Add LLM token

backend/app/services/llm/qwen_client.py

@@ -1,4 +1,8 @@
-"""Provide service-layer logic for qwen client."""
+"""Provide service-layer logic for qwen client.
+
+P0-0: ``chat()`` now accepts an optional ``tools`` list and parses ``tool_calls``
+from the model response so that callers can dispatch tool invocations.
+"""
 
 import time
 import json
@@ -7,6 +11,7 @@ from loguru import logger
 import httpx
 
 from .base_client import BaseLLMClient, LLMResponse, LLMConfig, LLMProvider
+from .tool_types import Tool, ToolCall
 # Keep provider-specific behavior explicit so debugging stays straightforward.
 
 
@@ -54,14 +59,20 @@ class QwenClient(BaseLLMClient):
         messages: List[Dict[str, str]],
         max_tokens: Optional[int] = None,
         temperature: Optional[float] = None,
+        tools: Optional[List[Tool]] = None,
         **kwargs
     ) -> LLMResponse:
-        """Handle chat for the Qwen Client instance."""
+        """Handle chat for the Qwen Client instance.
+
+        When ``tools`` is provided the request includes the tool definitions and
+        ``tool_choice="auto"``; any tool_calls returned by the model are parsed
+        into ``LLMResponse.tool_calls``.
+        """
         start_time = time.time()
 
         try:
             # Keep provider-specific behavior explicit so debugging stays straightforward.
-            payload = {
+            payload: Dict = {
                 "model": self.config.model,
                 "messages": messages,
                 "max_tokens": max_tokens or self.config.max_tokens,
@@ -70,6 +81,11 @@ class QwenClient(BaseLLMClient):
                 "stream": False
             }
 
+            # P0-0: inject tool definitions when provided.
+            if tools:
+                payload["tools"] = [t.to_openai_format() for t in tools]
+                payload["tool_choice"] = "auto"
+
             # Keep provider-specific behavior explicit so debugging stays straightforward.
             response = self._client.post("/chat/completions", json=payload)
             response.raise_for_status()
@@ -82,12 +98,24 @@ class QwenClient(BaseLLMClient):
             choices = data.get("choices", [{}])
             message = choices[0].get("message", {})
 
+            # P0-0: parse tool_calls returned by the model.
+            raw_tool_calls = message.get("tool_calls") or []
+            parsed_tool_calls: List[ToolCall] = []
+            for tc in raw_tool_calls:
+                fn = tc.get("function", {})
+                try:
+                    args = json.loads(fn.get("arguments", "{}"))
+                except json.JSONDecodeError:
+                    args = {}
+                parsed_tool_calls.append(ToolCall(id=tc.get("id", ""), name=fn.get("name", ""), arguments=args))
+
             return LLMResponse(
-                content=message.get("content", ""),
+                content=message.get("content", "") or "",
                 model=data.get("model", self.config.model),
                 usage=data.get("usage", {}),
                 finish_reason=choices[0].get("finish_reason", "stop"),
-                latency_ms=latency_ms
+                latency_ms=latency_ms,
+                tool_calls=parsed_tool_calls,
             )
 
         except httpx.HTTPStatusError as e: