feat: support function call

app/services/chat/message_converter.py

@@ -37,13 +37,15 @@ class OpenAIMessageConverter(MessageConverter):
             parts = []
 
             if isinstance(msg["content"], str):
-                parts.append({"text": msg["content"]})
+                # 请求 gemini 接口时如果包含 content 字段但内容为空时会返回 400 错误，所以需要判断是否为空
+                if msg["content"]:
+                    parts.append({"text": msg["content"]})
             elif isinstance(msg["content"], list):
                 for content in msg["content"]:
-                    if isinstance(content, str):
+                    if isinstance(content, str) and content:
                         parts.append({"text": content})
                     elif isinstance(content, dict):
-                        if content["type"] == "text":
+                        if content["type"] == "text" and content["text"]:
                             parts.append({"text": content["text"]})
                         elif content["type"] == "image_url":
                             parts.append(_convert_image(content["image_url"]["url"]))

app/services/chat/response_handler.py

@@ -1,7 +1,10 @@
 # app/services/chat/response_handler.py
 
+import json
+import random
+import string
 from abc import ABC, abstractmethod
-from typing import Dict, Any, Optional
+from typing import Dict, Any, List, Optional
 import time
 import uuid
 from app.core.config import settings
@@ -29,40 +32,38 @@ class GeminiResponseHandler(ResponseHandler):
 
 
 def _handle_openai_stream_response(response: Dict[str, Any], model: str, finish_reason: str) -> Dict[str, Any]:
-    text = _extract_text(response, model, stream=True)
+    text, tool_calls = _extract_result(response, model, stream=True, gemini_format=False)
+    if not text and not tool_calls:
+        delta = {}
+    else:
+        delta = {"content": text, "role": "assistant"}
+        if tool_calls:
+            delta["tool_calls"] = tool_calls
+
     return {
         "id": f"chatcmpl-{uuid.uuid4()}",
         "object": "chat.completion.chunk",
         "created": int(time.time()),
         "model": model,
-        "choices": [{
+        "choices": [{"index": 0, "delta": delta, "finish_reason": finish_reason}],
-            "index": 0,
-            "delta": {"content": text} if text else {},
-            "finish_reason": finish_reason
-        }]
     }
 
 
 def _handle_openai_normal_response(response: Dict[str, Any], model: str, finish_reason: str) -> Dict[str, Any]:
-    text = _extract_text(response, model, stream=False)
+    text, tool_calls = _extract_result(response, model, stream=False, gemini_format=False)
     return {
         "id": f"chatcmpl-{uuid.uuid4()}",
         "object": "chat.completion",
         "created": int(time.time()),
         "model": model,
-        "choices": [{
+        "choices": [
-            "index": 0,
+            {
-            "message": {
+                "index": 0,
-                "role": "assistant",
+                "message": {"role": "assistant", "content": text, "tool_calls": tool_calls},
-                "content": text
+                "finish_reason": finish_reason,
-            },
+            }
-            "finish_reason": finish_reason
+        ],
-        }],
+        "usage": {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
-        "usage": {
-            "prompt_tokens": 0,
-            "completion_tokens": 0,
-            "total_tokens": 0
-        }
     }
 
 
@@ -127,8 +128,8 @@ def _handle_openai_normal_image_response(image_str: str,model: str,finish_reason
     }
 
 
-def _extract_text(response: Dict[str, Any], model: str, stream: bool = False) -> str:
+def _extract_result(response: Dict[str, Any], model: str, stream: bool = False, gemini_format: bool = False) -> tuple[str, List[Dict[str, Any]]]:
-    text = ""
+    text, tool_calls = "", []
     if stream:
         if response.get("candidates"):
             candidate = response["candidates"][0]
@@ -212,6 +213,7 @@ def _extract_text(response: Dict[str, Any], model: str, stream: bool = False) ->
             else:
                 text = ""
             text = _add_search_link_text(model, candidate, text)
+            tool_calls = _extract_tool_calls(parts, gemini_format)
     else:
         if response.get("candidates"):
             candidate = response["candidates"][0]
@@ -234,23 +236,65 @@ def _extract_text(response: Dict[str, Any], model: str, stream: bool = False) ->
             else:
                 text = ""
                 for part in candidate["content"]["parts"]:
-                    text += part["text"]
+                    text += part.get("text", "")
             text = _add_search_link_text(model, candidate, text)
+            tool_calls = _extract_tool_calls(candidate["content"]["parts"], gemini_format)
         else:
             text = "暂无返回"
-    return text
+    return text, tool_calls
+
+def _extract_tool_calls(parts: List[Dict[str, Any]], gemini_format: bool) -> List[Dict[str, Any]]:
+    """提取工具调用信息"""
+    if not parts or not isinstance(parts, list):
+        return []
+
+    letters = string.ascii_lowercase + string.digits
+
+    tool_calls = list()
+    for i in range(len(parts)):
+        part = parts[i]
+        if not part or not isinstance(part, dict):
+            continue
+
+        item = part.get("functionCall", {})
+        if not item or not isinstance(item, dict):
+            continue
+
+        if gemini_format:
+            tool_calls.append(part)
+        else:
+            id = f"call_{''.join(random.sample(letters, 32))}"
+            name = item.get("name", "")
+            arguments = json.dumps(item.get("args", None) or {})
+
+            tool_calls.append(
+                {
+                    "index": i,
+                    "id": id,
+                    "type": "function",
+                    "function": {"name": name, "arguments": arguments},
+                }
+            )
+
+    return tool_calls
 
 
 def _handle_gemini_stream_response(response: Dict[str, Any], model: str, stream: bool) -> Dict[str, Any]:
-    text = _extract_text(response, model, stream=stream)
+    text, tool_calls = _extract_result(response, model, stream=stream, gemini_format=True)
-    content = {"parts": [{"text": text}], "role": "model"}
+    if tool_calls:
+        content = {"parts": tool_calls, "role": "model"}
+    else:
+        content = {"parts": [{"text": text}], "role": "model"}
     response["candidates"][0]["content"] = content
     return response
 
 
 def _handle_gemini_normal_response(response: Dict[str, Any], model: str, stream: bool) -> Dict[str, Any]:
-    text = _extract_text(response, model, stream=stream)
+    text, tool_calls = _extract_result(response, model, stream=stream, gemini_format=True)
-    content = {"parts": [{"text": text}], "role": "model"}
+    if tool_calls:
+        content = {"parts": tool_calls, "role": "model"}
+    else:
+        content = {"parts": [{"text": text}], "role": "model"}
     response["candidates"][0]["content"] = content
     return response
 

app/services/openai_chat_service.py

@@ -1,5 +1,6 @@
 # app/services/chat_service.py
 
+from copy import deepcopy
 import json
 from typing import Dict, Any, AsyncGenerator, List, Union
 from app.core.logger import get_openai_logger
@@ -39,6 +40,25 @@ def _build_tools(
         tools.append({"code_execution": {}})
     if model.endswith("-search"):
         tools.append({"googleSearch": {}})
+
+    # 将 request 中的 tools 合并到 tools 中
+    if request.tools:
+        function_declarations = []
+        for tool in request.tools:
+            if not tool or not isinstance(tool, dict):
+                continue
+
+            if tool.get("type", "") == "function" and tool.get("function"):
+                function = deepcopy(tool.get("function"))
+                parameters = function.get("parameters", {})
+                if parameters.get("type") == "object" and not parameters.get("properties", {}):
+                    function.pop("parameters", None)
+
+                function_declarations.append(function)
+
+        if function_declarations:
+            tools.append({"functionDeclarations": function_declarations})
+            
     return tools