Merge pull request #20: fix: add /api/chat endpoint for streaming (closes #10)

models/api_client.py

@@ -169,15 +169,18 @@ class APIClient:
         self, messages: list, model: str = None, **kwargs
     ) -> Dict[str, Any]:
         """
-        Get a completion from the LLM using Ollama chat endpoint
+        Get a completion from the LLM using Ollama chat endpoint.
+
+        Uses /api/chat with proper message list format to preserve
+        multi-turn conversation semantics (system, user, assistant roles).
 
         Args:
-            messages (list): List of message dictionaries (roles and content)
+            messages (list): List of message dictionaries with 'role' and 'content'
             model (str): Model to use
             **kwargs: Additional parameters for the API
 
         Returns:
-            dict: Response from the LLM
+            dict: Response from the LLM in OpenAI-compatible format
         """
         if model is None:
             # Reload config to get latest model setting
@@ -186,40 +189,53 @@ class APIClient:
             current_config = load_config()
             model = current_config.get("model", "qwen3-coder:30b")
 
-        # Convert messages to prompt format expected by Ollama generate endpoint
+        # Filter messages to only include valid roles for chat endpoint
-        prompt_text = ""
+        chat_messages = []
         for message in messages:
             role = message.get("role", "user")
             content = message.get("content", "")
+            if role in ("system", "user", "assistant"):
+                chat_messages.append({"role": role, "content": content})
 
-            # Format messages properly for the model
+        if not chat_messages:
-            if role == "system":
+            return {"error": "No valid messages provided for chat completion"}
-                prompt_text += f"System: {content}\n\n"
-            elif role == "assistant":
-                prompt_text += f"Assistant: {content}\n\n"
-            else:  # user
-                prompt_text += f"User: {content}\n\n"
 
-        # Add instruction for assistant response
+        # Use Ollama chat endpoint with proper message format
-        prompt_text += "Assistant:"
+        data = {
+            "model": model,
+            "messages": chat_messages,
+            "stream": False,
+            **kwargs,
+        }
 
-        # Prepare the data for Ollama generate endpoint
+        result = self._make_request("/api/chat", "POST", data)
-        data = {"model": model, "prompt": prompt_text, "stream": False, **kwargs}
-
-        result = self._make_request("/api/generate", "POST", data)
         if not result["success"]:
             return {"error": result["error"]}
 
-        # Extract the response from Ollama's generate format
+        # Extract the response from Ollama's chat format
         try:
             response_data = result["data"]
-            # In Ollama generate responses, the actual text is in the "response" field
+            # Ollama chat endpoint returns message in message.content
-            if "response" in response_data:
+            if "message" in response_data:
+                return {
+                    "choices": [
+                        {
+                            "message": {
+                                "role": response_data["message"].get(
+                                    "role", "assistant"
+                                ),
+                                "content": response_data["message"].get(
+                                    "content", ""
+                                ),
+                            }
+                        }
+                    ]
+                }
+            elif "response" in response_data:
                 return {
                     "choices": [{"message": {"content": response_data["response"]}}]
                 }
             else:
-                # If we get a different format, return what we found
                 return response_data
         except Exception as e:
             return {"error": f"Failed to process chat completion result: {str(e)}"}