Merge remote-tracking branch 'origin/main' into feature/upload-compatibility

app/core/constants.py

@@ -76,4 +76,15 @@ DEFAULT_SAFETY_SETTINGS = [
         {"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "threshold": "OFF"},
         {"category": "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold": "OFF"},
         {"category": "HARM_CATEGORY_CIVIC_INTEGRITY", "threshold": "BLOCK_NONE"},
-    ]
+    ]
+
+TTS_VOICE_NAMES = [
+    "Zephyr", "Puck", "Charon", "Kore",
+    "Fenrir", "Leda", "Orus", "Aoede",
+    "Callirhoe", "Autonoe", "Enceladus", "Iapetus",
+    "Umbriel", "Algieba", "Despina", "Erinome",
+    "Algenib", "Rasalgethi", "Laomedeia", "Achernar",
+    "Alnilam", "Schedar", "Gacrux", "Pulcherrima",
+    "Achird", "Zubenelgenubi", "Vindemiatrix", "Sadachbia",
+    "Sadaltager", "Sulafat"
+]

app/handler/response_handler.py

@@ -39,13 +39,13 @@ class GeminiResponseHandler(ResponseHandler):
 def _handle_openai_stream_response(
     response: Dict[str, Any], model: str, finish_reason: str, usage_metadata: Optional[Dict[str, Any]]
 ) -> Dict[str, Any]:
-    text, tool_calls, _ = _extract_result(
+    text, reasoning_content, tool_calls, _ = _extract_result(
         response, model, stream=True, gemini_format=False
     )
-    if not text and not tool_calls:
+    if not text and not tool_calls and not reasoning_content:
         delta = {}
     else:
-        delta = {"content": text, "role": "assistant"}
+        delta = {"content": text, "reasoning_content": reasoning_content, "role": "assistant"}
         if tool_calls:
             delta["tool_calls"] = tool_calls
     template_chunk = {
@@ -63,7 +63,7 @@ def _handle_openai_stream_response(
 def _handle_openai_normal_response(
     response: Dict[str, Any], model: str, finish_reason: str, usage_metadata: Optional[Dict[str, Any]]
 ) -> Dict[str, Any]:
-    text, tool_calls, _ = _extract_result(
+    text, reasoning_content, tool_calls, _ = _extract_result(
         response, model, stream=False, gemini_format=False
     )
     return {
@@ -77,6 +77,7 @@ def _handle_openai_normal_response(
                 "message": {
                     "role": "assistant",
                     "content": text,
+                    "reasoning_content": reasoning_content,
                     "tool_calls": tool_calls,
                 },
                 "finish_reason": finish_reason,
@@ -156,19 +157,21 @@ def _extract_result(
     model: str,
     stream: bool = False,
     gemini_format: bool = False,
-) -> tuple[str, List[Dict[str, Any]], Optional[bool]]:
-    text, tool_calls = "", []
-    thought = None
+) -> tuple[str, Optional[str], List[Dict[str, Any]], Optional[bool]]:
+    text, reasoning_content, tool_calls, thought = "", "", [], None
     if stream:
         if response.get("candidates"):
             candidate = response["candidates"][0]
             content = candidate.get("content", {})
             parts = content.get("parts", [])
             if not parts:
-                return "", [], None
+                return "", None, [], None
             if "text" in parts[0]:
                 text = parts[0].get("text")
                 if "thought" in parts[0]:
+                    if not gemini_format and settings.SHOW_THINKING_PROCESS:
+                        reasoning_content = text
+                        text = ""
                     thought = parts[0].get("thought")
             elif "executableCode" in parts[0]:
                 text = _format_code_block(parts[0]["executableCode"])
@@ -187,32 +190,18 @@ def _extract_result(
     else:
         if response.get("candidates"):
             candidate = response["candidates"][0]
-            if "thinking" in model:
-                if settings.SHOW_THINKING_PROCESS:
-                    if len(candidate["content"]["parts"]) == 2:
-                        text = (
-                            "> thinking\n\n"
-                            + candidate["content"]["parts"][0]["text"]
-                            + "\n\n---\n> output\n\n"
-                            + candidate["content"]["parts"][1]["text"]
-                        )
-                    else:
-                        text = candidate["content"]["parts"][0]["text"]
-                else:
-                    if len(candidate["content"]["parts"]) == 2:
-                        text = candidate["content"]["parts"][1]["text"]
-                    else:
-                        text = candidate["content"]["parts"][0]["text"]
-            else:
-                text = ""
-                if "parts" in candidate["content"]:
-                    for part in candidate["content"]["parts"]:
-                        if "text" in part:
+            text, reasoning_content = "", ""
+            if "parts" in candidate["content"]:
+                for part in candidate["content"]["parts"]:
+                    if "text" in part:
+                        if "thought" in part and settings.SHOW_THINKING_PROCESS:
+                            reasoning_content += part["text"]
+                        else:
                             text += part["text"]
-                            if "thought" in part and thought is None:
-                                thought = part.get("thought")
-                        elif "inlineData" in part:
-                            text += _extract_image_data(part)
+                        if "thought" in part and thought is None:
+                            thought = part.get("thought")
+                    elif "inlineData" in part:
+                        text += _extract_image_data(part)
 
             text = _add_search_link_text(model, candidate, text)
             tool_calls = _extract_tool_calls(
@@ -220,7 +209,7 @@ def _extract_result(
             )
         else:
             text = "暂无返回"
-    return text, tool_calls, thought
+    return text, reasoning_content, tool_calls, thought
 
 
 def _extract_image_data(part: dict) -> str:
@@ -294,7 +283,7 @@ def _extract_tool_calls(
 def _handle_gemini_stream_response(
     response: Dict[str, Any], model: str, stream: bool
 ) -> Dict[str, Any]:
-    text, tool_calls, thought = _extract_result(
+    text, reasoning_content, tool_calls, thought = _extract_result(
         response, model, stream=stream, gemini_format=True
     )
     if tool_calls:
@@ -311,16 +300,18 @@ def _handle_gemini_stream_response(
 def _handle_gemini_normal_response(
     response: Dict[str, Any], model: str, stream: bool
 ) -> Dict[str, Any]:
-    text, tool_calls, thought = _extract_result(
+    text, reasoning_content, tool_calls, thought = _extract_result(
         response, model, stream=stream, gemini_format=True
     )
+    parts = []
     if tool_calls:
-        content = {"parts": tool_calls, "role": "model"}
+        parts = tool_calls
     else:
-        part = {"text": text}
         if thought is not None:
-            part["thought"] = thought
-        content = {"parts": [part], "role": "model"}
+            parts.append({"text": reasoning_content,"thought": thought})
+        part = {"text": text}
+        parts.append(part)
+    content = {"parts": parts, "role": "model"}
     response["candidates"][0]["content"] = content
     return response
 

app/service/chat/gemini_chat_service.py

@@ -50,6 +50,33 @@ def _extract_file_references(contents: List[Dict[str, Any]]) -> List[str]:
                 logger.info(f"Found file reference: {file_id}")
     return file_names
 
+def _clean_json_schema_properties(obj: Any) -> Any:
+    """清理JSON Schema中Gemini API不支持的字段"""
+    if not isinstance(obj, dict):
+        return obj
+    
+    # Gemini API不支持的JSON Schema字段
+    unsupported_fields = {
+        "exclusiveMaximum", "exclusiveMinimum", "const", "examples", 
+        "contentEncoding", "contentMediaType", "if", "then", "else",
+        "allOf", "anyOf", "oneOf", "not", "definitions", "$schema",
+        "$id", "$ref", "$comment", "readOnly", "writeOnly"
+    }
+    
+    cleaned = {}
+    for key, value in obj.items():
+        if key in unsupported_fields:
+            continue
+        if isinstance(value, dict):
+            cleaned[key] = _clean_json_schema_properties(value)
+        elif isinstance(value, list):
+            cleaned[key] = [_clean_json_schema_properties(item) for item in value]
+        else:
+            cleaned[key] = value
+    
+    return cleaned
+
+
 def _build_tools(model: str, payload: Dict[str, Any]) -> List[Dict[str, Any]]:
     """构建工具"""
     
@@ -62,7 +89,15 @@ def _build_tools(model: str, payload: Dict[str, Any]) -> List[Dict[str, Any]]:
             for k, v in item.items():
                 if k == "functionDeclarations" and v and isinstance(v, list):
                     functions = record.get("functionDeclarations", [])
-                    functions.extend(v)
+                    # 清理每个函数声明中的不支持字段
+                    cleaned_functions = []
+                    for func in v:
+                        if isinstance(func, dict):
+                            cleaned_func = _clean_json_schema_properties(func)
+                            cleaned_functions.append(cleaned_func)
+                        else:
+                            cleaned_functions.append(func)
+                    functions.extend(cleaned_functions)
                     record["functionDeclarations"] = functions
                 else:
                     record[k] = v
@@ -136,6 +171,10 @@ def _build_payload(model: str, request: GeminiRequest) -> Dict[str, Any]:
         "systemInstruction": request_dict.get("systemInstruction"),
     }
 
+    # 确保 generationConfig 不为 None
+    if payload["generationConfig"] is None:
+        payload["generationConfig"] = {}
+
     if model.endswith("-image") or model.endswith("-image-generation"):
         payload.pop("systemInstruction")
         payload["generationConfig"]["responseModalities"] = ["Text", "Image"]
@@ -153,7 +192,13 @@ def _build_payload(model: str, request: GeminiRequest) -> Dict[str, Any]:
         if model.endswith("-non-thinking"):
             payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": 0} 
         elif model in settings.THINKING_BUDGET_MAP:
-            payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": settings.THINKING_BUDGET_MAP.get(model,1000)}
+            if settings.SHOW_THINKING_PROCESS:
+                payload["generationConfig"]["thinkingConfig"] = {
+                    "thinkingBudget": settings.THINKING_BUDGET_MAP.get(model,1000),
+                    "includeThoughts": True
+                }
+            else:
+                payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": settings.THINKING_BUDGET_MAP.get(model,1000)}
 
     return payload
 

app/service/chat/openai_chat_service.py

@@ -26,16 +26,43 @@ from app.service.key.key_manager import KeyManager
 logger = get_openai_logger()
 
 
-def _has_media_parts(contents: List[Dict[str, Any]]) -> bool:
-    """判断消息是否包含图片、音频或视频部分 (inline_data)"""
-    for content in contents:
-        if content and "parts" in content and isinstance(content["parts"], list):
-            for part in content["parts"]:
-                if isinstance(part, dict) and "inline_data" in part:
+def _has_media_parts(messages: List[Dict[str, Any]]) -> bool:
+    """判断消息是否包含多媒体部分"""
+    for message in messages:
+        if "parts" in message:
+            for part in message["parts"]:
+                if "image_url" in part or "inline_data" in part:
                     return True
     return False
 
 
+def _clean_json_schema_properties(obj: Any) -> Any:
+    """清理JSON Schema中Gemini API不支持的字段"""
+    if not isinstance(obj, dict):
+        return obj
+    
+    # Gemini API不支持的JSON Schema字段
+    unsupported_fields = {
+        "exclusiveMaximum", "exclusiveMinimum", "const", "examples", 
+        "contentEncoding", "contentMediaType", "if", "then", "else",
+        "allOf", "anyOf", "oneOf", "not", "definitions", "$schema",
+        "$id", "$ref", "$comment", "readOnly", "writeOnly"
+    }
+    
+    cleaned = {}
+    for key, value in obj.items():
+        if key in unsupported_fields:
+            continue
+        if isinstance(value, dict):
+            cleaned[key] = _clean_json_schema_properties(value)
+        elif isinstance(value, list):
+            cleaned[key] = [_clean_json_schema_properties(item) for item in value]
+        else:
+            cleaned[key] = value
+    
+    return cleaned
+
+
 def _build_tools(
     request: ChatRequest, messages: List[Dict[str, Any]]
 ) -> List[Dict[str, Any]]:
@@ -76,6 +103,8 @@ def _build_tools(
                 ):
                     function.pop("parameters", None)
 
+                # 清理函数中的不支持字段
+                function = _clean_json_schema_properties(function)
                 function_declarations.append(function)
 
         if function_declarations:
@@ -137,9 +166,13 @@ def _build_payload(
     if request.model.endswith("-non-thinking"):
         payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": 0}
     if request.model in settings.THINKING_BUDGET_MAP:
-        payload["generationConfig"]["thinkingConfig"] = {
-            "thinkingBudget": settings.THINKING_BUDGET_MAP.get(request.model, 1000)
-        }
+        if settings.SHOW_THINKING_PROCESS:
+            payload["generationConfig"]["thinkingConfig"] = {
+                "thinkingBudget": settings.THINKING_BUDGET_MAP.get(request.model, 1000),
+                "includeThoughts": True
+            }
+        else:
+            payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": settings.THINKING_BUDGET_MAP.get(request.model, 1000)}
 
     if (
         instruction

app/service/chat/vertex_express_chat_service.py

@@ -28,6 +28,33 @@ def _has_image_parts(contents: List[Dict[str, Any]]) -> bool:
     return False
 
 
+def _clean_json_schema_properties(obj: Any) -> Any:
+    """清理JSON Schema中Gemini API不支持的字段"""
+    if not isinstance(obj, dict):
+        return obj
+    
+    # Gemini API不支持的JSON Schema字段
+    unsupported_fields = {
+        "exclusiveMaximum", "exclusiveMinimum", "const", "examples", 
+        "contentEncoding", "contentMediaType", "if", "then", "else",
+        "allOf", "anyOf", "oneOf", "not", "definitions", "$schema",
+        "$id", "$ref", "$comment", "readOnly", "writeOnly"
+    }
+    
+    cleaned = {}
+    for key, value in obj.items():
+        if key in unsupported_fields:
+            continue
+        if isinstance(value, dict):
+            cleaned[key] = _clean_json_schema_properties(value)
+        elif isinstance(value, list):
+            cleaned[key] = [_clean_json_schema_properties(item) for item in value]
+        else:
+            cleaned[key] = value
+    
+    return cleaned
+
+
 def _build_tools(model: str, payload: Dict[str, Any]) -> List[Dict[str, Any]]:
     """构建工具"""
     
@@ -40,7 +67,15 @@ def _build_tools(model: str, payload: Dict[str, Any]) -> List[Dict[str, Any]]:
             for k, v in item.items():
                 if k == "functionDeclarations" and v and isinstance(v, list):
                     functions = record.get("functionDeclarations", [])
-                    functions.extend(v)
+                    # 清理每个函数声明中的不支持字段
+                    cleaned_functions = []
+                    for func in v:
+                        if isinstance(func, dict):
+                            cleaned_func = _clean_json_schema_properties(func)
+                            cleaned_functions.append(cleaned_func)
+                        else:
+                            cleaned_functions.append(func)
+                    functions.extend(cleaned_functions)
                     record["functionDeclarations"] = functions
                 else:
                     record[k] = v
@@ -98,10 +133,26 @@ def _build_payload(model: str, request: GeminiRequest) -> Dict[str, Any]:
         payload.pop("systemInstruction")
         payload["generationConfig"]["responseModalities"] = ["Text", "Image"]
         
-    if model.endswith("-non-thinking"):
-        payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": 0} 
-    if model in settings.THINKING_BUDGET_MAP:
-        payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": settings.THINKING_BUDGET_MAP.get(model,1000)}
+    # 处理思考配置：优先使用客户端提供的配置，否则使用默认配置
+    client_thinking_config = None
+    if request.generationConfig and request.generationConfig.thinkingConfig:
+        client_thinking_config = request.generationConfig.thinkingConfig
+    
+    if client_thinking_config is not None:
+        # 客户端提供了思考配置，直接使用
+        payload["generationConfig"]["thinkingConfig"] = client_thinking_config
+    else:
+        # 客户端没有提供思考配置，使用默认配置    
+        if model.endswith("-non-thinking"):
+            payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": 0} 
+        elif model in settings.THINKING_BUDGET_MAP:
+            if settings.SHOW_THINKING_PROCESS:
+                payload["generationConfig"]["thinkingConfig"] = {
+                    "thinkingBudget": settings.THINKING_BUDGET_MAP.get(model,1000),
+                    "includeThoughts": True
+                }
+            else:
+                payload["generationConfig"]["thinkingConfig"] = {"thinkingBudget": settings.THINKING_BUDGET_MAP.get(model,1000)}
 
     return payload
 

app/service/client/api_client.py

@@ -21,10 +21,6 @@ class ApiClient(ABC):
     async def stream_generate_content(self, payload: Dict[str, Any], model: str, api_key: str) -> AsyncGenerator[str, None]:
         pass
 
-    @abstractmethod
-    async def count_tokens(self, payload: Dict[str, Any], model: str, api_key: str) -> Dict[str, Any]:
-        pass
-
 
 class GeminiApiClient(ApiClient):
     """Gemini API客户端"""

app/service/tts/tts_service.py

@@ -8,6 +8,7 @@ from typing import Optional
 from google import genai
 
 from app.config.config import settings
+from app.core.constants import TTS_VOICE_NAMES
 from app.database.services import add_error_log, add_request_log
 from app.domain.openai_models import TTSRequest
 from app.log.logger import get_openai_logger
@@ -47,7 +48,7 @@ class TTSService:
                     "speech_config": {
                         "voice_config": {
                             "prebuilt_voice_config": {
-                                "voice_name": settings.TTS_VOICE_NAME
+                                "voice_name": request.voice if request.voice in TTS_VOICE_NAMES else settings.TTS_VOICE_NAME
                             }
                         }
                     },