Maas-Model-platform/backend/app/services/openai_compat_service.py

"""
OpenAI 兼容服务层

完整支持 /v1/chat/completions 和 /v1/models 接口
支持多种模型提供商的自动适配
"""
import asyncio
import logging
import time
import uuid
from decimal import Decimal
import httpx
from fastapi import UploadFile
from app.services.api_call_log_service import ApiCallLogService
from app.services.image_service import ImageGenerationService
from app.services.model_adapters import BaseAdapter, get_adapter, ModelProvider
from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple, Union

from sqlalchemy import desc
from sqlalchemy.orm import Session

from app.models.model import ModelNew as Model, ModelPriceNew as ModelPrice
from app.models.user import User
from app.schemas.openai_compat import (
    ChatCompletionsRequest,
    ModelInfo,
    EmbeddingsRequest, 
    EmbeddingsResponse, 
    EmbeddingData, 
    Usage,
    ModelsListResponse,
    ImageGenerationRequest,
    ImageGenerationResponse,
    ImageData,
    AudioTranscriptionResponse,
    AudioSpeechRequest,
    VideoGenerationRequest,
    VideoGenerationResponse,
    RerankRequest,
    RerankResponse,
    RerankResult,
)
from app.services.crypto_utils import decrypt_api_key
from app.services.system_config_manager import get_config_bool

logger = logging.getLogger(__name__)


# ─────────────────────────────────────────────
# 工具函数
# ─────────────────────────────────────────────

def parse_video_size(size: str) -> Tuple[str, str]:
    """
    解析视频尺寸，支持多种格式并转换为OpenAI标准格式和内部格式
    
    Args:
        size: 视频尺寸，支持以下格式：
            - OpenAI格式: "1280x720", "1920x1080", "720x1280"
            - 简写格式: "720P", "1080P", "720p", "1080p"
    
    Returns:
        (openai_format, internal_format)
        例如: ("1280x720", "720P")
    
    Raises:
        ValueError: 如果格式无效
    """
    import re
    
    size = size.strip()
    
    # 如果是OpenAI格式 (widthxheight)
    if 'x' in size.lower():
        match = re.match(r'^(\d+)x(\d+)$', size.lower())
        if not match:
            raise ValueError(f"Invalid size format: {size}. Expected format: 1280x720")
        
        width, height = int(match.group(1)), int(match.group(2))
        
        # 推断简写格式（基于高度）
        if height == 720:
            internal = "720P"
        elif height == 1080:
            internal = "1080P"
        else:
            # 对于非标准分辨率，使用高度作为简写
            internal = f"{height}P"
        
        return (f"{width}x{height}", internal)
    
    # 如果是简写格式 (720P, 1080P)
    else:
        match = re.match(r'^(\d+)p$', size.lower())
        if not match:
            raise ValueError(f"Invalid size format: {size}. Expected format: 720P or 1280x720")
        
        height = int(match.group(1))
        size_upper = f"{height}P"
        
        # 标准分辨率映射（16:9比例）
        if height == 720:
            return ("1280x720", "720P")
        elif height == 1080:
            return ("1920x1080", "1080P")
        else:
            # 对于非标准分辨率，假设16:9比例
            width = int(height * 16 / 9)
            return (f"{width}x{height}", size_upper)


class OpenAICompatError(Exception):
    """OpenAI 兼容服务错误"""

    def __init__(self, status_code: int, message: str, error_type: str = "invalid_request_error"):
        self.status_code = status_code
        self.message = message
        self.error_type = error_type
        super().__init__(message)


class OpenAICompatService:
    """OpenAI API 兼容服务"""

    def __init__(self, db: Session):
        self.db = db
        self._user_cache: dict = {}  # user_id → User，请求内缓存，避免重复查询

    def _get_user(self, user_id: str):
        """获取用户对象，同一请求内缓存，避免重复查询 users 表。"""
        if user_id not in self._user_cache:
            self._user_cache[user_id] = self.db.query(User).filter(User.id == user_id).first()
        return self._user_cache[user_id]

    # ─────────────────────────────────────────────
    # 主入口
    # ─────────────────────────────────────────────

    async def chat_completions(
        self,
        request: ChatCompletionsRequest,
        user_id: str,
        api_key_id: int,
        request_ip: Optional[str] = None,
    ) -> Union[Dict[str, Any], AsyncGenerator[str, None]]:
        """
        处理 Chat Completions 请求，包含日志记录与扣费。

        权限检查流程：
        1. 验证模型是否存在
        2. 检查模型 is_api_enabled（云端模型）
        3. 检查用户余额（云端模型）
        4. 获取用户 API Key（云端模型）
        5. 检查用户对本地模型的访问权限（本地模型）
        """
        log_service = ApiCallLogService(self.db)

        model = self._find_model(request.model, user_id)

        if not model:
            raise OpenAICompatError(
                status_code=404,
                message=f"The model '{request.model}' does not exist",
                error_type="model_not_found",
            )

        if not model.is_local and not model.is_api_enabled:
            raise OpenAICompatError(
                status_code=403,
                message=f"Model '{request.model}' does not support API calls",
                error_type="model_not_available",
            )

        # realtime 模型使用 WebSocket 实时流协议，不支持此 REST 接口
        if "realtime" in request.model.lower() and not model.is_local:
            raise OpenAICompatError(
                status_code=400,
                message=f"Model '{request.model}' is a real-time streaming model that uses WebSocket protocol. "
                        f"It cannot be called via /api/v1/chat/completions. "
                        f"Please use the WebSocket API instead.",
                error_type="model_not_supported",
            )

        # 检查本地模型的访问权限
        if model.is_local:
            from app.services.user_local_model_permission_service import UserLocalModelPermissionService
            permission_service = UserLocalModelPermissionService(self.db)
            if not await permission_service.check_user_model_access(user_id, model.id):
                raise OpenAICompatError(
                    status_code=403,
                    message=f"You don't have permission to access model '{request.model}'",
                    error_type="permission_error",
                )

        user_api_key: Optional[str] = None
        if not model.is_local:
            user = self._get_user(user_id)
            if not user:
                raise OpenAICompatError(
                    status_code=401,
                    message="User not found",
                    error_type="authentication_error",
                )

            # 优先使用模型自带的 api_key（爬虫同步的），没有则 fallback 到用户自己配置的 apikey
            if model.encrypted_api_key:
                from app.services.crypto_utils import decrypt_api_key
                decrypted = decrypt_api_key(model.encrypted_api_key)
                user_api_key = decrypted if decrypted else None

            if not user_api_key:
                user_api_key = user.apikey

            if not user_api_key:
                raise OpenAICompatError(
                    status_code=403,
                    message="User API key not configured. Please configure your API key in settings.",
                    error_type="api_key_not_configured",
                )

        # ── OCR 模型校验：必须包含图片 ────────────────────────────────────
        OCR_MODELS = ("qwen-vl-ocr",)
        if model.model_code in OCR_MODELS and not model.is_local:
            has_image = any(
                isinstance(msg.content, list) and
                any(isinstance(part, dict) and part.get("type") == "image_url"
                    for part in msg.content)
                for msg in request.messages
            )
            # 也兼容 Pydantic 对象形式
            if not has_image:
                from app.schemas.openai_compat import ContentPartImage
                has_image = any(
                    isinstance(msg.content, list) and
                    any(isinstance(part, ContentPartImage) for part in msg.content)
                    for msg in request.messages
                )
            if not has_image:
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{model.model_code}' is an OCR model and requires at least one image in the messages. "
                            f"Please include an image_url content part in your user message.",
                    error_type="invalid_request_error",
                )

        # ── 流式 ──────────────────────────────────────────────────────────
        if request.stream:
            raw_stream = await self._call_local_model(model, request) if model.is_local \
                else await self._call_cloud_model(model, request, user_api_key)

            async def _stream_with_billing() -> AsyncGenerator[str, None]:
                input_text = "".join(
                    [m.content for m in request.messages if isinstance(m.content, str)]
                )
                input_tokens = max(int(len(input_text) * 1.2), 1)
                output_tokens = 0
                stream_error: Optional[Exception] = None
                try:
                    async for chunk in raw_stream:
                        yield chunk
                        if isinstance(chunk, str) and chunk.startswith("data: ") \
                                and "data: [DONE]" not in chunk:
                            try:
                                import json as _json
                                data_dict = _json.loads(chunk[6:])
                                delta = data_dict.get("choices", [{}])[0] \
                                    .get("delta", {}).get("content", "")
                                if delta:
                                    output_tokens += max(int(len(delta) * 1.2), 1)
                                # 优先使用上游返回的真实 usage（部分模型在最后一个 chunk 里带）
                                usage = data_dict.get("usage")
                                if usage:
                                    input_tokens = usage.get("prompt_tokens", input_tokens)
                                    output_tokens = usage.get("completion_tokens", output_tokens)
                            except Exception:
                                pass
                except (GeneratorExit, asyncio.CancelledError):
                    # 客户端断开/任务取消 - 仍按已产生的 token 扣费
                    raise
                except Exception as exc:
                    stream_error = exc
                    raise
                finally:
                    # 关键：无论流正常结束、客户端中断、还是异常都要扣费
                    # 防止 "token 已被消耗但未扣费" 的资损
                    try:
                        if stream_error is not None:
                            # 上游错误 - 记录失败日志
                            log_service.create_log(
                                user_id=user_id, api_key_id=api_key_id,
                                model_id=model.id, model_name=request.model,
                                is_local=model.is_local,
                                input_tokens=0, output_tokens=0,
                                bill=0, status="failed",
                                error_message=str(stream_error), request_ip=request_ip,
                            )
                        else:
                            bill = self.calculate_bill(model, input_tokens, output_tokens)
                            log_service.create_log(
                                user_id=user_id, api_key_id=api_key_id,
                                model_id=model.id, model_name=request.model,
                                is_local=model.is_local,
                                input_tokens=input_tokens, output_tokens=output_tokens,
                                bill=float(bill), status="success", request_ip=request_ip,
                            )
                    except Exception as fin_exc:
                        logger.error("流式响应收尾日志记录失败: %s", fin_exc)

            return _stream_with_billing()

        # ── 非流式 ────────────────────────────────────────────────────────
        try:
            result = await self._call_local_model(model, request) if model.is_local \
                else await self._call_cloud_model(model, request, user_api_key)

            input_tokens, output_tokens = self.extract_usage_from_response(result)
            bill = self.calculate_bill(model, input_tokens, output_tokens)

            log_service.create_log(
                user_id=user_id, api_key_id=api_key_id,
                model_id=model.id, model_name=request.model,
                is_local=model.is_local,
                input_tokens=input_tokens, output_tokens=output_tokens,
                bill=float(bill), status="success", request_ip=request_ip,
            )

            return result
        except OpenAICompatError:
            raise
        except Exception as exc:
            error_msg = str(exc) or repr(exc)
            logger.warning(
                "[CHAT_COMPLETION_FAILED] model=%s user_id=%s error_type=%s error=%s",
                request.model, user_id, type(exc).__name__, error_msg,
            )
            log_service.create_log(
                user_id=user_id, api_key_id=api_key_id,
                model_id=model.id, model_name=request.model,
                is_local=model.is_local,
                input_tokens=0, output_tokens=0,
                bill=0, status="failed",
                error_message=error_msg, request_ip=request_ip,
            )
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="upstream_error")

    # ─────────────────────────────────────────────
    # 构建请求体（完整参数透传）
    # ─────────────────────────────────────────────

    def _build_request_body(
        self, request: ChatCompletionsRequest, model_name: str
    ) -> Dict[str, Any]:
        """将 ChatCompletionsRequest 转换为上游 API 请求体，透传所有非 None 参数"""

        # 序列化消息（content 支持 str 或 list）
        messages = []
        for msg in request.messages:
            m: Dict[str, Any] = {"role": msg.role}

            if msg.content is None:
                m["content"] = None
            elif isinstance(msg.content, str):
                m["content"] = msg.content
            else:
                # 多模态内容列表
                parts = []
                for part in msg.content:
                    part_dict = part.model_dump(exclude_none=True)
                    # 校验 image_url 格式
                    if part_dict.get("type") == "image_url":
                        url = (part_dict.get("image_url") or {}).get("url", "")
                        if url and not url.startswith("data:"):
                            import os
                            ext = os.path.splitext(url.split("?")[0].lower())[1]
                            SUPPORTED_IMAGE_EXTS = {'.jpg', '.jpeg', '.png', '.webp', '.gif', '.bmp'}
                            if ext and ext not in SUPPORTED_IMAGE_EXTS:
                                raise OpenAICompatError(
                                    status_code=400,
                                    message=f"Unsupported image format '{ext}'. "
                                            f"Supported formats: jpg, jpeg, png, webp, gif, bmp.",
                                    error_type="invalid_request_error",
                                )
                    parts.append(part_dict)
                m["content"] = parts

            if msg.name is not None:
                m["name"] = msg.name
            if msg.tool_calls is not None:
                m["tool_calls"] = [tc.model_dump() for tc in msg.tool_calls]
            if msg.tool_call_id is not None:
                m["tool_call_id"] = msg.tool_call_id

            messages.append(m)

        body: Dict[str, Any] = {
            "model": model_name,
            "messages": messages,
            "stream": request.stream,
        }

        # 可选参数：只有非 None 才透传
        optional_fields = [
            "temperature", "top_p", "n", "stop",
            "presence_penalty", "frequency_penalty",
            "logit_bias", "logprobs", "top_logprobs",
            "seed", "user", "service_tier", "store", "metadata",
            "parallel_tool_calls",
        ]
        for field in optional_fields:
            val = getattr(request, field, None)
            if val is not None:
                body[field] = val

        # max_tokens / max_completion_tokens（优先新版）
        if request.max_completion_tokens is not None:
            body["max_completion_tokens"] = request.max_completion_tokens
        elif request.max_tokens is not None:
            body["max_tokens"] = request.max_tokens

        # 流式选项
        if request.stream and request.stream_options is not None:
            body["stream_options"] = request.stream_options.model_dump(exclude_none=True)

        # 工具调用
        if request.tools is not None:
            body["tools"] = [t.model_dump(exclude_none=True) for t in request.tools]
        if request.tool_choice is not None:
            if isinstance(request.tool_choice, str):
                body["tool_choice"] = request.tool_choice
            else:
                body["tool_choice"] = request.tool_choice.model_dump()

        # 响应格式
        if request.response_format is not None:
            body["response_format"] = request.response_format.model_dump(exclude_none=True)

        return body

    # ─────────────────────────────────────────────
    # 本地模型调用
    # ─────────────────────────────────────────────

    async def _call_local_model(
        self, model: Model, request: ChatCompletionsRequest
    ) -> Union[Dict[str, Any], AsyncGenerator[str, None]]:
        # 从缓存获取模型信息
        from app.services.cache_service import CacheService
        model_data = await CacheService.get_model(model.id)
        
        if model_data:
            base_url = model_data.get("base_url", "").rstrip("/")
            local_api_key = model_data.get("local_api_key")
        else:
            # 从数据库获取
            base_url = (model.base_url or "").rstrip("/")
            local_api_key = model.local_api_key
            # 缓存模型信息
            await CacheService.set_model(model.id, {
                "base_url": base_url,
                "local_api_key": local_api_key,
                "is_local": model.is_local,
                "name": model.name
            })

        if not base_url:
            raise OpenAICompatError(
                status_code=500,
                message="本地模型未配置 Base URL",
                error_type="configuration_error",
            )

        headers: Dict[str, str] = {"Content-Type": "application/json"}
        # 本地模型不使用用户的API密钥，而是使用模型配置的API密钥
        if local_api_key:
            api_key = decrypt_api_key(local_api_key)
            if api_key:
                headers["Authorization"] = f"Bearer {api_key}"

        # 使用模型的name字段作为实际模型名称
        actual_name = model.name

        body = self._build_request_body(request, actual_name)
        api_url = f"{base_url}/chat/completions"

        if request.stream:
            return self._stream_response(api_url, headers, body)
        else:
            return await self._non_stream_response(api_url, headers, body)

    # ─────────────────────────────────────────────
    # 云端模型调用（阿里云百炼 OpenAI 兼容模式）
    # ─────────────────────────────────────────────

    async def _call_cloud_model(
        self, model: Model, request: ChatCompletionsRequest, user_api_key: str
    ) -> Union[Dict[str, Any], AsyncGenerator[str, None]]:
        api_url = "https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions"
        headers = {
            "Content-Type": "application/json",
            "Authorization": f"Bearer {user_api_key}",
        }
        body = self._build_request_body(request, model.model_code)

        if request.stream:
            return self._stream_response(api_url, headers, body)
        else:
            return await self._non_stream_response(api_url, headers, body)

    # ─────────────────────────────────────────────
    # HTTP 请求封装
    # ─────────────────────────────────────────────

    async def _non_stream_response(
        self, api_url: str, headers: Dict[str, str], body: Dict[str, Any],
        timeout: float = 300.0,
    ) -> Dict[str, Any]:
        model_name = body.get("model", "unknown")
        try:
            async with httpx.AsyncClient(timeout=timeout) as client:
                resp = await client.post(api_url, headers=headers, json=body)
                if resp.status_code >= 400:
                    error_msg = self._extract_upstream_error(resp)
                    logger.warning(
                        "[UPSTREAM_ERROR] model=%s status=%d url=%s response=%s",
                        model_name, resp.status_code, api_url, resp.text[:500],
                    )
                    raise OpenAICompatError(
                        status_code=resp.status_code,
                        message=error_msg,
                        error_type="upstream_error",
                    )
                return resp.json()
        except httpx.ReadTimeout:
            logger.warning("[UPSTREAM_TIMEOUT] model=%s url=%s timeout=%ss", model_name, api_url, timeout)
            raise OpenAICompatError(
                status_code=504,
                message=f"模型 '{model_name}' 响应超时（{timeout}s），请稍后重试或换个模型",
                error_type="timeout_error",
            )

    async def _stream_response(
        self, api_url: str, headers: Dict[str, str], body: Dict[str, Any]
    ) -> AsyncGenerator[str, None]:
        """SSE 流式响应生成器"""
        model_name = body.get("model", "unknown")
        try:
            async with httpx.AsyncClient(
                timeout=httpx.Timeout(30.0, read=None)
            ) as client:
                async with client.stream("POST", api_url, headers=headers, json=body) as resp:
                    if resp.status_code >= 400:
                        error_body = await resp.aread()
                        error_text = error_body.decode("utf-8", errors="replace")

                        logger.warning(
                            "[UPSTREAM_STREAM_ERROR] model=%s status=%d url=%s response=%s",
                            model_name, resp.status_code, api_url, error_text[:500],
                        )

                        # 使用统一的错误提取方法
                        error_detail = self._extract_upstream_error(resp)

                        # 针对特定状态码提供友好提示
                        if resp.status_code == 401:
                            raise OpenAICompatError(
                                status_code=500,
                                message=f"模型 '{model_name}' 认证失败（401 Unauthorized）。\n"
                                        f"请在管理后台检查该模型的 API Key 配置是否正确。\n"
                                        f"错误详情: {error_detail}",
                                error_type="authentication_error",
                            )
                        elif resp.status_code == 404:
                            raise OpenAICompatError(
                                status_code=500,
                                message=f"模型 '{model_name}' 接口不存在（404 Not Found）。请检查 Base URL 和模型名称是否正确。\n"
                                        f"错误详情: {error_detail}",
                                error_type="not_found_error",
                            )
                        else:
                            raise OpenAICompatError(
                                status_code=resp.status_code,
                                message=f"模型 '{model_name}' 调用失败（{resp.status_code}）: {error_detail}",
                                error_type="upstream_error",
                            )

                    async for line in resp.aiter_lines():
                        if line.startswith("data: "):
                            data = line[6:]
                            if data.strip() == "[DONE]":
                                yield "data: [DONE]\n\n"
                                return
                            yield f"data: {data}\n\n"
                        elif line.strip():
                            yield f"data: {line}\n\n"
        except httpx.ConnectTimeout:
            logger.warning("[UPSTREAM_CONNECT_TIMEOUT] model=%s url=%s", model_name, api_url)
            raise OpenAICompatError(
                status_code=504,
                message=f"模型 '{model_name}' 连接超时，请稍后重试",
                error_type="timeout_error",
            )

    def _extract_upstream_error(self, resp: httpx.Response) -> str:
        """从上游错误响应中提取错误信息，确保始终返回有意义的内容"""
        raw_text = ""
        try:
            raw_text = resp.text
            data = resp.json()

            # 标准 OpenAI 格式: {"error": {"message": "...", "type": "...", "code": "..."}}
            err = data.get("error", {})
            if isinstance(err, dict):
                msg = err.get("message", "")
                err_type = err.get("type", "")
                code = err.get("code", "")
                parts = [p for p in [msg, f"type={err_type}" if err_type else "", f"code={code}" if code else ""] if p]
                if parts:
                    return " | ".join(parts)

            # DashScope 格式: {"code": "...", "message": "...", "request_id": "..."}
            code = data.get("code", "")
            msg = data.get("message", "")
            request_id = data.get("request_id", "")
            if msg:
                parts = [msg]
                if code:
                    parts.append(f"code={code}")
                if request_id:
                    parts.append(f"request_id={request_id}")
                return " | ".join(parts)

            # 兜底：返回整个 JSON
            if data:
                return str(data)
        except Exception:
            pass

        # JSON 解析失败或为空，返回原始文本
        if raw_text:
            return raw_text[:500]
        return f"Upstream error {resp.status_code} (empty response body)"

    # ─────────────────────────────────────────────
    # Models 列表
    # ─────────────────────────────────────────────

    def get_available_models(self, user_id: str, key_type: str = "public") -> ModelsListResponse:
        """返回用户可用的模型，根据API密钥类型过滤"""
        models_data: List[ModelInfo] = []

        # 根据密钥类型返回相应的模型
        if key_type == "public":
            # 公钥只能访问云端模型
            cloud_models = (
                self.db.query(Model)
                .filter(Model.is_local == False, Model.is_api_enabled == True)
                .all()
            )
            for m in cloud_models:
                models_data.append(
                    ModelInfo(
                        id=m.model_code,
                        object="model",
                        created=int(m.created_at.timestamp()) if m.created_at else int(time.time()),
                        owned_by=m.supplier or "platform",
                    )
                )
        elif key_type == "local":
            # 检查本地模型是否启用
            if get_config_bool("enable_local_models", True):
                # 如果本地模型启用，返回所有本地模型
                local_models = (
                    self.db.query(Model)
                    .filter(
                        Model.is_local == True
                    )
                    .all()
                )
                for m in local_models:
                    models_data.append(
                        ModelInfo(
                            id=f"local:{m.id}",
                            object="model",
                            created=int(m.created_at.timestamp()) if m.created_at else int(time.time()),
                            owned_by="local",
                        )
                    )
            else:
                # 如果本地模型未启用，返回用户有权限的模型
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                permissions = permission_service.get_user_model_permissions(user_id)
                permitted_model_ids = [perm["model_id"] for perm in permissions if perm["has_access"]]
                
                local_models = (
                    self.db.query(Model)
                    .filter(
                        Model.is_local == True,
                        Model.id.in_(permitted_model_ids)
                    )
                    .all()
                )
                for m in local_models:
                    models_data.append(
                        ModelInfo(
                            id=f"local:{m.id}",
                            object="model",
                            created=int(m.created_at.timestamp()) if m.created_at else int(time.time()),
                            owned_by="local",
                        )
                    )

        return ModelsListResponse(object="list", data=models_data)

    # ─────────────────────────────────────────────
    # 工具方法
    # ─────────────────────────────────────────────

    async def _handle_local_model_request(
        self,
        api_url: str,
        headers: Dict[str, str],
        payload: Dict[str, Any],
        model_name: str,
        base_url: str,
        endpoint_type: str = "chat",
        timeout: float = 60.0,
        return_raw_response: bool = False
    ) -> Union[Dict[str, Any], httpx.Response]:
        """
        统一处理本地模型 HTTP 请求，直接透传 OpenAI 格式
        
        注意：本地模型必须是 OpenAI 兼容的，不做任何格式适配
        
        Args:
            api_url: 请求URL
            headers: 请求头
            payload: 请求体（OpenAI 格式）
            model_name: 模型名称（用于错误提示）
            base_url: 模型的 base_url
            endpoint_type: 端点类型（保留参数，暂未使用）
            timeout: 超时时间
            return_raw_response: 是否返回原始响应对象（用于处理音频等二进制数据）
            
        Returns:
            响应的 JSON 数据（OpenAI 格式）或原始 httpx.Response 对象
            
        Raises:
            OpenAICompatError: 请求失败时抛出
        """
        async with httpx.AsyncClient(timeout=timeout) as client:
            try:
                response = await client.post(api_url, headers=headers, json=payload)
                response.raise_for_status()
                
                # 如果需要原始响应（例如音频数据），直接返回
                if return_raw_response:
                    return response
                
                # 直接返回 JSON 响应，不做任何转换
                result = response.json()
                return result
                
            except httpx.HTTPStatusError as e:
                if e.response.status_code == 401:
                    # 认证失败
                    error_detail = ""
                    try:
                        error_data = e.response.json()
                        if isinstance(error_data, dict):
                            error_detail = error_data.get("error", {}).get("message", "") or error_data.get("message", "")
                    except:
                        error_detail = e.response.text[:200]
                    
                    raise OpenAICompatError(
                        status_code=500,
                        message=f"本地模型 '{model_name}' 认证失败（401 Unauthorized）。\n"
                                f"请在管理后台检查该模型的 API Key 配置是否正确。\n"
                                f"错误详情: {error_detail}",
                        error_type="authentication_error",
                    )
                elif e.response.status_code == 404:
                    raise OpenAICompatError(
                        status_code=500,
                        message=f"本地模型 '{model_name}' 接口不存在（404 Not Found）。请检查 Base URL 和模型名称是否正确。",
                        error_type="not_found_error",
                    )
                else:
                    # 其他 HTTP 错误
                    error_detail = ""
                    try:
                        error_data = e.response.json()
                        if isinstance(error_data, dict):
                            error_detail = error_data.get("error", {}).get("message", "") or error_data.get("message", "")
                    except:
                        error_detail = e.response.text[:200]
                    
                    raise OpenAICompatError(
                        status_code=e.response.status_code,
                        message=f"本地模型调用失败（{e.response.status_code}）: {error_detail}",
                        error_type="upstream_error",
                    )
            except httpx.TimeoutException:
                raise OpenAICompatError(
                    status_code=504,
                    message=f"本地模型 '{model_name}' 请求超时。请检查网络连接或增加超时时间。",
                    error_type="timeout_error",
                )
            except httpx.RequestError as e:
                raise OpenAICompatError(
                    status_code=500,
                    message=f"本地模型 '{model_name}' 请求失败: {str(e)}",
                    error_type="request_error",
                )

    def _find_model(self, model_name: str, user_id: str) -> Optional[Model]:
        # 优先识别 local:{id} 格式（精确匹配）
        if model_name.startswith("local:"):
            try:
                model_id = int(model_name[6:])
            except ValueError:
                return None
            return (
                self.db.query(Model)
                .filter(Model.id == model_id, Model.is_local == True)
                .first()
            )

        # supplier/name 格式（本地模型）
        if "/" in model_name:
            parts = model_name.split("/", 1)
            if len(parts) == 2:
                supplier, name = parts
                local_with_supplier = (
                    self.db.query(Model)
                    .filter(
                        Model.supplier == supplier,
                        Model.display_name == name,
                        Model.is_local == True,
                    )
                    .order_by(Model.created_at.desc())
                    .first()
                )
                if local_with_supplier:
                    return local_with_supplier

        # 云端模型按 model_code 精确匹配
        cloud = (
            self.db.query(Model)
            .filter(Model.model_code == model_name, Model.is_local == False)
            .first()
        )
        if cloud:
            return cloud

        # 本地模型按 display_name 查找
        return (
            self.db.query(Model)
            .filter(Model.display_name == model_name, Model.is_local == True)
            .order_by(Model.created_at.desc())
            .first()
        )


    def calculate_bill(
        self, model: Model, input_tokens: int, output_tokens: int
    ) -> Decimal:
        """API 调用免费，始终返回 0"""
        return Decimal("0")

    def extract_usage_from_response(self, response: Dict[str, Any]) -> Tuple[int, int]:
        usage = response.get("usage", {})
        input_tokens = usage.get("prompt_tokens", 0)
        output_tokens = usage.get("completion_tokens", 0)
        return input_tokens, output_tokens

    async def _validate_model_and_balance(self, model_name: str, user_id: str) -> Model:
        """
        验证模型状态

        Args:
            model_name: 模型名称
            user_id: 用户ID

        Returns:
            Model: 验证通过的模型对象

        Raises:
            OpenAICompatError: 模型不存在或不可用时抛出
        """
        model = self._find_model(model_name, user_id)

        if not model:
            raise OpenAICompatError(
                status_code=404,
                message=f"Model '{model_name}' not found",
                error_type="model_not_found"
            )

        if not model.is_local and not model.is_api_enabled:
            raise OpenAICompatError(
                status_code=403,
                message=f"Model '{model_name}' does not support API calls",
                error_type="model_not_available"
            )

        return model
    
    async def embeddings(
        self,
        request: EmbeddingsRequest,
        user_id: str,
        api_key_id: int,
        request_ip: Optional[str] = None
    ) -> EmbeddingsResponse:
        """
        处理文本嵌入(Embeddings)请求
        
        Args:
            request: 嵌入请求对象
            user_id: 用户ID
            api_key_id: 用于日志记录
            request_ip: 请求来源IP
            
        Returns:
            EmbeddingsResponse: 包含向量数据的响应对象
            
        Raises:
            OpenAICompatError: 处理失败或鉴权失败时抛出
            
        """
        log_service = ApiCallLogService(self.db)
        
        try:
            model = await self._validate_model_and_balance(request.model, user_id)
            
            # 检查模型类型是否支持向量嵌入
            from app.models.model import ModelCategory
            if not any(int(c) in [int(ModelCategory.EMBEDDING), int(ModelCategory.LLM), int(ModelCategory.MULTIMODAL)] for c in (model.categories or [])):
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{request.model}' does not support embeddings",
                    error_type="model_not_supported",
                )
            
            # 检查本地模型权限
            if model.is_local:
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                if not await permission_service.check_user_model_access(user_id, model.id):
                    raise OpenAICompatError(
                        status_code=403,
                        message=f"You don't have permission to access model '{request.model}'",
                        error_type="permission_error",
                    )
            else:
                user = self._get_user(user_id)
                if not user:
                    raise OpenAICompatError(
                        status_code=403, 
                        message="User API key not configured.", 
                        error_type="api_key_not_configured"
                    )
                # 优先使用模型自带的 api_key（爬虫同步的），没有则 fallback 到用户自己配置的 apikey
                effective_api_key: Optional[str] = None
                if model.encrypted_api_key:
                    from app.services.crypto_utils import decrypt_api_key as _decrypt
                    decrypted = _decrypt(model.encrypted_api_key)
                    effective_api_key = decrypted if decrypted else None
                if not effective_api_key:
                    effective_api_key = user.apikey
                if not effective_api_key:
                    raise OpenAICompatError(
                        status_code=403, 
                        message="User API key not configured.", 
                        error_type="api_key_not_configured"
                    )

            texts = [request.input] if isinstance(request.input, str) else request.input
            
            if model.is_local:
                # 从缓存获取模型信息
                from app.services.cache_service import CacheService
                from app.services.crypto_utils import decrypt_api_key
                model_data = await CacheService.get_model(model.id)
                
                if model_data:
                    base_url = model_data.get("base_url", "").rstrip("/")
                    local_api_key = model_data.get("local_api_key")
                else:
                    # 从数据库获取
                    base_url = (model.base_url or "").rstrip("/")
                    local_api_key = model.local_api_key
                    # 缓存模型信息
                    await CacheService.set_model(model.id, {
                        "base_url": base_url,
                        "local_api_key": local_api_key,
                        "is_local": model.is_local,
                        "name": model.name
                    })

                if not base_url:
                    raise OpenAICompatError(
                        status_code=500,
                        message="本地模型未配置 Base URL",
                        error_type="configuration_error",
                    )

                # 构建请求头
                headers = {"Content-Type": "application/json"}
                if local_api_key:
                    api_key = decrypt_api_key(local_api_key)
                    if api_key:
                        headers["Authorization"] = f"Bearer {api_key}"
                
                # 构建请求体（OpenAI 格式）
                payload = {
                    "model": model.name,
                    "input": texts
                }
                if request.dimensions:
                    payload["dimensions"] = request.dimensions
                if request.user:
                    payload["user"] = request.user

                # 使用统一的方法发送请求
                api_url = f"{base_url}/embeddings"
                result_data = await self._handle_local_model_request(
                    api_url=api_url,
                    headers=headers,
                    payload=payload,
                    model_name=model.name,
                    base_url=base_url,
                    endpoint_type="embedding",
                    timeout=30.0
                )

                # 处理响应
                embeddings_list = result_data.get("data", [])
                usage_data = result_data.get("usage", {})
                total_tokens = usage_data.get("total_tokens", 0)
                
                data_list = []
                for item in embeddings_list:
                    data_list.append(
                        EmbeddingData(
                            index=item.get("index", 0),
                            embedding=item.get("embedding", [])
                        )
                    )
            else:
                # 云端模型处理
                # 根据模型类型选择端点：多模态 embedding vs 文本 embedding
                # 多模态模型：名称含 vl/vision/multimodal
                code_lower = model.model_code.lower()
                is_multimodal = any(kw in code_lower for kw in ("vl", "vision", "multimodal"))
                if is_multimodal:
                    api_url = "https://dashscope.aliyuncs.com/api/v1/services/embeddings/multimodal-embedding/multimodal-embedding"
                    payload = {
                        "model": model.model_code,
                        "input": {
                            "contents": [{"text": t} for t in texts]
                        }
                    }
                else:
                    api_url = "https://dashscope.aliyuncs.com/api/v1/services/embeddings/text-embedding/text-embedding"
                    payload = {
                        "model": model.model_code,
                        "input": {
                            "texts": texts
                        }
                    }
                    if request.dimensions:
                        payload.setdefault("parameters", {})["dimension"] = request.dimensions
                headers = {
                    "Content-Type": "application/json",
                    "Authorization": f"Bearer {effective_api_key}"
                }
                # 多模态 embedding 的 dimension 放在 parameters 里
                if is_multimodal and request.dimensions:
                    payload.setdefault("parameters", {})["dimension"] = request.dimensions

                async with httpx.AsyncClient(timeout=30.0) as client:
                    response = await client.post(api_url, headers=headers, json=payload)
                    response.raise_for_status()
                    result_data = response.json()

                output = result_data.get("output", {})
                embeddings_list = output.get("embeddings", [])
                usage_data = result_data.get("usage", {})
                total_tokens = usage_data.get("total_tokens", 0)
                
                data_list = []
                for item in embeddings_list:
                    data_list.append(
                        EmbeddingData(
                            index=item.get("text_index", 0),
                            embedding=item.get("embedding", [])
                        )
                    )

            bill = Decimal("0")

            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=request.model,
                is_local=model.is_local if model else False,
                input_tokens=total_tokens,
                output_tokens=0,
                bill=float(bill),
                status="success",
                request_ip=request_ip
            )

            return EmbeddingsResponse(
                model=request.model,
                data=data_list,
                usage=Usage(
                    prompt_tokens=total_tokens,
                    completion_tokens=0,
                    total_tokens=total_tokens
                )
            )
            
        except Exception as e:
            error_msg = str(e) if not isinstance(e, OpenAICompatError) else e.message
            model = self._find_model(request.model, user_id)
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=request.model,
                is_local=model.is_local if model else False,
                input_tokens=0,
                output_tokens=0,
                bill=0,
                status="failed",
                error_message=error_msg,
                request_ip=request_ip
            )
            if isinstance(e, OpenAICompatError):
                raise e
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="embeddings_error")
    
    async def image_generations(
        self, request: ImageGenerationRequest, user_id: str, api_key_id: int, request_ip: str = None
    ) -> ImageGenerationResponse:
        """
        处理图像生成请求
        
        调用底层ImageGenerationService完成真实的图像生成与业务扣费。
        
        Args:
            request: 图像生成请求对象
            user_id: 用户ID
            api_key_id: API Key ID
            request_ip: 请求来源IP
            
        Returns:
            ImageGenerationResponse: 符合OpenAI规范的图像响应对象
            
        Raises:
            OpenAICompatError: 处理失败或鉴权失败时抛出
        """
        log_service = ApiCallLogService(self.db)
        
        try:
            # 验证模型状态与用户基础余额
            model = await self._validate_model_and_balance(request.model, user_id)
            
            # 检查模型类型是否支持图像生成（文生图）
            from app.models.model import ModelCategory
            if not any(int(c) in [int(ModelCategory.IMAGE_GEN), int(ModelCategory.MULTIMODAL)] for c in (model.categories or [])):
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{request.model}' does not support image generation. Use a model with category IMAGE_GEN or MULTIMODAL.",
                    error_type="model_not_supported",
                )
            
            # 检查本地模型权限
            if model.is_local:
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                if not await permission_service.check_user_model_access(user_id, model.id):
                    raise OpenAICompatError(
                        status_code=403,
                        message=f"You don't have permission to access model '{request.model}'",
                        error_type="permission_error",
                    )
            
            if model.is_local:
                # 从缓存获取模型信息
                from app.services.cache_service import CacheService
                from app.services.crypto_utils import decrypt_api_key
                model_data = await CacheService.get_model(model.id)
                
                if model_data:
                    base_url = model_data.get("base_url", "").rstrip("/")
                    local_api_key = model_data.get("local_api_key")
                else:
                    # 从数据库获取
                    base_url = (model.base_url or "").rstrip("/")
                    local_api_key = model.local_api_key
                    # 缓存模型信息
                    await CacheService.set_model(model.id, {
                        "base_url": base_url,
                        "local_api_key": local_api_key,
                        "is_local": model.is_local,
                        "name": model.name
                    })

                if not base_url:
                    raise OpenAICompatError(
                        status_code=500,
                        message="本地模型未配置 Base URL",
                        error_type="configuration_error",
                    )

                # 构建请求头
                headers = {"Content-Type": "application/json"}
                if local_api_key:
                    try:
                        api_key = decrypt_api_key(local_api_key)
                        if api_key:
                            headers["Authorization"] = f"Bearer {api_key}"
                        else:
                            raise OpenAICompatError(
                                status_code=500,
                                message="本地模型 API Key 解密失败",
                                error_type="configuration_error",
                            )
                    except Exception as e:
                        raise OpenAICompatError(
                            status_code=500,
                            message=f"本地模型 API Key 处理失败: {str(e)}",
                            error_type="configuration_error",
                        )
                
                # 构建请求体（OpenAI 格式）
                payload = {
                    "model": model.name,
                    "prompt": request.prompt,
                    "n": request.n or 1,
                    "size": request.size or "1024x1024"
                }
                if request.quality:
                    payload["quality"] = request.quality
                if request.response_format:
                    payload["response_format"] = request.response_format
                if request.style:
                    payload["style"] = request.style
                if request.user:
                    payload["user"] = request.user

                # 使用统一的方法发送请求
                api_url = f"{base_url}/images/generations"
                result_data = await self._handle_local_model_request(
                    api_url=api_url,
                    headers=headers,
                    payload=payload,
                    model_name=model.name,
                    base_url=base_url,
                    endpoint_type="image",
                    timeout=60.0
                )

                # 处理响应
                images = []
                for item in result_data.get("data", []):
                    if item.get("url"):
                        images.append(item.get("url"))
                    elif item.get("b64_json"):
                        # 处理base64编码的图像
                        import base64
                        from app.services.oss_service import get_oss_service
                        oss_service = get_oss_service()
                        image_bytes = base64.b64decode(item.get("b64_json"))
                        url = oss_service.upload_file(
                            image_bytes, 
                            prefix="ai-images/generations", 
                            original_filename=f"generated_{time.time()}.png"
                        )
                        images.append(url)

                if not images:
                    raise OpenAICompatError(
                        status_code=500,
                        message="图像生成失败：未返回图像",
                        error_type="image_generation_error",
                    )

                result = ImageGenerationResponse(
                    created=int(time.time()),
                    data=[ImageData(url=url) for url in images]
                )

                bill = Decimal("0")
            else:
                # 云端模型处理
                # 获取用户API Key并实例化底层服务
                user = self._get_user(user_id)
                dashscope_api_key = user.apikey if user and user.apikey else ""
                real_image_service = ImageGenerationService(self.db, api_key=dashscope_api_key)
                
                # 适配尺寸参数
                mapped_size = request.size.replace("x", "*") if request.size else "1024*1024"
                
                # 调用底层图像生成服务
                result_obj = await real_image_service.text_to_image(
                    user_id=user_id,
                    prompt=request.prompt,
                    model=model.model_code,
                    n=request.n or 1,
                    size=mapped_size
                )
                
                if not result_obj.success:
                    raise OpenAICompatError(
                        status_code=500, 
                        message=result_obj.error or "图像生成失败", 
                        error_type="image_generation_error"
                    )
                
                result = ImageGenerationResponse(
                    created=int(time.time()),
                    data=[ImageData(url=url) for url in result_obj.images]
                )

                bill = result_obj.bill
            
            # 记录日志
            log_service.create_log(
                user_id=user_id, 
                api_key_id=api_key_id, 
                model_id=model.id if model else None,
                model_name=request.model, 
                is_local=model.is_local if model else False,
                input_tokens=0, 
                output_tokens=len(result.data), 
                bill=float(bill),
                status="success", 
                request_ip=request_ip
            )
            
            return result
            
        except Exception as e:
            error_msg = str(e) if not isinstance(e, OpenAICompatError) else e.message
            model = self._find_model(request.model, user_id)
            log_service.create_log(
                user_id=user_id, 
                api_key_id=api_key_id, 
                model_id=model.id if model else None,
                model_name=request.model, 
                is_local=model.is_local if model else False,
                input_tokens=0, 
                output_tokens=0, 
                bill=0,
                status="failed", 
                error_message=error_msg, 
                request_ip=request_ip
            )
            if isinstance(e, OpenAICompatError): 
                raise e
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="image_generation_error")

    async def image_edits(
        self,
        image: Union[str, UploadFile],
        prompt: str,
        mask: Optional[Union[str, UploadFile]],
        model_name: str,
        n: int,
        size: str,
        user_id: str,
        api_key_id: int,
        request_ip: Optional[str] = None
    ) -> ImageGenerationResponse:
        """
        处理图像编辑(图生图)请求
        
        接收上传图片，转存OSS后调用底层ImageGenerationService处理。
        
        Args:
            image: 用户上传的原始图片
            prompt: 对新图像的文本描述
            mask: 可选的遮罩图
            model_name: 模型名称
            n: 生成数量
            size: 生成尺寸
            user_id: 用户ID
            api_key_id: API Key ID
            request_ip: 请求来源IP
            
        Returns:
            ImageGenerationResponse: 包含生成图片URL的响应对象
            
        Raises:
            OpenAICompatError: 处理失败或鉴权失败时抛出
            
        需求: OpenAI兼容-图生图
        """
        log_service = ApiCallLogService(self.db)
        
        try:
            model = await self._validate_model_and_balance(model_name, user_id)
            
            # 检查模型类型是否支持图像编辑（图生图）
            from app.models.model import ModelCategory
            if not any(int(c) in [int(ModelCategory.IMAGE_EDIT), int(ModelCategory.MULTIMODAL)] for c in (model.categories or [])):
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{model_name}' does not support image editing. Use a model with category IMAGE_EDIT or MULTIMODAL.",
                    error_type="model_not_supported",
                )
            
            # 检查本地模型权限
            if model.is_local:
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                if not await permission_service.check_user_model_access(user_id, model.id):
                    raise OpenAICompatError(
                        status_code=403,
                        message=f"You don't have permission to access model '{model_name}'",
                        error_type="permission_error",
                    )
            
            import base64
            from app.services.oss_service import get_oss_service
            oss_service = get_oss_service()
            
            # 处理图像数据
            if isinstance(image, str):
                # 检查是否是URL
                if image.startswith(('http://', 'https://')):
                    # 直接使用URL
                    image_url = image
                else:
                    # 解码base64字符串
                    image_bytes = base64.b64decode(image)
                    image_url = oss_service.upload_file(
                        image_bytes, 
                        prefix="ai-images/edits", 
                        original_filename="edit_source.png"
                    )
            else:
                # 处理UploadFile对象
                image_bytes = await image.read()
                image_url = oss_service.upload_file(
                    image_bytes, 
                    prefix="ai-images/edits", 
                    original_filename=image.filename or "edit_source.png"
                )
            image_urls = [image_url]
            
            # 处理遮罩数据
            if mask:
                if isinstance(mask, str):
                    # 检查是否是URL
                    if mask.startswith(('http://', 'https://')):
                        # 直接使用URL
                        mask_url = mask
                    else:
                        # 解码base64字符串
                        mask_bytes = base64.b64decode(mask)
                        mask_url = oss_service.upload_file(
                            mask_bytes, 
                            prefix="ai-images/edits", 
                            original_filename="edit_mask.png"
                        )
                else:
                    # 处理UploadFile对象
                    mask_bytes = await mask.read()
                    mask_url = oss_service.upload_file(
                        mask_bytes, 
                        prefix="ai-images/edits", 
                        original_filename=mask.filename or "edit_mask.png"
                    )
                image_urls.append(mask_url)
            
            if model.is_local:
                # 从缓存获取模型信息
                from app.services.cache_service import CacheService
                from app.services.crypto_utils import decrypt_api_key
                model_data = await CacheService.get_model(model.id)
                
                if model_data:
                    base_url = model_data.get("base_url", "").rstrip("/")
                    local_api_key = model_data.get("local_api_key")
                else:
                    # 从数据库获取
                    base_url = (model.base_url or "").rstrip("/")
                    local_api_key = model.local_api_key
                    # 缓存模型信息
                    await CacheService.set_model(model.id, {
                        "base_url": base_url,
                        "local_api_key": local_api_key,
                        "is_local": model.is_local,
                        "name": model.name
                    })

                if not base_url:
                    raise OpenAICompatError(
                        status_code=500,
                        message="本地模型未配置 Base URL",
                        error_type="configuration_error",
                    )

                # 构建请求头
                headers = {"Content-Type": "application/json"}
                if local_api_key:
                    api_key = decrypt_api_key(local_api_key)
                    if api_key:
                        headers["Authorization"] = f"Bearer {api_key}"
                
                # 构建请求体（OpenAI 格式）
                payload = {
                    "model": model.name,
                    "prompt": prompt,
                    "n": n,
                    "size": size or "1024x1024"
                }
                
                # 处理图像和遮罩
                if len(image_urls) == 1:
                    payload["image"] = image_urls[0]
                elif len(image_urls) == 2:
                    payload["image"] = image_urls[0]
                    payload["mask"] = image_urls[1]
                
                if prompt:
                    payload["prompt"] = prompt

                # 使用统一的方法发送请求
                api_url = f"{base_url}/images/edits"
                result_data = await self._handle_local_model_request(
                    api_url=api_url,
                    headers=headers,
                    payload=payload,
                    model_name=model.name,
                    base_url=base_url,
                    endpoint_type="image",
                    timeout=60.0
                )

                # 处理响应
                images = []
                for item in result_data.get("data", []):
                    if item.get("url"):
                        images.append(item.get("url"))
                    elif item.get("b64_json"):
                        # 处理base64编码的图像
                        image_bytes = base64.b64decode(item.get("b64_json"))
                        url = oss_service.upload_file(
                            image_bytes, 
                            prefix="ai-images/edits", 
                            original_filename=f"edited_{time.time()}.png"
                        )
                        images.append(url)

                if not images:
                    raise OpenAICompatError(
                        status_code=500,
                        message="图像编辑失败：未返回图像",
                        error_type="image_edit_error",
                    )

                result = ImageGenerationResponse(
                    created=int(time.time()),
                    data=[ImageData(url=url) for url in images]
                )

                bill = Decimal("0")
            else:
                # 云端模型处理
                user = self._get_user(user_id)
                dashscope_api_key = user.apikey if user and user.apikey else ""
                
                from app.services.image_service import ImageGenerationService
                real_image_service = ImageGenerationService(self.db, api_key=dashscope_api_key)
                
                mapped_size = size.replace("x", "*") if size else "1024*1024"
                
                result_obj = await real_image_service.image_to_image(
                    user_id=user_id,
                    image_urls=image_urls,
                    prompt=prompt,
                    model=model.model_code,
                    n=n,
                    size=mapped_size
                )
                
                if not result_obj.success:
                    raise OpenAICompatError(
                        status_code=500, 
                        message=result_obj.error or "图生图编辑失败", 
                        error_type="image_edit_error"
                    )
                    
                result = ImageGenerationResponse(
                    created=int(time.time()),
                    data=[ImageData(url=url) for url in result_obj.images]
                )

                bill = result_obj.bill
            
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=model_name,
                is_local=model.is_local if model else False,
                input_tokens=0,
                output_tokens=len(result.data),
                bill=float(bill),
                status="success",
                request_ip=request_ip
            )
            
            return result
            
        except Exception as e:
            error_msg = str(e) if not isinstance(e, OpenAICompatError) else e.message
            model_obj = self._find_model(model_name, user_id)
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model_obj.id if model_obj else None,
                model_name=model_name,
                is_local=model_obj.is_local if model_obj else False,
                input_tokens=0,
                output_tokens=0,
                bill=0,
                status="failed",
                error_message=error_msg,
                request_ip=request_ip
            )
            if isinstance(e, OpenAICompatError):
                raise e
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="image_edit_error")
        
    async def audio_transcriptions(
        self,
        file: Union[str, UploadFile],
        model_name: str,
        language: Optional[str],
        user_id: str,
        api_key_id: int,
        request_ip: Optional[str] = None
    ) -> AudioTranscriptionResponse:
        """
        处理语音识别(STT)请求
        
        接收上传的音频文件，转换为Base64编码后调用底层ASRService处理。
        包含模型名称从OpenAI(whisper-1)到DashScope原生模型的映射。
        
        Args:
            file: 客户端上传的音频文件
            model_name: 模型名称
            language: 语言代码 (ISO-639-1)
            user_id: 用户ID
            api_key_id: API Key ID
            request_ip: 请求来源IP
            
        Returns:
            AudioTranscriptionResponse: 包含识别文本的响应对象
            
        Raises:
            OpenAICompatError: 处理失败或鉴权失败时抛出
            
        需求: OpenAI兼容-语音转文字
        """
        log_service = ApiCallLogService(self.db)
        
        try:
            actual_model = "qwen3-asr-flash" if model_name in ["whisper-1", "whisper-large-v3"] else model_name

            # realtime 模型使用 WebSocket 实时流协议，不支持此文件上传接口
            if "realtime" in actual_model.lower():
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{actual_model}' is a real-time streaming model that uses WebSocket protocol. "
                            f"It cannot be called via /api/v1/audio/transcriptions. "
                            f"Please use the WebSocket API instead.",
                    error_type="model_not_supported",
                )

            model = await self._validate_model_and_balance(actual_model, user_id)
            
            # 检查模型类型是否支持语音识别
            from app.models.model import ModelCategory
            if not any(int(c) in [int(ModelCategory.STT), int(ModelCategory.MULTIMODAL)] for c in (model.categories or [])):
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{model_name}' does not support speech transcription",
                    error_type="model_not_supported",
                )
            
            # 检查本地模型权限
            if model.is_local:
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                if not await permission_service.check_user_model_access(user_id, model.id):
                    raise OpenAICompatError(
                        status_code=403,
                        message=f"You don't have permission to access model '{model_name}'",
                        error_type="permission_error",
                    )
            
            import base64
            import httpx
            if isinstance(file, str):
                # 检查是否是URL
                if file.startswith(('http://', 'https://')):
                    # 从URL下载音频文件并转换为base64
                    async with httpx.AsyncClient() as client:
                        response = await client.get(file)
                        response.raise_for_status()
                        audio_bytes = response.content
                        audio_base64 = base64.b64encode(audio_bytes).decode('utf-8')
                else:
                    # 使用base64字符串
                    audio_base64 = file
            else:
                # 处理UploadFile对象
                audio_bytes = await file.read()
                audio_base64 = base64.b64encode(audio_bytes).decode('utf-8')
            
            if model.is_local:
                # 从缓存获取模型信息
                from app.services.cache_service import CacheService
                from app.services.crypto_utils import decrypt_api_key
                model_data = await CacheService.get_model(model.id)
                
                if model_data:
                    base_url = model_data.get("base_url", "").rstrip("/")
                    local_api_key = model_data.get("local_api_key")
                else:
                    # 从数据库获取
                    base_url = (model.base_url or "").rstrip("/")
                    local_api_key = model.local_api_key
                    # 缓存模型信息
                    await CacheService.set_model(model.id, {
                        "base_url": base_url,
                        "local_api_key": local_api_key,
                        "is_local": model.is_local,
                        "name": model.name
                    })

                if not base_url:
                    raise OpenAICompatError(
                        status_code=500,
                        message="本地模型未配置 Base URL",
                        error_type="configuration_error",
                    )

                # 构建请求头
                headers = {"Content-Type": "application/json"}
                if local_api_key:
                    api_key = decrypt_api_key(local_api_key)
                    if api_key:
                        headers["Authorization"] = f"Bearer {api_key}"
                
                # 构建请求体（OpenAI 格式）
                payload = {
                    "model": model.name,
                    "file": audio_base64
                }
                if language:
                    payload["language"] = language

                # 使用统一的方法发送请求
                api_url = f"{base_url}/audio/transcriptions"
                result_data = await self._handle_local_model_request(
                    api_url=api_url,
                    headers=headers,
                    payload=payload,
                    model_name=model.name,
                    base_url=base_url,
                    endpoint_type="audio_stt",
                    timeout=60.0
                )

                # 处理响应
                text = result_data.get("text", "")
                duration_seconds = len(audio_base64) // 10000  # 粗略估算
                text_length = len(text)

                bill = Decimal("0")
            else:
                # 云端模型处理
                user = self._get_user(user_id)
                dashscope_api_key = user.apikey if user and user.apikey else ""
                
                from app.services.asr_service import ASRService
                from app.schemas.audio_schema import ASRRequest
                
                real_asr_service = ASRService(self.db, user_id=user_id, api_key=dashscope_api_key)
                internal_req = ASRRequest(
                    model=actual_model,
                    audio_base64=audio_base64,
                    language=language
                )
                
                asr_response = await real_asr_service.recognize(internal_req)
                
                text = asr_response.text
                duration_seconds = asr_response.usage.seconds if asr_response.usage else 0
                text_length = len(text) if text else 0

                bill = Decimal("0")
            
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=model_name,
                is_local=model.is_local if model else False,
                input_tokens=duration_seconds,
                output_tokens=text_length,
                bill=float(bill),
                status="success",
                request_ip=request_ip
            )
            
            return AudioTranscriptionResponse(text=text)
            
        except Exception as e:
            error_msg = str(e) if not isinstance(e, OpenAICompatError) else e.message
            model_obj = self._find_model(model_name, user_id)
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model_obj.id if model_obj else None,
                model_name=model_name,
                is_local=model_obj.is_local if model_obj else False,
                input_tokens=0,
                output_tokens=0,
                bill=0,
                status="failed",
                error_message=error_msg,
                request_ip=request_ip
            )
            if isinstance(e, OpenAICompatError):
                raise e
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="stt_error")
    
    async def audio_speech(
        self,
        request: AudioSpeechRequest,
        user_id: str,
        api_key_id: int,
        request_ip: Optional[str] = None
    ) -> Tuple[AsyncGenerator, str]:
        """
        处理文字转语音(TTS)请求
        
        执行全量音色映射与模型能力校验。
        通过底层 TTSService 生成语音并转存 OSS，随后转换为流式下发。
        
        Args:
            request: TTS请求对象
            user_id: 用户ID
            api_key_id: API Key ID
            request_ip: 请求来源IP
            
        Returns:
            Tuple[AsyncGenerator, str]: 音频二进制生成器与 MIME 类型
            
        Raises:
            OpenAICompatError: 模型不支持所选音色或处理失败时抛出
            
        需求: OpenAI兼容-文字转语音
        """
        log_service = ApiCallLogService(self.db)
        
        try:
            actual_model = "cosyvoice-v3-flash" if request.model in ["tts-1", "tts-1-hd"] else request.model

            # realtime 模型使用 WebSocket 实时流协议，不支持此 REST 接口
            if "realtime" in actual_model.lower():
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{actual_model}' is a real-time streaming model that uses WebSocket protocol. "
                            f"It cannot be called via /api/v1/audio/speech. "
                            f"Please use the WebSocket API instead.",
                    error_type="model_not_supported",
                )

            # cosyvoice-clone 系列：voice 参数就是 voice_id，不做映射，直接透传
            is_clone = "clone" in actual_model.lower()

            voice_map = {
                "alloy": "longxiaochun_v3",
                "echo": "longcheng_v3",
                "shimmer": "longwan_v3",
                "onyx": "longhua_v3",
                "nova": "longxiaoxia_v3",
                "fable": "longshu_v3",
                "sunny": "longanyang",
                "lively": "longanhuan",
                "cute_girl": "longhuhu_v3",
                "cute_boy": "longniuniu_v3",
                "bubble": "longpaopao_v3",
                "naughty": "longjielidou_v3",
                "bold_girl": "longxian_v3",
                "cantonese_f": "longjiaxin_v3",
                "cantonese_m": "longanyue_v3",
                "dongbei": "longlaotie_v3",
                "shanbei": "longshange_v3",
                "korean": "loongkyong_v3",
                "japanese": "loongriko_v3",
                "news_m": "longfei_v3",
                "news_f": "longxiaoxia_v3",
                "story_m": "longxiu_v3",
                "story_f": "longmiao_v3",
                "customer_service": "longyingxiao_v3",
                "monkey": "longhouge_v3",
                "robot": "longjiqi_v3",
                "daiyu": "longdaiyu_v3",
                "uncle": "longlaobo_v3",
                "aunt": "longlaoyi_v3"
            }
            
            actual_voice = request.voice if is_clone else voice_map.get(request.voice.lower(), request.voice)

            if "plus" in actual_model.lower():
                plus_allowed_voices = ["longanyang", "longanhuan"]
                if actual_voice not in plus_allowed_voices:
                    raise OpenAICompatError(
                        status_code=400,
                        message=f"Model '{actual_model}' only supports voices: {plus_allowed_voices}. Requested: '{actual_voice}'.",
                        error_type="invalid_request_error"
                    )

            model = await self._validate_model_and_balance(actual_model, user_id)
            
            # 检查模型类型是否支持语音合成
            from app.models.model import ModelCategory
            if not any(int(c) in [int(ModelCategory.TTS), int(ModelCategory.MULTIMODAL)] for c in (model.categories or [])):
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{request.model}' does not support speech synthesis",
                    error_type="model_not_supported",
                )
            
            # 检查本地模型权限
            if model.is_local:
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                if not await permission_service.check_user_model_access(user_id, model.id):
                    raise OpenAICompatError(
                        status_code=403,
                        message=f"You don't have permission to access model '{request.model}'",
                        error_type="permission_error",
                    )
            
            if model.is_local:
                # 从缓存获取模型信息
                from app.services.cache_service import CacheService
                from app.services.crypto_utils import decrypt_api_key
                model_data = await CacheService.get_model(model.id)
                
                if model_data:
                    base_url = model_data.get("base_url", "").rstrip("/")
                    local_api_key = model_data.get("local_api_key")
                else:
                    # 从数据库获取
                    base_url = (model.base_url or "").rstrip("/")
                    local_api_key = model.local_api_key
                    # 缓存模型信息
                    await CacheService.set_model(model.id, {
                        "base_url": base_url,
                        "local_api_key": local_api_key,
                        "is_local": model.is_local,
                        "name": model.name
                    })

                if not base_url:
                    raise OpenAICompatError(
                        status_code=500,
                        message="本地模型未配置 Base URL",
                        error_type="configuration_error",
                    )

                # 构建请求头
                headers = {"Content-Type": "application/json"}
                if local_api_key:
                    api_key = decrypt_api_key(local_api_key)
                    if api_key:
                        headers["Authorization"] = f"Bearer {api_key}"
                
                # 构建请求体（OpenAI 格式）
                # 注意：本地模型使用原始的 OpenAI 音色名称，不进行映射
                payload = {
                    "model": model.name,
                    "input": request.input,
                    "voice": request.voice,  # 使用原始音色名称，不映射
                    "response_format": request.response_format or "mp3",
                    "speed": request.speed or 1.0
                }

                # 使用统一的适配器方法发送请求（获取原始响应）
                api_url = f"{base_url}/audio/speech"
                response = await self._handle_local_model_request(
                    api_url=api_url,
                    headers=headers,
                    payload=payload,
                    model_name=model.name,
                    base_url=base_url,
                    endpoint_type="audio_tts",
                    timeout=60.0,
                    return_raw_response=True
                )
                
                # 检查响应类型
                content_type = response.headers.get("content-type", "")
                
                import httpx
                if "application/json" in content_type:
                    # 响应是 JSON，包含音频 URL
                    result_data = response.json()
                    audio_url = result_data.get("audio_url") or result_data.get("url")
                    
                    if not audio_url:
                        raise OpenAICompatError(
                            status_code=500,
                            message="本地模型返回的 JSON 中未找到音频 URL",
                            error_type="invalid_response",
                        )
                    
                    # 下载音频并流式返回
                    async def generate_audio():
                        async with httpx.AsyncClient() as client:
                            async with client.stream("GET", audio_url) as audio_response:
                                audio_response.raise_for_status()
                                async for chunk in audio_response.aiter_bytes():
                                    yield chunk
                else:
                    # 响应直接是音频数据（如硅基流动）
                    audio_bytes = response.content
                    
                    async def generate_audio():
                        # 直接返回音频数据
                        yield audio_bytes

                media_type = f"audio/{request.response_format or 'mp3'}" if (request.response_format or 'mp3') != 'mp3' else "audio/mpeg"
                bill = Decimal("0")
            else:
                # 云端模型处理
                user = self._get_user(user_id)
                dashscope_api_key = user.apikey if user and user.apikey else ""
                
                from app.services.tts_service import TTSService
                from app.schemas.audio_schema import TTSRequest
                
                real_tts_service = TTSService(self.db, user_id=user_id, api_key=dashscope_api_key)
                
                internal_req = TTSRequest(
                    text=request.input,
                    model=actual_model,
                    voice=actual_voice,
                    format=request.response_format or "mp3",
                    sample_rate=24000
                )
                
                tts_response = await real_tts_service.synthesize(internal_req)
                
                import httpx
                async def generate_audio():
                    async with httpx.AsyncClient() as client:
                        async with client.stream("GET", tts_response.audio_url) as response:
                            response.raise_for_status()
                            async for chunk in response.aiter_bytes():
                                yield chunk
                                
                media_type = f"audio/{request.response_format}" if request.response_format != 'mp3' else "audio/mpeg"
                bill = Decimal("0")

            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=request.model,
                is_local=model.is_local if model else False,
                input_tokens=len(request.input),
                output_tokens=0,
                bill=float(bill),
                status="success",
                request_ip=request_ip
            )

            return generate_audio(), media_type
            
        except Exception as e:
            error_msg = str(e) if not isinstance(e, OpenAICompatError) else e.message
            model_obj = self._find_model(request.model, user_id)
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model_obj.id if model_obj else None,
                model_name=request.model,
                is_local=model_obj.is_local if model_obj else False,
                input_tokens=0,
                output_tokens=0,
                bill=0,
                status="failed",
                error_message=error_msg,
                request_ip=request_ip
            )
            if isinstance(e, OpenAICompatError):
                raise e
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="tts_error")
    
    async def video_generations(
        self,
        request: VideoGenerationRequest,
        user_id: str,
        api_key_id: int,
        request_ip: Optional[str] = None
    ) -> VideoGenerationResponse:
        """
        处理视频生成请求
        
        调用底层VideoService提交异步任务，并通过轮询将其封装为同步阻塞接口。
        
        Args:
            request: 视频生成请求对象
            user_id: 用户ID
            api_key_id: API Key ID
            request_ip: 请求来源IP
            
        Returns:
            VideoGenerationResponse: 包含最终视频URL的响应对象
            
        Raises:
            OpenAICompatError: 模型不支持或生成失败时抛出
            
        需求: OpenAI兼容-视频生成
        """
        import time
        import asyncio
        from app.services.video_service import VideoService
        from app.schemas.video_schema import VideoGenerateRequest
        
        log_service = ApiCallLogService(self.db)
        
        try:
            model = await self._validate_model_and_balance(request.model, user_id)
            
            # 检查模型类型是否支持视频生成
            from app.models.model import ModelCategory
            if not any(int(c) in [int(ModelCategory.VIDEO_GEN), int(ModelCategory.MULTIMODAL), int(ModelCategory.LLM)] for c in (model.categories or [])):
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{request.model}' does not support video generation",
                    error_type="model_not_supported",
                )
            
            # 检查本地模型权限
            if model.is_local:
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                if not await permission_service.check_user_model_access(user_id, model.id):
                    raise OpenAICompatError(
                        status_code=403,
                        message=f"You don't have permission to access model '{request.model}'",
                        error_type="permission_error",
                    )
            
            if model.is_local:
                # 从缓存获取模型信息
                from app.services.cache_service import CacheService
                from app.services.crypto_utils import decrypt_api_key
                model_data = await CacheService.get_model(model.id)
                
                if model_data:
                    base_url = model_data.get("base_url", "").rstrip("/")
                    local_api_key = model_data.get("local_api_key")
                else:
                    # 从数据库获取
                    base_url = (model.base_url or "").rstrip("/")
                    local_api_key = model.local_api_key
                    # 缓存模型信息
                    await CacheService.set_model(model.id, {
                        "base_url": base_url,
                        "local_api_key": local_api_key,
                        "is_local": model.is_local,
                        "name": model.name
                    })

                if not base_url:
                    raise OpenAICompatError(
                        status_code=500,
                        message="本地模型未配置 Base URL",
                        error_type="configuration_error",
                    )

                # 构建请求头
                headers = {"Content-Type": "application/json"}
                if local_api_key:
                    api_key = decrypt_api_key(local_api_key)
                    if api_key:
                        headers["Authorization"] = f"Bearer {api_key}"
                
                # 构建请求体（OpenAI 格式）
                payload = {
                    "model": model.name,
                    "prompt": request.prompt,
                    "size": request.size or "1280x720",  # 使用OpenAI标准格式
                    "duration": request.duration or 5
                }

                # 使用统一的适配器方法发送请求
                api_url = f"{base_url}/videos/generations"
                result_data = await self._handle_local_model_request(
                    api_url=api_url,
                    headers=headers,
                    payload=payload,
                    model_name=model.name,
                    base_url=base_url,
                    endpoint_type="video",
                    timeout=60.0
                )

                # 处理响应
                videos = []
                for item in result_data.get("data", []):
                    if item.get("url"):
                        videos.append(item.get("url"))

                if not videos:
                    raise OpenAICompatError(
                        status_code=500,
                        message="视频生成失败：未返回视频",
                        error_type="video_generation_error",
                    )

                # 组装结果
                from app.schemas.openai_compat import VideoData
                result = VideoGenerationResponse(
                    created=int(time.time()),
                    data=[VideoData(url=url, content_type="video/mp4") for url in videos]
                )

                bill = Decimal("0")
            else:
                # 云端模型处理
                user = self._get_user(user_id)
                from app.services.crypto_utils import get_effective_api_key
                dashscope_api_key = get_effective_api_key(self.db, request.model, user.apikey if user else "") if user else ""
                
                real_video_service = VideoService(self.db, user_id=int(user_id) if str(user_id).isdigit() else user_id, api_key=dashscope_api_key)
                
                # 解析并转换size格式
                try:
                    openai_size, internal_size = parse_video_size(request.size or "1280x720")
                except ValueError as e:
                    raise OpenAICompatError(
                        status_code=400,
                        message=str(e),
                        error_type="invalid_request_error"
                    )
                
                internal_req = VideoGenerateRequest(
                    prompt=request.prompt,
                    resolution=internal_size,  # 使用内部格式 "720P"
                    duration=request.duration or 5,
                    prompt_extend=True,
                    watermark=False
                )
                
                # 提交异步任务
                task_resp = await real_video_service.generate(internal_req)
                task_id = task_resp.task_id
                
                # 阻塞轮询
                max_retries = 120
                poll_interval = 5
                
                final_video_url = None
                
                for _ in range(max_retries):
                    await asyncio.sleep(poll_interval)
                    status_result = await real_video_service.get_task_status(task_id)
                    
                    if status_result.task_status == "SUCCEEDED":
                        final_video_url = status_result.video_url
                        break
                    elif status_result.task_status == "FAILED":
                        raise OpenAICompatError(
                            status_code=500,
                            message=status_result.error_message or "底层视频生成任务失败",
                            error_type="video_generation_error"
                        )
                
                if not final_video_url:
                    raise OpenAICompatError(
                        status_code=504,
                        message="视频生成任务超时，请稍后再试或通过平台任务列表查看结果",
                        error_type="timeout_error"
                    )
                
                # 组装结果
                from app.schemas.openai_compat import VideoData
                result = VideoGenerationResponse(
                    created=int(time.time()),
                    data=[VideoData(url=final_video_url, content_type="video/mp4")]
                )

                bill = Decimal("0")
            
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=request.model,
                is_local=model.is_local if model else False,
                input_tokens=0,
                output_tokens=request.duration or 5,
                bill=float(bill),
                status="success",
                request_ip=request_ip
            )
            
            return result
            
        except Exception as e:
            error_msg = str(e) if not isinstance(e, OpenAICompatError) else e.message
            model_obj = self._find_model(request.model, user_id)
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model_obj.id if model_obj else None,
                model_name=request.model,
                is_local=model_obj.is_local if model_obj else False,
                input_tokens=0,
                output_tokens=0,
                bill=0,
                status="failed",
                error_message=error_msg,
                request_ip=request_ip
            )
            if isinstance(e, OpenAICompatError):
                raise e
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="video_generation_error")
    
    async def image_to_video_generations(
        self,
        image: Union[str, UploadFile],
        prompt: str,
        model_name: str,
        size: str,
        user_id: str,
        api_key_id: int,
        request_ip: Optional[str] = None
    ) -> VideoGenerationResponse:
        """
        处理图生视频(I2V)请求
        
        接收上传图片，转存OSS获取URL，调用底层VideoService提交图生视频异步任务，
        并轮询任务状态封装为同步接口返回。
        
        Args:
            image: 客户端上传的参考图像
            prompt: 对视频的文本描述
            model_name: 模型名称 (如 wan2.6-i2v)
            size: 视频分辨率 (如 720P)
            user_id: 用户ID
            api_key_id: API Key ID
            request_ip: 请求来源IP
            
        Returns:
            VideoGenerationResponse: 包含最终视频URL的响应对象
            
        Raises:
            OpenAICompatError: 处理失败时抛出
            
        需求: OpenAI兼容-图生视频
        """
        import time
        import asyncio
        import base64
        from app.services.oss_service import get_oss_service
        
        log_service = ApiCallLogService(self.db)
        
        try:
            model = await self._validate_model_and_balance(model_name, user_id)
            
            # 检查模型类型是否支持视频生成
            from app.models.model import ModelCategory
            if not any(int(c) in [int(ModelCategory.VIDEO_GEN), int(ModelCategory.MULTIMODAL), int(ModelCategory.LLM)] for c in (model.categories or [])):
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{model_name}' does not support video generation",
                    error_type="model_not_supported",
                )
            
            # 检查本地模型权限
            if model.is_local:
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                if not await permission_service.check_user_model_access(user_id, model.id):
                    raise OpenAICompatError(
                        status_code=403,
                        message=f"You don't have permission to access model '{model_name}'",
                        error_type="permission_error",
                    )
            
            # 处理图像数据
            oss_service = get_oss_service()
            if isinstance(image, str):
                # 检查是否是URL
                if image.startswith(('http://', 'https://')):
                    # 直接使用URL
                    image_url = image
                else:
                    # 解码base64字符串
                    image_bytes = base64.b64decode(image)
                    image_url = oss_service.upload_file(
                        image_bytes, 
                        prefix="ai-videos/i2v-source", 
                        original_filename="i2v_source.png"
                    )
            else:
                # 处理UploadFile对象
                image_bytes = await image.read()
                image_url = oss_service.upload_file(
                    image_bytes, 
                    prefix="ai-videos/i2v-source", 
                    original_filename=image.filename or "i2v_source.png"
                )
            
            if model.is_local:
                # 从缓存获取模型信息
                from app.services.cache_service import CacheService
                from app.services.crypto_utils import decrypt_api_key
                model_data = await CacheService.get_model(model.id)
                
                if model_data:
                    base_url = model_data.get("base_url", "").rstrip("/")
                    local_api_key = model_data.get("local_api_key")
                else:
                    # 从数据库获取
                    base_url = (model.base_url or "").rstrip("/")
                    local_api_key = model.local_api_key
                    # 缓存模型信息
                    await CacheService.set_model(model.id, {
                        "base_url": base_url,
                        "local_api_key": local_api_key,
                        "is_local": model.is_local,
                        "name": model.name
                    })

                if not base_url:
                    raise OpenAICompatError(
                        status_code=500,
                        message="本地模型未配置 Base URL",
                        error_type="configuration_error",
                    )

                # 构建请求头
                headers = {"Content-Type": "application/json"}
                if local_api_key:
                    api_key = decrypt_api_key(local_api_key)
                    if api_key:
                        headers["Authorization"] = f"Bearer {api_key}"
                
                # 构建请求体（OpenAI 格式）
                payload = {
                    "model": model.name,
                    "prompt": prompt,
                    "image": image_url,
                    "size": size or "1280x720",  # 使用OpenAI标准格式
                    "duration": 5
                }

                # 使用统一的适配器方法发送请求
                api_url = f"{base_url}/videos/generations"
                result_data = await self._handle_local_model_request(
                    api_url=api_url,
                    headers=headers,
                    payload=payload,
                    model_name=model.name,
                    base_url=base_url,
                    endpoint_type="video",
                    timeout=60.0
                )

                # 处理响应
                videos = []
                for item in result_data.get("data", []):
                    if item.get("url"):
                        videos.append(item.get("url"))

                if not videos:
                    raise OpenAICompatError(
                        status_code=500,
                        message="图生视频失败：未返回视频",
                        error_type="video_generation_error",
                    )

                # 组装结果
                from app.schemas.openai_compat import VideoData
                result = VideoGenerationResponse(
                    created=int(time.time()),
                    data=[VideoData(url=url, content_type="video/mp4") for url in videos]
                )

                bill = Decimal("0")
            else:
                # 云端模型处理
                user = self._get_user(user_id)
                from app.services.crypto_utils import get_effective_api_key
                dashscope_api_key = get_effective_api_key(self.db, model_name, user.apikey if user else "") if user else ""
                
                from app.services.video_service import VideoService
                from app.schemas.video_schema import VideoGenerateRequest
                
                real_video_service = VideoService(
                    self.db, 
                    user_id=int(user_id) if str(user_id).isdigit() else user_id, 
                    api_key=dashscope_api_key
                )
                
                # 解析并转换size格式
                try:
                    openai_size, internal_size = parse_video_size(size or "1280x720")
                except ValueError as e:
                    raise OpenAICompatError(
                        status_code=400,
                        message=str(e),
                        error_type="invalid_request_error"
                    )
                
                internal_req = VideoGenerateRequest(
                    prompt=prompt,
                    first_frame_url=image_url,
                    resolution=internal_size,  # 使用内部格式 "720P"
                    duration=5,
                    prompt_extend=True,
                    watermark=False
                )
                
                # 提交异步任务
                task_resp = await real_video_service.generate(internal_req)
                task_id = task_resp.task_id
                
                # 阻塞轮询
                max_retries = 120
                poll_interval = 5
                final_video_url = None
                
                for _ in range(max_retries):
                    await asyncio.sleep(poll_interval)
                    status_result = await real_video_service.get_task_status(task_id)
                    
                    if status_result.task_status == "SUCCEEDED":
                        final_video_url = status_result.video_url
                        break
                    elif status_result.task_status == "FAILED":
                        raise OpenAICompatError(
                            status_code=500,
                            message=status_result.error_message or "底层图生视频任务失败",
                            error_type="video_generation_error"
                        )
                
                if not final_video_url:
                    raise OpenAICompatError(
                        status_code=504,
                        message="图生视频任务超时，请稍后再试",
                        error_type="timeout_error"
                    )
                
                # 组装结果
                from app.schemas.openai_compat import VideoData
                result = VideoGenerationResponse(
                    created=int(time.time()),
                    data=[VideoData(url=final_video_url, content_type="video/mp4")]
                )

                bill = Decimal("0")
            
            # 记录兼容层日志
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=model_name,
                is_local=model.is_local if model else False,
                input_tokens=0,
                output_tokens=5, 
                bill=float(bill),
                status="success",
                request_ip=request_ip
            )
            
            return result
            
        except Exception as e:
            error_msg = str(e) if not isinstance(e, OpenAICompatError) else e.message
            model_obj = self._find_model(model_name, user_id)
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model_obj.id if model_obj else None,
                model_name=model_name,
                is_local=model_obj.is_local if model_obj else False,
                input_tokens=0,
                output_tokens=0,
                bill=0,
                status="failed",
                error_message=error_msg,
                request_ip=request_ip
            )
            if isinstance(e, OpenAICompatError):
                raise e
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="video_generation_error")

    async def rerank(
        self,
        request: RerankRequest,
        user_id: str,
        api_key_id: int,
        request_ip: Optional[str] = None
    ) -> RerankResponse:
        """
        处理重排序(Rerank)请求
        
        Args:
            request: 重排序请求对象
            user_id: 用户ID
            api_key_id: 用于日志记录
            request_ip: 请求来源IP
            
        Returns:
            RerankResponse: 包含排序结果的响应对象
            
        Raises:
            OpenAICompatError: 处理失败或鉴权失败时抛出
        """
        log_service = ApiCallLogService(self.db)
        
        try:
            model = await self._validate_model_and_balance(request.model, user_id)
            
            # 检查模型类型是否支持重排序
            from app.models.model import ModelCategory
            if not any(int(c) in [int(ModelCategory.RERANK), int(ModelCategory.EMBEDDING), int(ModelCategory.LLM)] for c in (model.categories or [])):
                raise OpenAICompatError(
                    status_code=400,
                    message=f"Model '{request.model}' does not support reranking",
                    error_type="model_not_supported",
                )
            
            # 检查本地模型权限
            if model.is_local:
                from app.services.user_local_model_permission_service import UserLocalModelPermissionService
                permission_service = UserLocalModelPermissionService(self.db)
                if not await permission_service.check_user_model_access(user_id, model.id):
                    raise OpenAICompatError(
                        status_code=403,
                        message=f"You don't have permission to access model '{request.model}'",
                        error_type="permission_error",
                    )
            else:
                user = self._get_user(user_id)
                if not user or not user.apikey:
                    raise OpenAICompatError(
                        status_code=403, 
                        message="User API key not configured.", 
                        error_type="api_key_not_configured"
                    )

            if model.is_local:
                # 本地模型处理
                from app.services.cache_service import CacheService
                from app.services.crypto_utils import decrypt_api_key
                model_data = await CacheService.get_model(model.id)
                
                if model_data:
                    base_url = model_data.get("base_url", "").rstrip("/")
                    local_api_key = model_data.get("local_api_key")
                else:
                    # 从数据库获取
                    base_url = (model.base_url or "").rstrip("/")
                    local_api_key = model.local_api_key
                    # 缓存模型信息
                    await CacheService.set_model(model.id, {
                        "base_url": base_url,
                        "local_api_key": local_api_key,
                        "is_local": model.is_local,
                        "name": model.name
                    })

                if not base_url:
                    raise OpenAICompatError(
                        status_code=500,
                        message="本地模型未配置 Base URL",
                        error_type="configuration_error",
                    )

                # 构建请求头
                headers = {"Content-Type": "application/json"}
                if local_api_key:
                    api_key = decrypt_api_key(local_api_key)
                    if api_key:
                        headers["Authorization"] = f"Bearer {api_key}"
                
                # 构建请求体（OpenAI 格式）
                payload = {
                    "model": model.name,
                    "query": request.query,
                    "documents": request.documents
                }
                if request.top_n is not None:
                    payload["top_n"] = request.top_n
                if request.return_documents is not None:
                    payload["return_documents"] = request.return_documents
                if request.user:
                    payload["user"] = request.user

                # 使用统一的方法发送请求
                api_url = f"{base_url}/rerank"
                result_data = await self._handle_local_model_request(
                    api_url=api_url,
                    headers=headers,
                    payload=payload,
                    model_name=model.name,
                    base_url=base_url,
                    endpoint_type="rerank",
                    timeout=30.0
                )

                # 处理响应
                import logging
                logger = logging.getLogger(__name__)
                
                results_list = result_data.get("data", [])
                usage_data = result_data.get("usage", {})
                
                # 如果 data 为空，尝试从 results 字段获取（某些模型如硅基流动使用 results）
                if not results_list:
                    results_list = result_data.get("results", [])
                    logger.debug(f"Using 'results' field, found {len(results_list)} items")
                
                # 尝试从不同位置获取 token 信息
                if not usage_data or usage_data.get("total_tokens", 0) == 0:
                    # 尝试从 meta.tokens 获取（硅基流动格式）
                    meta = result_data.get("meta", {})
                    tokens = meta.get("tokens", {})
                    if tokens:
                        input_tokens = tokens.get("input_tokens", 0)
                        output_tokens = tokens.get("output_tokens", 0)
                        total_tokens = input_tokens + output_tokens
                        logger.debug(f"Using meta.tokens: input={input_tokens}, output={output_tokens}, total={total_tokens}")
                    else:
                        total_tokens = 0
                else:
                    total_tokens = usage_data.get("total_tokens", 0)
                
                data_list = []
                for item in results_list:
                    # 获取文档内容（处理不同格式）
                    doc_content = None
                    if request.return_documents:
                        doc = item.get("document")
                        if isinstance(doc, dict):
                            # 如果 document 是对象，尝试获取 text 字段
                            doc_content = doc.get("text", "")
                        elif isinstance(doc, str):
                            # 如果 document 是字符串，直接使用
                            doc_content = doc
                    
                    result_item = RerankResult(
                        index=item.get("index", 0),
                        relevance_score=item.get("relevance_score", 0.0)
                    )
                    if doc_content:
                        result_item.document = doc_content
                    data_list.append(result_item)
            else:
                # 云端模型处理（阿里云百炼）
                api_url = "https://dashscope.aliyuncs.com/api/v1/services/rerank/text-rerank/text-rerank"
                headers = {
                    "Content-Type": "application/json",
                    "Authorization": f"Bearer {user.apikey}"
                }
                
                payload = {
                    "model": model.model_code,
                    "input": {
                        "query": request.query,
                        "documents": request.documents
                    }
                }
                if request.top_n is not None:
                    payload["parameters"] = {"top_n": request.top_n}

                async with httpx.AsyncClient(timeout=30.0) as client:
                    response = await client.post(api_url, headers=headers, json=payload)
                    response.raise_for_status()
                    result_data = response.json()

                output = result_data.get("output", {})
                results_list = output.get("results", [])
                usage_data = result_data.get("usage", {})
                total_tokens = usage_data.get("total_tokens", 0)
                
                data_list = []
                for item in results_list:
                    result_item = RerankResult(
                        index=item.get("index", 0),
                        relevance_score=item.get("relevance_score", 0.0)
                    )
                    if request.return_documents:
                        result_item.document = request.documents[item.get("index", 0)]
                    data_list.append(result_item)

            # 记录日志
            bill = Decimal("0")
            call_log = log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=request.model,
                is_local=model.is_local if model else False,
                input_tokens=total_tokens,
                output_tokens=0,
                bill=float(bill),
                status="success",
                request_ip=request_ip
            )

            return RerankResponse(
                model=request.model,
                data=data_list,
                usage=Usage(
                    prompt_tokens=total_tokens,
                    completion_tokens=0,
                    total_tokens=total_tokens
                )
            )
            
        except Exception as e:
            error_msg = str(e) if not isinstance(e, OpenAICompatError) else e.message
            model = self._find_model(request.model, user_id)
            log_service.create_log(
                user_id=user_id,
                api_key_id=api_key_id,
                model_id=model.id if model else None,
                model_name=request.model,
                is_local=model.is_local if model else False,
                input_tokens=0,
                output_tokens=0,
                bill=0,
                status="failed",
                error_message=error_msg,
                request_ip=request_ip
            )
            if isinstance(e, OpenAICompatError):
                raise e
            raise OpenAICompatError(status_code=500, message=error_msg, error_type="rerank_error")