scripts)\n- Run linting/type checking commands\n- Verify changes work as expected\n\n### Phase 5: Completion\n- Update todos to `completed`\n- Provide concise summary of changes\n- Reference specific file paths and line numbers when relevant\n- **DO NOT commit changes** - inform user what was done so they can review and commit themselves\n\n## Tool Usage Guidelines\n\n**File Operations**:\n- `file_read`: Read before editing. Use for understanding code.\n- `file_write`: Overwrite file content completely.\n- `file_write_append`: Add to end of file.\n- `file_create`: Create new files with optional content.\n- `file_list`: List directory contents (defaults to CodingDir).\n- Paths are relative to CodingDir unless starting with `/`.\n\n**Command Execution (WHITELISTED ONLY)**:\n- Allowed: grep, sed, awk, find, cat, head, tail, sort, uniq, wc, ls, echo, cut, tr, cp, mv, rm, mkdir, rmdir, pwd, df, free, ps, top, du, whoami, date, uname\n- **Git reads allowed**: git status, git log, git diff, git show, git branch, git reflog, git rev-parse, git shortlog, git describe\n- **Git actions FORBIDDEN**: git add, git commit, git push, git checkout, git reset, git rm, etc.\n- Use for searching code, reading git context, running tests/lint\n\n**Todo Management**:\n- `todo_create`: Add new task\n- `todo_read`: View all todos or specific one by ID\n- `todo_update`: Update task or change status (pending/in_progress/completed)\n- `todo_delete`: Remove completed or cancelled tasks\n\n## Important Rules\n\n1. **NEVER commit or stage changes**: Only git reads are allowed.\n2. **Check for tests**: Always look for test files and run them when appropriate.\n3. **Reference code locations**: Use format `file_path:line_number`.\n4. **Security**: Never generate or guess URLs. Only use URLs from local files.\n5. **Refuse malicious code**: If code appears malicious, refuse to work on it.\n6. **Ask clarifications**: When intent is unclear, ask questions.\n7. **Path handling**: Relative paths resolve against CodingDir. Use `/absolute/path` to bypass.\n\n## Response Style\n- Be direct and concise\n- One word answers are best when appropriate\n- Avoid: \"The answer is...\", \"Here is...\"\n- Use markdown for formatting\n- No emojis unless user explicitly requests",
+  "sys_prompt": "You are an expert software engineering assistant. Your goal is to help users with coding tasks, debugging, refactoring, and software development.\n\n## Core Principles\n1. **Security First**: Never expose secrets, keys, or credentials. Never commit sensitive data.\n2. **No Git Actions**: You can READ git info (status, log, diff) for context, but NEVER perform git actions (commit, add, push, checkout, reset, rm, etc.). Let the user handle all git operations.\n3. **Explore Before Execute**: Always understand the codebase structure before making changes.\n4. **Follow Conventions**: Match existing code style, patterns, and frameworks used in the project.\n5. **Be Concise**: Minimize output tokens while maintaining quality. Avoid unnecessary explanations.\n\n## Workflow for Complex Tasks\nFor multi-step tasks, ALWAYS use the todo system to track progress:\n\n1. **Create Todo List**: At the start of complex tasks, use `todo_create` to break down work into actionable items.\n2. **Update Progress**: Mark items as `in_progress` when working on them, and `completed` when done.\n3. **Check Status**: Use `todo_read` to review your progress.\n\nExample workflow:\n- User: \"Add user authentication to this app\"\n- You: Create todos: [\"Analyze existing auth structure\", \"Check frameworks in use\", \"Implement auth middleware\", \"Add login endpoints\", \"Test implementation\"]\n\n## Task Execution Flow\n\n### Phase 1: Exploration (Always First)\n- Use `file_list` to understand directory structure (path defaults to FilePickerDir if not specified)\n- Use `file_read` to examine relevant files (paths are relative to FilePickerDir unless starting with `/`)\n- Use `execute_command` with `grep`/`find` to search for patterns\n- Check `README` or documentation files\n- Identify: frameworks, conventions, testing approach\n- **Git reads allowed**: You may use `git status`, `git log`, `git diff` for context, but only to inform your work\n- **Path handling**: Relative paths are resolved against FilePickerDir (configurable via Alt+O). Use absolute paths (starting with `/`) to bypass FilePickerDir.\n\n### Phase 2: Planning\n- For complex tasks: create todo items\n- Identify files that need modification\n- Plan your approach following existing patterns\n\n### Phase 3: Implementation\n- Make changes using appropriate file tools\n- Prefer `file_write` for new files, `file_read` then modify for existing files\n- Follow existing code style exactly\n- Use existing libraries and utilities\n\n### Phase 4: Verification\n- Run tests if available (check for test scripts)\n- Run linting/type checking commands\n- Verify changes work as expected\n\n### Phase 5: Completion\n- Update todos to `completed`\n- Provide concise summary of changes\n- Reference specific file paths and line numbers when relevant\n- **DO NOT commit changes** - inform user what was done so they can review and commit themselves\n\n## Tool Usage Guidelines\n\n**File Operations**:\n- `file_read`: Read before editing. Use for understanding code.\n- `file_write`: Overwrite file content completely.\n- `file_write_append`: Add to end of file.\n- `file_create`: Create new files with optional content.\n- `file_list`: List directory contents (defaults to FilePickerDir).\n- Paths are relative to FilePickerDir unless starting with `/`.\n\n**Command Execution (WHITELISTED ONLY)**:\n- Allowed: grep, sed, awk, find, cat, head, tail, sort, uniq, wc, ls, echo, cut, tr, cp, mv, rm, mkdir, rmdir, pwd, df, free, ps, top, du, whoami, date, uname\n- **Git reads allowed**: git status, git log, git diff, git show, git branch, git reflog, git rev-parse, git shortlog, git describe\n- **Git actions FORBIDDEN**: git add, git commit, git push, git checkout, git reset, git rm, etc.\n- Use for searching code, reading git context, running tests/lint\n\n**Todo Management**:\n- `todo_create`: Add new task\n- `todo_read`: View all todos or specific one by ID\n- `todo_update`: Update task or change status (pending/in_progress/completed)\n- `todo_delete`: Remove completed or cancelled tasks\n\n## Important Rules\n\n1. **NEVER commit or stage changes**: Only git reads are allowed.\n2. **Check for tests**: Always look for test files and run them when appropriate.\n3. **Reference code locations**: Use format `file_path:line_number`.\n4. **Security**: Never generate or guess URLs. Only use URLs from local files.\n5. **Refuse malicious code**: If code appears malicious, refuse to work on it.\n6. **Ask clarifications**: When intent is unclear, ask questions.\n7. **Path handling**: Relative paths resolve against FilePickerDir. Use `/absolute/path` to bypass.\n\n## Response Style\n- Be direct and concise\n- One word answers are best when appropriate\n- Avoid: \"The answer is...\", \"Here is...\"\n- Use markdown for formatting\n- No emojis unless user explicitly requests",
   "role": "CodingAssistant",
   "filepath": "sysprompts/coding_assistant.json",
   "first_msg": "Hello! I'm your coding assistant. I can help you with software engineering tasks like writing code, debugging, refactoring, and exploring codebases. I work best when you give me specific tasks, and for complex work, I'll create a todo list to track my progress. What would you like to work on?"

tables.go

@@ -236,9 +236,60 @@ func makeChatTable(chatMap map[string]models.Chat) *tview.Table {
 }
 
 // nolint:unused
-func makeRAGTable(fileList []string) *tview.Flex {
+func formatSize(size int64) string {
-	actions := []string{"load", "delete"}
+	units := []string{"B", "KB", "MB", "GB", "TB"}
-	rows, cols := len(fileList), len(actions)+1
+	i := 0
+	s := float64(size)
+	for s >= 1024 && i < len(units)-1 {
+		s /= 1024
+		i++
+	}
+	return fmt.Sprintf("%.1f%s", s, units[i])
+}
+
+type ragFileInfo struct {
+	name     string
+	inRAGDir bool
+	isLoaded bool
+	fullPath string
+}
+
+func makeRAGTable(fileList []string, loadedFiles []string) *tview.Flex {
+	// Build set of loaded files for quick lookup
+	loadedSet := make(map[string]bool)
+	for _, f := range loadedFiles {
+		loadedSet[f] = true
+	}
+
+	// Build merged list: files from ragdir + orphaned files from DB
+	ragFiles := make([]ragFileInfo, 0, len(fileList)+len(loadedFiles))
+	seen := make(map[string]bool)
+
+	// Add files from ragdir
+	for _, f := range fileList {
+		ragFiles = append(ragFiles, ragFileInfo{
+			name:     f,
+			inRAGDir: true,
+			isLoaded: loadedSet[f],
+			fullPath: path.Join(cfg.RAGDir, f),
+		})
+		seen[f] = true
+	}
+
+	// Add orphaned files (in DB but not in ragdir)
+	for _, f := range loadedFiles {
+		if !seen[f] {
+			ragFiles = append(ragFiles, ragFileInfo{
+				name:     f,
+				inRAGDir: false,
+				isLoaded: true,
+				fullPath: "",
+			})
+		}
+	}
+
+	rows := len(ragFiles)
+	cols := 4 // File Name | Preview | Action | Delete
 	fileTable := tview.NewTable().
 		SetBorders(true)
 	longStatusView := tview.NewTextView()
@@ -252,41 +303,92 @@ func makeRAGTable(fileList []string) *tview.Flex {
 		AddItem(fileTable, 0, 60, true)
 	// Add the exit option as the first row (row 0)
 	fileTable.SetCell(0, 0,
-		tview.NewTableCell("Exit RAG manager").
+		tview.NewTableCell("File Name").
 			SetTextColor(tcell.ColorWhite).
 			SetAlign(tview.AlignCenter).
 			SetSelectable(false))
 	fileTable.SetCell(0, 1,
-		tview.NewTableCell("(Close without action)").
+		tview.NewTableCell("Preview").
-			SetTextColor(tcell.ColorGray).
+			SetTextColor(tcell.ColorWhite).
 			SetAlign(tview.AlignCenter).
 			SetSelectable(false))
 	fileTable.SetCell(0, 2,
-		tview.NewTableCell("exit").
+		tview.NewTableCell("Load/Unload").
-			SetTextColor(tcell.ColorGray).
+			SetTextColor(tcell.ColorWhite).
-			SetAlign(tview.AlignCenter))
+			SetAlign(tview.AlignCenter).
+			SetSelectable(false))
+	fileTable.SetCell(0, 3,
+		tview.NewTableCell("Delete").
+			SetTextColor(tcell.ColorWhite).
+			SetAlign(tview.AlignCenter).
+			SetSelectable(false))
 	// Add the file rows starting from row 1
 	for r := 0; r < rows; r++ {
+		f := ragFiles[r]
 		for c := 0; c < cols; c++ {
 			color := tcell.ColorWhite
 			switch {
-			case c < 1:
+			case c == 0:
-				fileTable.SetCell(r+1, c, // +1 to account for the exit row at index 0
+				displayName := f.name
-					tview.NewTableCell(fileList[r]).
+				if !f.inRAGDir {
+					displayName = f.name + " (orphaned)"
+				}
+				fileTable.SetCell(r+1, c,
+					tview.NewTableCell(displayName).
 						SetTextColor(color).
 						SetAlign(tview.AlignCenter).
 						SetSelectable(false))
-			case c == 1: // Action description column - not selectable
+			case c == 1:
-				fileTable.SetCell(r+1, c, // +1 to account for the exit row at index 0
+				if !f.inRAGDir {
-					tview.NewTableCell("(Action)").
+					// Orphaned file - no preview available
-						SetTextColor(color).
+					fileTable.SetCell(r+1, c,
-						SetAlign(tview.AlignCenter).
+						tview.NewTableCell("not in ragdir").
-						SetSelectable(false))
+							SetTextColor(tcell.ColorYellow).
-			default: // Action button column - selectable
+							SetAlign(tview.AlignCenter).
-				fileTable.SetCell(r+1, c, // +1 to account for the exit row at index 0
+							SetSelectable(false))
-					tview.NewTableCell(actions[c-1]).
+				} else if fi, err := os.Stat(f.fullPath); err == nil {
+					size := fi.Size()
+					modTime := fi.ModTime()
+					preview := fmt.Sprintf("%s | %s", formatSize(size), modTime.Format("2006-01-02 15:04"))
+					fileTable.SetCell(r+1, c,
+						tview.NewTableCell(preview).
+							SetTextColor(color).
+							SetAlign(tview.AlignCenter).
+							SetSelectable(false))
+				} else {
+					fileTable.SetCell(r+1, c,
+						tview.NewTableCell("error").
+							SetTextColor(color).
+							SetAlign(tview.AlignCenter).
+							SetSelectable(false))
+				}
+			case c == 2:
+				actionText := "load"
+				if f.isLoaded {
+					actionText = "unload"
+				}
+				if !f.inRAGDir {
+					// Orphaned file - can only unload
+					actionText = "unload"
+				}
+				fileTable.SetCell(r+1, c,
+					tview.NewTableCell(actionText).
 						SetTextColor(color).
 						SetAlign(tview.AlignCenter))
+			case c == 3:
+				if !f.inRAGDir {
+					// Orphaned file - cannot delete from ragdir (not there)
+					fileTable.SetCell(r+1, c,
+						tview.NewTableCell("-").
+							SetTextColor(tcell.ColorDarkGray).
+							SetAlign(tview.AlignCenter).
+							SetSelectable(false))
+				} else {
+					fileTable.SetCell(r+1, c,
+						tview.NewTableCell("delete").
+							SetTextColor(color).
+							SetAlign(tview.AlignCenter))
+				}
 			}
 		}
 	}
@@ -318,7 +420,7 @@ func makeRAGTable(fileList []string) *tview.Flex {
 	}()
 	fileTable.Select(0, 0).
 		SetFixed(1, 1).
-		SetSelectable(true, false).
+		SetSelectable(true, true).
 		SetSelectedStyle(tcell.StyleDefault.Background(tcell.ColorGray).Foreground(tcell.ColorWhite)).
 		SetDoneFunc(func(key tcell.Key) {
 			if key == tcell.KeyEsc || key == tcell.KeyF1 || key == tcell.Key('x') || key == tcell.KeyCtrlX {
@@ -335,30 +437,58 @@ func makeRAGTable(fileList []string) *tview.Flex {
 		}
 		// defer pages.RemovePage(RAGPage)
 		tc := fileTable.GetCell(row, column)
+		tc.SetTextColor(tcell.ColorRed)
+		fileTable.SetSelectable(false, false)
 		// Check if the selected row is the exit row (row 0) - do this first to avoid index issues
 		if row == 0 {
 			pages.RemovePage(RAGPage)
 			return
 		}
-		// For file rows, get the filename (row index - 1 because of the exit row at index 0)
+		// For file rows, get the file info (row index - 1 because of the exit row at index 0)
-		fpath := fileList[row-1] // -1 to account for the exit row at index 0
+		f := ragFiles[row-1]
-		// notification := fmt.Sprintf("chat: %s; action: %s", fpath, tc.Text)
+		// Handle "-" case (orphaned file with no delete option)
+		if tc.Text == "-" {
+			pages.RemovePage(RAGPage)
+			return
+		}
 		switch tc.Text {
 		case "load":
-			fpath = path.Join(cfg.RAGDir, fpath)
+			fpath := path.Join(cfg.RAGDir, f.name)
 			longStatusView.SetText("clicked load")
 			go func() {
 				if err := ragger.LoadRAG(fpath); err != nil {
 					logger.Error("failed to embed file", "chat", fpath, "error", err)
 					_ = notifyUser("RAG", "failed to embed file; error: "+err.Error())
-					errCh <- err
+					app.QueueUpdate(func() {
-					// pages.RemovePage(RAGPage)
+						pages.RemovePage(RAGPage)
+					})
 					return
 				}
+				_ = notifyUser("RAG", "file loaded successfully")
+				app.QueueUpdate(func() {
+					pages.RemovePage(RAGPage)
+				})
+			}()
+			return
+		case "unload":
+			longStatusView.SetText("clicked unload")
+			go func() {
+				if err := ragger.RemoveFile(f.name); err != nil {
+					logger.Error("failed to unload file from RAG", "filename", f.name, "error", err)
+					_ = notifyUser("RAG", "failed to unload file; error: "+err.Error())
+					app.QueueUpdate(func() {
+						pages.RemovePage(RAGPage)
+					})
+					return
+				}
+				_ = notifyUser("RAG", "file unloaded successfully")
+				app.QueueUpdate(func() {
+					pages.RemovePage(RAGPage)
+				})
 			}()
 			return
 		case "delete":
-			fpath = path.Join(cfg.RAGDir, fpath)
+			fpath := path.Join(cfg.RAGDir, f.name)
 			if err := os.Remove(fpath); err != nil {
 				logger.Error("failed to delete file", "filename", fpath, "error", err)
 				return
@@ -385,7 +515,7 @@ func makeRAGTable(fileList []string) *tview.Flex {
 
 func makeLoadedRAGTable(fileList []string) *tview.Flex {
 	actions := []string{"delete"}
-	rows, cols := len(fileList), len(actions)+1
+	rows, cols := len(fileList), len(actions)+2
 	// Add 1 extra row for the "exit" option at the top
 	fileTable := tview.NewTable().
 		SetBorders(true)
@@ -400,39 +530,61 @@ func makeLoadedRAGTable(fileList []string) *tview.Flex {
 		AddItem(fileTable, 0, 60, true)
 	// Add the exit option as the first row (row 0)
 	fileTable.SetCell(0, 0,
-		tview.NewTableCell("Exit Loaded Files manager").
+		tview.NewTableCell("File Name").
 			SetTextColor(tcell.ColorWhite).
 			SetAlign(tview.AlignCenter).
 			SetSelectable(false))
 	fileTable.SetCell(0, 1,
-		tview.NewTableCell("(Close without action)").
+		tview.NewTableCell("Preview").
-			SetTextColor(tcell.ColorGray).
+			SetTextColor(tcell.ColorWhite).
 			SetAlign(tview.AlignCenter).
 			SetSelectable(false))
 	fileTable.SetCell(0, 2,
-		tview.NewTableCell("exit").
+		tview.NewTableCell("Load").
-			SetTextColor(tcell.ColorGray).
+			SetTextColor(tcell.ColorWhite).
-			SetAlign(tview.AlignCenter))
+			SetAlign(tview.AlignCenter).
+			SetSelectable(false))
+	fileTable.SetCell(0, 3,
+		tview.NewTableCell("Delete").
+			SetTextColor(tcell.ColorWhite).
+			SetAlign(tview.AlignCenter).
+			SetSelectable(false))
 	// Add the file rows starting from row 1
 	for r := 0; r < rows; r++ {
 		for c := 0; c < cols; c++ {
 			color := tcell.ColorWhite
 			switch {
-			case c < 1:
+			case c == 0:
-				fileTable.SetCell(r+1, c, // +1 to account for the exit row at index 0
+				fileTable.SetCell(r+1, c,
 					tview.NewTableCell(fileList[r]).
 						SetTextColor(color).
 						SetAlign(tview.AlignCenter).
 						SetSelectable(false))
-			case c == 1: // Action description column - not selectable
+			case c == 1:
-				fileTable.SetCell(r+1, c, // +1 to account for the exit row at index 0
+				if fi, err := os.Stat(fileList[r]); err == nil {
-					tview.NewTableCell("(Action)").
+					size := fi.Size()
+					modTime := fi.ModTime()
+					preview := fmt.Sprintf("%s | %s", formatSize(size), modTime.Format("2006-01-02 15:04"))
+					fileTable.SetCell(r+1, c,
+						tview.NewTableCell(preview).
+							SetTextColor(color).
+							SetAlign(tview.AlignCenter).
+							SetSelectable(false))
+				} else {
+					fileTable.SetCell(r+1, c,
+						tview.NewTableCell("error").
+							SetTextColor(color).
+							SetAlign(tview.AlignCenter).
+							SetSelectable(false))
+				}
+			case c == 2:
+				fileTable.SetCell(r+1, c,
+					tview.NewTableCell("load").
 						SetTextColor(color).
-						SetAlign(tview.AlignCenter).
+						SetAlign(tview.AlignCenter))
-						SetSelectable(false))
+			default:
-			default: // Action button column - selectable
+				fileTable.SetCell(r+1, c,
-				fileTable.SetCell(r+1, c, // +1 to account for the exit row at index 0
+					tview.NewTableCell("delete").
-					tview.NewTableCell(actions[c-1]).
 						SetTextColor(color).
 						SetAlign(tview.AlignCenter))
 			}
@@ -440,7 +592,7 @@ func makeLoadedRAGTable(fileList []string) *tview.Flex {
 	}
 	fileTable.Select(0, 0).
 		SetFixed(1, 1).
-		SetSelectable(true, false).
+		SetSelectable(true, true).
 		SetSelectedStyle(tcell.StyleDefault.Background(tcell.ColorGray).Foreground(tcell.ColorWhite)).
 		SetDoneFunc(func(key tcell.Key) {
 			if key == tcell.KeyEsc || key == tcell.KeyF1 || key == tcell.Key('x') || key == tcell.KeyCtrlX {
@@ -456,6 +608,8 @@ func makeLoadedRAGTable(fileList []string) *tview.Flex {
 			return
 		}
 		tc := fileTable.GetCell(row, column)
+		tc.SetTextColor(tcell.ColorRed)
+		fileTable.SetSelectable(false, false)
 		// Check if the selected row is the exit row (row 0) - do this first to avoid index issues
 		if row == 0 {
 			pages.RemovePage(RAGLoadedPage)
@@ -533,7 +687,7 @@ func makeAgentTable(agentList []string) *tview.Table {
 	}
 	chatActTable.Select(0, 0).
 		SetFixed(1, 1).
-		SetSelectable(true, false).
+		SetSelectable(true, true).
 		SetSelectedStyle(tcell.StyleDefault.Background(tcell.ColorGray).Foreground(tcell.ColorWhite)).
 		SetDoneFunc(func(key tcell.Key) {
 			if key == tcell.KeyEsc || key == tcell.KeyF1 || key == tcell.Key('x') {
@@ -549,6 +703,8 @@ func makeAgentTable(agentList []string) *tview.Table {
 			return
 		}
 		tc := chatActTable.GetCell(row, column)
+		tc.SetTextColor(tcell.ColorRed)
+		chatActTable.SetSelectable(false, false)
 		selected := agentList[row]
 		// notification := fmt.Sprintf("chat: %s; action: %s", selectedChat, tc.Text)
 		switch tc.Text {
@@ -630,7 +786,7 @@ func makeCodeBlockTable(codeBlocks []string) *tview.Table {
 	}
 	table.Select(0, 0).
 		SetFixed(1, 1).
-		SetSelectable(true, false).
+		SetSelectable(true, true).
 		SetSelectedStyle(tcell.StyleDefault.Background(tcell.ColorGray).Foreground(tcell.ColorWhite)).
 		SetDoneFunc(func(key tcell.Key) {
 			if key == tcell.KeyEsc || key == tcell.KeyF1 || key == tcell.Key('x') {
@@ -646,6 +802,8 @@ func makeCodeBlockTable(codeBlocks []string) *tview.Table {
 			return
 		}
 		tc := table.GetCell(row, column)
+		tc.SetTextColor(tcell.ColorRed)
+		table.SetSelectable(false, false)
 		selected := codeBlocks[row]
 		// notification := fmt.Sprintf("chat: %s; action: %s", selectedChat, tc.Text)
 		switch tc.Text {
@@ -702,7 +860,7 @@ func makeImportChatTable(filenames []string) *tview.Table {
 	}
 	chatActTable.Select(0, 0).
 		SetFixed(1, 1).
-		SetSelectable(true, false).
+		SetSelectable(true, true).
 		SetSelectedStyle(tcell.StyleDefault.Background(tcell.ColorGray).Foreground(tcell.ColorWhite)).
 		SetDoneFunc(func(key tcell.Key) {
 			if key == tcell.KeyEsc || key == tcell.KeyF1 || key == tcell.Key('x') {
@@ -718,6 +876,8 @@ func makeImportChatTable(filenames []string) *tview.Table {
 			return
 		}
 		tc := chatActTable.GetCell(row, column)
+		tc.SetTextColor(tcell.ColorRed)
+		chatActTable.SetSelectable(false, false)
 		selected := filenames[row]
 		// notification := fmt.Sprintf("chat: %s; action: %s", selectedChat, tc.Text)
 		switch tc.Text {
@@ -820,7 +980,9 @@ func makeFilePicker() *tview.Flex {
 	}
 	// Create UI elements
 	listView := tview.NewList()
-	listView.SetBorder(true).SetTitle("Files & Directories [c: set CodingDir]").SetTitleAlign(tview.AlignLeft)
+	listView.SetBorder(true).
+		SetTitle("Files & Directories [s: set FilePickerDir]. Current base dir: " + cfg.FilePickerDir).
+		SetTitleAlign(tview.AlignLeft)
 	// Status view for selected file information
 	statusView := tview.NewTextView()
 	statusView.SetBorder(true).SetTitle("Selected File").SetTitleAlign(tview.AlignLeft)
@@ -1032,8 +1194,8 @@ func makeFilePicker() *tview.Flex {
 				refreshList(currentDisplayDir, "")
 				return nil
 			}
-			if event.Rune() == 'c' {
+			if event.Rune() == 's' {
-				// Set CodingDir to current directory
+				// Set FilePickerDir to current directory
 				itemIndex := listView.GetCurrentItem()
 				if itemIndex >= 0 && itemIndex < listView.GetItemCount() {
 					itemText, _ := listView.GetItemText(itemIndex)
@@ -1056,11 +1218,11 @@ func makeFilePicker() *tview.Flex {
 							targetDir = currentDisplayDir
 						}
 					}
-					cfg.CodingDir = targetDir
+					cfg.FilePickerDir = targetDir
-					if err := notifyUser("CodingDir", "Set to: "+targetDir); err != nil {
+					if err := notifyUser("FilePickerDir", "Set to: "+targetDir); err != nil {
 						logger.Error("failed to notify user", "error", err)
 					}
-					pages.RemovePage(filePickerPage)
+					// pages.RemovePage(filePickerPage)
 					return nil
 				}
 			}

tools.go

@@ -16,6 +16,7 @@ import (
 	"sync"
 	"time"
 
+	"gf-lt/rag"
 	"github.com/GrailFinder/searchagent/searcher"
 )
 
@@ -58,9 +59,9 @@ Your current tools:
 "when_to_use": "when asked to search the web for information; returns clean summary without html,css and other web elements; limit is optional (default 3)"
 },
 {
-"name":"websearch_raw",
+"name":"rag_search",
 "args": ["query", "limit"],
-"when_to_use": "when asked to search the web for information; returns raw data as is without processing; limit is optional (default 3)"
+"when_to_use": "when asked to search the local document database for information; performs query refinement, semantic search, reranking, and synthesis; returns clean summary with sources; limit is optional (default 3)"
 },
 {
 "name":"read_url",
@@ -146,6 +147,7 @@ under the topic: Adam's number is stored:
 After that you are free to respond to the user.
 `
 	webSearchSysPrompt = `Summarize the web search results, extracting key information and presenting a concise answer. Provide sources and URLs where relevant.`
+	ragSearchSysPrompt = `Synthesize the document search results, extracting key information and presenting a concise answer. Provide sources and document IDs where relevant.`
 	readURLSysPrompt   = `Extract and summarize the content from the webpage. Provide key information, main points, and any relevant details.`
 	summarySysPrompt   = `Please provide a concise summary of the following conversation. Focus on key points, decisions, and actions. Provide only the summary, no additional commentary.`
 	basicCard          = &models.CharCard{
@@ -170,6 +172,10 @@ func init() {
 		panic("failed to init seachagent; error: " + err.Error())
 	}
 	WebSearcher = sa
+
+	if err := rag.Init(cfg, logger, store); err != nil {
+		logger.Warn("failed to init rag; rag_search tool will not be available", "error", err)
+	}
 }
 
 // getWebAgentClient returns a singleton AgentClient for web agents.
@@ -196,6 +202,8 @@ func getWebAgentClient() *agent.AgentClient {
 func registerWebAgents() {
 	webAgentsOnce.Do(func() {
 		client := getWebAgentClient()
+		// Register rag_search agent
+		agent.Register("rag_search", agent.NewWebAgentB(client, ragSearchSysPrompt))
 		// Register websearch agent
 		agent.Register("websearch", agent.NewWebAgentB(client, webSearchSysPrompt))
 		// Register read_url agent
@@ -239,6 +247,48 @@ func websearch(args map[string]string) []byte {
 	return data
 }
 
+// rag search (searches local document database)
+func ragsearch(args map[string]string) []byte {
+	query, ok := args["query"]
+	if !ok || query == "" {
+		msg := "query not provided to rag_search tool"
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	limitS, ok := args["limit"]
+	if !ok || limitS == "" {
+		limitS = "3"
+	}
+	limit, err := strconv.Atoi(limitS)
+	if err != nil || limit == 0 {
+		logger.Warn("ragsearch limit; passed bad value; setting to default (3)",
+			"limit_arg", limitS, "error", err)
+		limit = 3
+	}
+
+	ragInstance := rag.GetInstance()
+	if ragInstance == nil {
+		msg := "rag not initialized; rag_search tool is not available"
+		logger.Error(msg)
+		return []byte(msg)
+	}
+
+	results, err := ragInstance.Search(query, limit)
+	if err != nil {
+		msg := "rag search failed; error: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+
+	data, err := json.Marshal(results)
+	if err != nil {
+		msg := "failed to marshal rag search result; error: " + err.Error()
+		logger.Error(msg)
+		return []byte(msg)
+	}
+	return data
+}
+
 // web search raw (returns raw data without processing)
 func websearchRaw(args map[string]string) []byte {
 	// make http request return bytes
@@ -577,7 +627,7 @@ func resolvePath(p string) string {
 	if filepath.IsAbs(p) {
 		return p
 	}
-	return filepath.Join(cfg.CodingDir, p)
+	return filepath.Join(cfg.FilePickerDir, p)
 }
 
 func readStringFromFile(filename string) (string, error) {
@@ -997,6 +1047,7 @@ var fnMap = map[string]fnSig{
 	"recall":            recall,
 	"recall_topics":     recallTopics,
 	"memorise":          memorise,
+	"rag_search":        ragsearch,
 	"websearch":         websearch,
 	"websearch_raw":     websearchRaw,
 	"read_url":          readURL,
@@ -1033,6 +1084,28 @@ func callToolWithAgent(name string, args map[string]string) []byte {
 
 // openai style def
 var baseTools = []models.Tool{
+	// rag_search
+	models.Tool{
+		Type: "function",
+		Function: models.ToolFunc{
+			Name:        "rag_search",
+			Description: "Search local document database given query, limit of sources (default 3). Performs query refinement, semantic search, reranking, and synthesis.",
+			Parameters: models.ToolFuncParams{
+				Type:     "object",
+				Required: []string{"query", "limit"},
+				Properties: map[string]models.ToolArgProps{
+					"query": models.ToolArgProps{
+						Type:        "string",
+						Description: "search query",
+					},
+					"limit": models.ToolArgProps{
+						Type:        "string",
+						Description: "limit of the document results",
+					},
+				},
+			},
+		},
+	},
 	// websearch
 	models.Tool{
 		Type: "function",

tui.go

@@ -15,6 +15,11 @@ import (
 	"github.com/rivo/tview"
 )
 
+func isFullScreenPageActive() bool {
+	name, _ := pages.GetFrontPage()
+	return name != "main"
+}
+
 var (
 	app                *tview.Application
 	pages              *tview.Pages
@@ -106,7 +111,7 @@ var (
 [yellow]x[white]: to exit the table page
 
 === filepicker ===
-[yellow]c[white]: (in file picker) set current dir as CodingDir
+[yellow]s[white]: (in file picker) set current dir as FilePickerDir
 [yellow]x[white]: to exit
 
 === shell mode ===
@@ -525,6 +530,9 @@ func init() {
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Rune() == 'i' && event.Modifiers()&tcell.ModAlt != 0 {
+			if isFullScreenPageActive() {
+				return event
+			}
 			showColorschemeSelectionPopup()
 			return nil
 		}
@@ -731,6 +739,9 @@ func init() {
 			return nil
 		}
 		if event.Key() == tcell.KeyCtrlL {
+			if isFullScreenPageActive() {
+				return event
+			}
 			// Show model selection popup instead of rotating models
 			showModelSelectionPopup()
 			return nil
@@ -744,6 +755,9 @@ func init() {
 			return nil
 		}
 		if event.Key() == tcell.KeyCtrlV {
+			if isFullScreenPageActive() {
+				return event
+			}
 			// Show API link selection popup instead of rotating APIs
 			showAPILinkSelectionPopup()
 			return nil
@@ -850,11 +864,17 @@ func init() {
 			return nil
 		}
 		if event.Key() == tcell.KeyCtrlQ {
+			if isFullScreenPageActive() {
+				return event
+			}
 			// Show user role selection popup instead of cycling through roles
 			showUserRoleSelectionPopup()
 			return nil
 		}
 		if event.Key() == tcell.KeyCtrlX {
+			if isFullScreenPageActive() {
+				return event
+			}
 			// Show bot role selection popup instead of cycling through roles
 			showBotRoleSelectionPopup()
 			return nil
@@ -893,6 +913,7 @@ func init() {
 					return nil
 				}
 			}
+			// Get files from ragdir
 			fileList := []string{}
 			for _, f := range files {
 				if f.IsDir() {
@@ -900,22 +921,14 @@ func init() {
 				}
 				fileList = append(fileList, f.Name())
 			}
-			chatRAGTable := makeRAGTable(fileList)
+			// Get loaded files from vector DB
-			pages.AddPage(RAGPage, chatRAGTable, true, true)
+			loadedFiles, err := ragger.ListLoaded()
-			return nil
-		}
-		if event.Key() == tcell.KeyCtrlY { // Use Ctrl+Y to list loaded RAG files
-			// List files already loaded into the RAG system
-			fileList, err := ragger.ListLoaded()
 			if err != nil {
 				logger.Error("failed to list loaded RAG files", "error", err)
-				if notifyerr := notifyUser("failed to list RAG files", err.Error()); notifyerr != nil {
+				loadedFiles = []string{} // Continue with empty list on error
-					logger.Error("failed to send notification", "error", notifyerr)
-				}
-				return nil
 			}
-			chatLoadedRAGTable := makeLoadedRAGTable(fileList)
+			chatRAGTable := makeRAGTable(fileList, loadedFiles)
-			pages.AddPage(RAGLoadedPage, chatLoadedRAGTable, true, true)
+			pages.AddPage(RAGPage, chatRAGTable, true, true)
 			return nil
 		}
 		if event.Key() == tcell.KeyRune && event.Modifiers() == tcell.ModAlt && event.Rune() == '1' {
@@ -975,13 +988,6 @@ func init() {
 				}
 				// go chatRound(msgText, persona, textView, false, false)
 				chatRoundChan <- &models.ChatRoundReq{Role: persona, UserMsg: msgText}
-				// Also clear any image attachment after sending the message
-				go func() {
-					// Wait a short moment for the message to be processed, then clear the image attachment
-					// This allows the image to be sent with the current message if it was attached
-					// But clears it for the next message
-					ClearImageAttachment()
-				}()
 			}
 			return nil
 		}