More work towards rawdogging ollama

Author: radare

src/repl/conf.go

@@ -290,25 +290,27 @@ func (r *REPL) resolvePromptPath(promptName string) (string, error) {
 func (r *REPL) handleSetCommand(args []string) error {
 	// Special handling for specific configOptions
 	if len(args) >= 3 {
+		val := strings.Join(args[2:], " ")
 		switch args[1] {
+		case "deterministic":
+			b, _ := strconv.ParseBool(val)
+			r.config.Deterministic = b
+			return nil
 		case "rawdog":
-			fmt.Println("RAWDOG")
-			r.config.Rawdog = true
+			b, _ := strconv.ParseBool(val)
+			r.config.Rawdog = b
 			return nil
 		case "promptfile":
-			filePath := args[2]
-			return r.loadSystemPrompt(filePath)
+			return r.loadSystemPrompt(val)
 		case "systemprompt":
-			promptText := strings.Join(args[2:], " ")
-			r.systemPrompt = promptText
-			r.configOptions.Set("systemprompt", promptText)
-			fmt.Printf("System prompt set (%d chars)\r\n", len(promptText))
+			r.systemPrompt = val
+			r.configOptions.Set("systemprompt", val)
+			fmt.Printf("System prompt set (%d chars)\r\n", len(val))
 			return nil
 		case "model":
-			model := strings.Join(args[2:], " ")
-			return r.setModel(model)
+			return r.setModel(val)
 		case "provider":
-			provider := strings.ToLower(args[2])
+			provider := strings.ToLower(val)
 			return r.setProvider(provider)
 		}
 	}

src/repl/llm/llm.go

@@ -3,6 +3,7 @@ package llm
 import (
 	"bytes"
 	"context"
+	"encoding/json"
 	"fmt"
 	"io"
 	"net/http"
@@ -362,4 +363,14 @@ func buildURL(defaultURL, baseURL, host, port, suffix string) string {
 	return fmt.Sprintf("http://%s:%s%s", host, port, suffix)
 }
 
-// ==================== PROVIDER IMPLEMENTATIONS ====================
+func MarshalNoEscape(v any) ([]byte, error) {
+	var buf bytes.Buffer
+	enc := json.NewEncoder(&buf)
+	enc.SetEscapeHTML(false)
+	err := enc.Encode(v)
+	if err != nil {
+		return nil, err
+	}
+	// Remove trailing newline added by Encoder
+	return bytes.TrimRight(buf.Bytes(), "\n"), nil
+}

src/repl/llm/ollama.go

@@ -49,7 +49,7 @@ func (c *LLMClient) sendOllamaWithImages(ctx context.Context, messages []Message
 		}
 	}
 
-	jsonData, err := json.Marshal(request)
+	jsonData, err := MarshalNoEscape(request)
 	if err != nil {
 		return "", err
 	}
@@ -168,7 +168,6 @@ func (p *OllamaProvider) ListModels(ctx context.Context) ([]Model, error) {
 
 func (p *OllamaProvider) SendMessage(ctx context.Context, messages []Message, stream bool) (string, error) {
 	if p.config.Rawdog {
-		stream = false
 		messageline := "" // <start_of_turn>user\nhello world<end_of_turn>\n<start_of_turn>model\n"
 		for _, msg := range messages {
 			messageline += msg.Content.(string)
@@ -184,17 +183,17 @@ func (p *OllamaProvider) SendMessage(ctx context.Context, messages []Message, st
 			Prompt: messageline,
 		}
 		// Apply deterministic settings if enabled
-		if p.config.Deterministic {
+		if p.config.Deterministic || true {
 			request.Options = map[string]float64{
 				"repeat_last_n":  0,
 				"top_p":          0.0,
 				"top_k":          1.0,
 				"temperature":    0.0,
 				"repeat_penalty": 1.0,
-				"seed":           123,
+				"seed":           0,
 			}
 		}
-		jsonData, err := json.Marshal(request)
+		jsonData, err := MarshalNoEscape(request)
 		if err != nil {
 			return "", err
 		}
@@ -211,7 +210,9 @@ func (p *OllamaProvider) SendMessage(ctx context.Context, messages []Message, st
 		}
 
 		// stream-mode
-		// if stream { return llmMakeStreamingRequest(ctx, "POST", url, headers, jsonData, p.parseStream) }
+		if stream {
+			return llmMakeStreamingRequest(ctx, "POST", url, headers, jsonData, p.parseStream)
+		}
 
 		// non-stream
 		respBody, err := llmMakeRequest(ctx, "POST", url, headers, jsonData)
@@ -258,7 +259,7 @@ func (p *OllamaProvider) SendMessage(ctx context.Context, messages []Message, st
 		}
 	}
 
-	jsonData, err := json.Marshal(request)
+	jsonData, err := MarshalNoEscape(request)
 	if err != nil {
 		return "", err
 	}
@@ -319,19 +320,37 @@ func (p *OllamaProvider) parseStream(reader io.Reader) (string, error) {
 			continue
 		}
 
-		var response struct {
-			Message struct {
-				Content string `json:"content""`
-			} `json:"message""`
-			Done bool `json:"done""`
-		}
+		isDone := false
+		content := ""
+		if p.config.Rawdog {
+			var response struct {
+				Response string `json:"response""`
+				Done     bool   `json:"done""`
+			}
 
-		if err := json.Unmarshal([]byte(line), &response); err != nil {
-			continue
-		}
+			if err := json.Unmarshal([]byte(line), &response); err != nil {
+				continue
+			}
 
-		// Format content based on markdown setting
-		content := response.Message.Content
+			// Format content based on markdown setting
+			content = response.Response
+			isDone = response.Done
+		} else {
+			var response struct {
+				Message struct {
+					Content string `json:"content""`
+				} `json:"message""`
+				Done bool `json:"done""`
+			}
+
+			if err := json.Unmarshal([]byte(line), &response); err != nil {
+				continue
+			}
+
+			// Format content based on markdown setting
+			content = response.Message.Content
+			isDone = response.Done
+		}
 		if !markdownEnabled {
 			// Standard formatting - just replace newlines for terminal display
 			content = strings.ReplaceAll(content, "\n", "\n\r")
@@ -340,9 +359,9 @@ func (p *OllamaProvider) parseStream(reader io.Reader) (string, error) {
 			content = FormatStreamingChunk(content, markdownEnabled)
 		}
 		fmt.Print(content)
-		fullResponse.WriteString(response.Message.Content)
+		fullResponse.WriteString(content) // response.Message.Content)
 
-		if response.Done {
+		if isDone {
 			break
 		}
 	}

src/repl/repl.go

@@ -1741,7 +1741,7 @@ func (r *REPL) sendToAI(input string) error {
 	}
 
 	// If reasoning is disabled, append /no_think to the last message sent to the LLM
-	if !r.reasoningEnabled {
+	if !r.reasoningEnabled && r.configOptions.GetBool("rawdog") {
 		// Create a copy of the messages for the API call with /no_think appended
 		messagesCopy := make([]llm.Message, len(messages))
 		copy(messagesCopy, messages)