Implement proxy for Ollama chat API

README.md

@@ -145,10 +145,54 @@ $ curl http://127.0.0.1:5555/api/embed \
 {"model":"text-embedding-004","embeddings":[[0.04824496,0.0117766075,-0.011552069,-0.018164534,-0.0026110192,0.05092675,0.08172899,0.007869772,0.054475933,0.026131334,-0.06593486,-0.002256868,0.038781915,...]]}
 ```
 
+Create chat completions:
+
+```sh
+$ curl http://127.0.0.1:5555/api/chat \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "gemini-1.5-pro",
+    "messages": [
+      { "role": "user", "content": "Hello, world!" }
+    ]
+  }'
+```
+
+Chat response example:
+
+```json
+{
+  "model": "gemini-1.5-pro",
+  "created_at": "2024-07-28T15:00:00Z",
+  "message": { "role": "assistant", "content": "Hello back to you!" },
+  "done": true,
+  "total_duration": 123456789,
+  "prompt_eval_count": 5,
+  "eval_count": 10
+}
+```
+
+Advanced usage:
+
+- Use the `format` parameter to request JSON or enforce a JSON schema.
+- Use the `options` parameter to set model generation parameters: `temperature`, `num_predict`, `top_k`, and `top_p`.
+
+```sh
+$ curl http://127.0.0.1:5555/api/chat \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "gemini-1.5-pro",
+    "messages": [...],
+    "format": "json",
+    "options": { "temperature": 0.5, "num_predict": 100 }
+  }'
+```
+
 ### Known Ollama Limitations
 * Streaming is not yet supported.
 * Images are not supported.
-* Response format is not supported.
+* Response format is supported only for chat API.
+* Tools are not supported for chat API.
 * Model parameters not supported by Gemini are ignored.
 
 ## Notes

ollama/ollama.go

@@ -18,16 +18,17 @@ package ollama
 
 import (
 	"encoding/json"
+	"fmt"
+	"github.com/google-gemini/proxy-to-gemini/internal"
+	"github.com/google/generative-ai-go/genai"
+	"github.com/gorilla/mux"
 	"io"
 	"net/http"
 	"strings"
 	"time"
-
-	"github.com/google-gemini/proxy-to-gemini/internal"
-	"github.com/google/generative-ai-go/genai"
-	"github.com/gorilla/mux"
 )
 
+// handlers provides HTTP handlers for the Ollama proxy API.
 type handlers struct {
 	client *genai.Client
 }
@@ -36,6 +37,7 @@ func RegisterHandlers(r *mux.Router, client *genai.Client) {
 	handlers := &handlers{client: client}
 	r.HandleFunc("/api/generate", handlers.generateHandler)
 	r.HandleFunc("/api/embed", handlers.embedHandler)
+	r.HandleFunc("/api/chat", handlers.chatHandler)
 }
 
 func (h *handlers) generateHandler(w http.ResponseWriter, r *http.Request) {
@@ -142,6 +144,193 @@ func (h *handlers) embedHandler(w http.ResponseWriter, r *http.Request) {
 	}
 }
 
+// ChatRequest represents a chat completion request for the Ollama API.
+type ChatRequest struct {
+	Model    string          `json:"model,omitempty"`
+	Messages []ChatMessage   `json:"messages,omitempty"`
+	Format   json.RawMessage `json:"format,omitempty"`
+	Options  Options         `json:"options,omitempty"`
+}
+
+// ChatMessage represents a single message in a chat.
+type ChatMessage struct {
+	Role    string   `json:"role,omitempty"`
+	Content string   `json:"content,omitempty"`
+	Images  []string `json:"images,omitempty"`
+}
+
+// ChatResponse represents a chat completion response for the Ollama API.
+type ChatResponse struct {
+	Model              string      `json:"model,omitempty"`
+	CreatedAt          time.Time   `json:"created_at,omitempty"`
+	Message            ChatMessage `json:"message,omitempty"`
+	Done               bool        `json:"done,omitempty"`
+	PromptEvalCount    int32       `json:"prompt_eval_count"`
+	EvalCount          int32       `json:"eval_count"`
+	TotalDuration      int64       `json:"total_duration"`
+	LoadDuration       int64       `json:"load_duration,omitempty"`
+	PromptEvalDuration int64       `json:"prompt_eval_duration,omitempty"`
+	EvalDuration       int64       `json:"eval_duration"`
+}
+
+func sanitizeJson(s string) string {
+	s = strings.ReplaceAll(s, "\n", "")
+	s = strings.TrimPrefix(s, "```json")
+	s = strings.TrimSuffix(s, "```")
+	s = strings.ReplaceAll(s, "'", "\\'")
+	return s
+}
+
+// chatHandler handles POST /api/chat requests.
+func (h *handlers) chatHandler(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodPost {
+		internal.ErrorHandler(w, r, http.StatusMethodNotAllowed, "method not allowed")
+		return
+	}
+	body, err := io.ReadAll(r.Body)
+	if err != nil {
+		internal.ErrorHandler(w, r, http.StatusInternalServerError, "failed to read request body: %v", err)
+		return
+	}
+	defer r.Body.Close()
+
+	var req ChatRequest
+	if err := json.Unmarshal(body, &req); err != nil {
+		internal.ErrorHandler(w, r, http.StatusInternalServerError, "failed to unmarshal chat request: %v", err)
+		return
+	}
+
+	// Handle advanced format parameter: JSON mode or JSON schema enforcement
+	expectJson := false
+	if len(req.Format) > 0 {
+		expectJson = true
+		var formatVal interface{}
+		if err := json.Unmarshal(req.Format, &formatVal); err != nil {
+			internal.ErrorHandler(w, r, http.StatusBadRequest, "invalid format parameter: %v", err)
+			return
+		}
+		var instr string
+		switch v := formatVal.(type) {
+		case string:
+			if v == "json" {
+				instr = "Please respond with valid JSON."
+			} else {
+				instr = fmt.Sprintf("Please respond with format: %s.", v)
+			}
+		default:
+			schemaBytes, err := json.MarshalIndent(v, "", "  ")
+			if err != nil {
+				schemaBytes = req.Format
+			}
+			instr = fmt.Sprintf("Please format your response according to the following JSON schema:\n%s", string(schemaBytes))
+		}
+		// Integrate with existing system message if present
+		found := false
+		for i, m := range req.Messages {
+			if m.Role == "system" {
+				req.Messages[i].Content = m.Content + "\n\n" + instr
+				found = true
+				break
+			}
+		}
+		if !found {
+			req.Messages = append([]ChatMessage{{Role: "system", Content: instr}}, req.Messages...)
+		}
+	}
+
+	model := h.client.GenerativeModel(req.Model)
+	model.GenerationConfig = genai.GenerationConfig{
+		Temperature:     req.Options.Temperature,
+		MaxOutputTokens: req.Options.NumPredict,
+		TopK:            req.Options.TopK,
+		TopP:            req.Options.TopP,
+	}
+	if req.Options.Stop != nil {
+		model.GenerationConfig.StopSequences = []string{*req.Options.Stop}
+	}
+
+	chat := model.StartChat()
+	var lastPart genai.Part
+	for i, m := range req.Messages {
+		if m.Role == "system" {
+			model.SystemInstruction = &genai.Content{
+				Role:  m.Role,
+				Parts: []genai.Part{genai.Text(m.Content)},
+			}
+			continue
+		}
+		if i == len(req.Messages)-1 {
+			lastPart = genai.Text(m.Content)
+			break
+		}
+		chat.History = append(chat.History, &genai.Content{
+			Role:  m.Role,
+			Parts: []genai.Part{genai.Text(m.Content)},
+		})
+	}
+
+	// Measure time spent generating the chat response
+	start := time.Now()
+
+	gresp, err := chat.SendMessage(r.Context(), lastPart)
+	if err != nil {
+		internal.ErrorHandler(w, r, http.StatusInternalServerError, "failed to send chat message: %v", err)
+		return
+	}
+	var builder strings.Builder
+	if len(gresp.Candidates) > 0 {
+		for _, part := range gresp.Candidates[0].Content.Parts {
+			if txt, ok := part.(genai.Text); ok {
+				builder.WriteString(string(txt))
+			}
+		}
+	}
+
+	var resp ChatResponse
+	if expectJson {
+		resp = ChatResponse{
+			Model:     req.Model,
+			CreatedAt: time.Now(),
+			Message: ChatMessage{
+				Role:    gresp.Candidates[0].Content.Role,
+				Content: sanitizeJson(builder.String()),
+			},
+			Done: true,
+		}
+	} else {
+		resp = ChatResponse{
+			Model:     req.Model,
+			CreatedAt: time.Now(),
+			Message: ChatMessage{
+				Role:    gresp.Candidates[0].Content.Role,
+				Content: builder.String(),
+			},
+			Done: true,
+		}
+	}
+
+	if gresp.UsageMetadata != nil {
+		resp.PromptEvalCount = gresp.UsageMetadata.PromptTokenCount
+		// Compute number of tokens in the response.
+		if gresp.UsageMetadata.CandidatesTokenCount > 0 {
+			resp.EvalCount = gresp.UsageMetadata.CandidatesTokenCount
+		} else if gresp.UsageMetadata.TotalTokenCount >= gresp.UsageMetadata.PromptTokenCount {
+			// Fallback: use total tokens minus prompt tokens
+			resp.EvalCount = gresp.UsageMetadata.TotalTokenCount - gresp.UsageMetadata.PromptTokenCount
+		}
+	}
+	// Populate duration metadata (in nanoseconds)
+	elapsed := time.Since(start).Nanoseconds()
+	resp.TotalDuration = elapsed
+	resp.LoadDuration = 0
+	resp.PromptEvalDuration = 0
+	resp.EvalDuration = elapsed
+	if err := json.NewEncoder(w).Encode(&resp); err != nil {
+		internal.ErrorHandler(w, r, http.StatusInternalServerError, "failed to encode chat response: %v", err)
+		return
+	}
+}
+
 type GenerateRequest struct {
 	Model   string  `json:"model,omitempty"`
 	Prompt  string  `json:"prompt,omitempty"`