Separate agent tuning from model setup

ai_knowledge_base_v2/apps/backend/app/services/model_service.py

@@ -59,6 +59,23 @@ class ModelClientService:
             return {"ok": False, "message": str(exc), "answer": ""}
         return {"ok": True, "message": "模型配置可用", "answer": answer[:500]}
 
+    @staticmethod
+    def debug_model(model: ModelConfig, rag_result: RagResult, overrides: dict[str, Any]) -> dict[str, Any]:
+        debug_model = _copy_model_with_overrides(model, overrides)
+        try:
+            answer = _call_configured_model(debug_model, rag_result, allow_no_hit=True)
+        except ExternalServiceError as exc:
+            return {"ok": False, "message": str(exc), "answer": "", "prompt": rag_result.prompt}
+        return {
+            "ok": True,
+            "message": "Agent 调试完成",
+            "answer": answer[:4000],
+            "prompt": rag_result.prompt,
+            "modelName": model.model_name,
+            "retrieveCount": len(rag_result.chunks),
+            "knowledgeIds": rag_result.knowledge_ids,
+        }
+
 
 def _mock_answer(rag_result: RagResult) -> str:
     if not rag_result.is_hit:
@@ -75,6 +92,35 @@ def _mock_answer(rag_result: RagResult) -> str:
     )
 
 
+def _copy_model_with_overrides(model: ModelConfig, overrides: dict[str, Any]) -> ModelConfig:
+    debug_model = ModelConfig(
+        provider=model.provider,
+        display_name=model.display_name,
+        api_type=model.api_type,
+        model_name=model.model_name,
+        base_url=model.base_url,
+        api_url=model.api_url,
+        api_key=model.api_key,
+        auth_type=model.auth_type,
+        api_version=model.api_version,
+        temperature=overrides.get("temperature"),
+        top_p=overrides.get("top_p"),
+        top_k=overrides.get("top_k"),
+        presence_penalty=overrides.get("presence_penalty"),
+        frequency_penalty=overrides.get("frequency_penalty"),
+        max_token=overrides.get("max_token"),
+        context_window=model.context_window,
+        stream_enabled=model.stream_enabled,
+        response_format=model.response_format,
+        extra_params=model.extra_params,
+        remark=model.remark,
+        timeout_second=model.timeout_second,
+        enabled=model.enabled,
+    )
+    debug_model.id = model.id
+    return debug_model
+
+
 def _rough_token_count(text: str) -> int:
     return max(1, len(text.strip()) // 2)