LQAgentPlatform/foundation/ai/models/model_handler.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

"""
AI模型处理器

用于管理生成、与嵌入模型的创建和配置

支持的模型类型：
- doubao: 豆包模型
- qwen: 通义千问模型
- deepseek: DeepSeek模型
- gemini: Gemini模型
- lq_qwen3_8b: 本地Qwen3-8B模型
- lq_qwen3_8b_lq_lora: 本地Qwen3-8B-lq-lora模型
- lq_qwen3_4b: 本地Qwen3-4B模型
- qwen_local_14b: 本地Qwen3-14B模型
- lq_qwen3_8b_emd: 本地Qwen3-Embedding-8B嵌入模型
- siliconflow_embed: 硅基流动Qwen3-Embedding-8B嵌入模型
- lq_bge_reranker_v2_m3: 本地BGE-reranker-v2-m3重排序模型
"""

# 禁用 transformers 的深度学习框架检测，避免启动时耗时扫描
import os
os.environ["TRANSFORMERS_VERBOSITY"] = "error"
os.environ["HF_HUB_DISABLE_TELEMETRY"] = "1"

import requests
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from foundation.infrastructure.config.config import config_handler
from foundation.observability.logger.loggering import server_logger as logger


class ModelConnectionError(Exception):
    """模型连接错误"""
    pass


class ModelConfigError(Exception):
    """模型配置错误"""
    pass


class ModelAPIError(Exception):
    """模型API调用错误"""
    pass


class ModelHandler:
    """
    AI模型处理器类，用于管理多种AI模型的创建和配置
    """

    # 模型连接超时时间配置（秒）
    CONNECTION_TIMEOUT = 30
    REQUEST_TIMEOUT = 120
    MAX_RETRIES = 2

    def __init__(self):
        """
        初始化模型处理器

        加载配置处理器，用于后续读取各种模型的配置信息
        """
        self.config = config_handler
        self._model_cache = {}  # 模型实例缓存

    def _check_connection(self, base_url: str, api_key: str = None, timeout: int = 5) -> bool:
        """
        检查模型服务连接是否可用

        Args:
            base_url: 模型服务地址
            api_key: API密钥（可选）
            timeout: 超时时间（秒）

        Returns:
            bool: 连接是否可用
        """
        try:
            # 构造健康检查URL
            health_url = f"{base_url.rstrip('/')}/models"

            headers = {}
            if api_key and api_key != "dummy":
                headers["Authorization"] = f"Bearer {api_key}"

            response = requests.get(
                health_url,
                headers=headers,
                timeout=timeout
            )

            # 200-299 都认为可用
            return 200 <= response.status_code < 300

        except requests.exceptions.Timeout:
            logger.warning(f"连接超时: {base_url}")
            return False
        except requests.exceptions.ConnectionError as e:
            logger.warning(f"连接错误: {base_url}, 错误: {e}")
            return False
        except Exception as e:
            logger.warning(f"连接检查异常: {base_url}, 错误: {e}")
            return False

    def _handle_model_error(self, model_name: str, error: Exception, fallback_model=None):
        """
        统一处理模型错误

        Args:
            model_name: 模型名称
            error: 异常对象
            fallback_model: 降级模型实例（可选）

        Returns:
            降级模型实例，如果不可用则返回None
        """
        error_type = type(error).__name__
        error_msg = str(error)

        logger.error(f"模型初始化失败 [{model_name}]: {error_type} - {error_msg}")

        # 如果提供了降级模型，记录日志并返回
        if fallback_model:
            logger.warning(f"使用降级模型: {fallback_model.__class__.__name__}")
            return fallback_model

        # 如果没有降级模型，返回None让调用方处理
        return None

    def get_models(self):
        """
        获取AI模型实例

        Returns:
            ChatOpenAI: 配置好的AI模型实例

        Note:
            根据配置文件中的MODEL_TYPE参数选择对应模型
            支持的模型类型：doubao, qwen, deepseek, lq_qwen3_8b, lq_qwen3_8b_lora, lq_qwen3_4b, qwen_local_14b
            默认返回豆包模型
        """
        model_type = self.config.get("model", "MODEL_TYPE")
        logger.info(f"正在初始化AI模型，模型类型: {model_type}")

        # 检查缓存
        cache_key = f"chat_{model_type}"
        if cache_key in self._model_cache:
            logger.info(f"使用缓存的模型: {model_type}")
            return self._model_cache[cache_key]

        model = None

        try:
            if model_type == "doubao":
                model = self._get_doubao_model()
            elif model_type == "gemini":
                model = self._get_gemini_model()
            elif model_type == "qwen":
                model = self._get_qwen_model()
            elif model_type == "deepseek":
                model = self._get_deepseek_model()
            elif model_type == "lq_qwen3_8b":
                model = self._get_lq_qwen3_8b_model()
            elif model_type == "lq_qwen3_8b_lq_lora":
                model = self._get_lq_qwen3_8b_lora_model()
            elif model_type == "lq_qwen3_4b":
                model = self._get_lq_qwen3_4b_model()
            elif model_type == "qwen_local_14b":
                model = self._get_qwen_local_14b_model()
            else:
                # 默认返回gemini
                logger.warning(f"未知的模型类型 '{model_type}'，使用默认gemini模型")
                model = self._get_gemini_model()

            if model:
                self._model_cache[cache_key] = model
                logger.info(f"AI模型初始化完成: {model_type}")
                return model
            else:
                raise ModelAPIError(f"模型初始化返回None: {model_type}")

        except Exception as e:
            logger.error(f"获取模型失败 [{model_type}]: {e}")

            # 尝试使用gemini作为降级方案
            if model_type != "gemini":
                logger.info("尝试使用Gemini模型作为降级方案")
                try:
                    fallback_model = self._get_gemini_model()
                    if fallback_model:
                        self._model_cache[cache_key] = fallback_model
                        logger.warning(f"已切换到Gemini降级模型")
                        return fallback_model
                except Exception as fallback_error:
                    logger.error(f"降级模型也失败: {fallback_error}")

            # 如果所有模型都失败，抛出异常
            raise ModelConnectionError(f"无法初始化任何模型服务: {e}")

    def get_embedding_model(self):
        """
        获取Embedding模型实例

        Returns:
            OpenAIEmbeddings: 配置好的Embedding模型实例

        Note:
            根据配置文件中的EMBEDDING_MODEL_TYPE参数选择对应模型
            支持的模型类型：lq_qwen3_8b_emd, siliconflow_embed
            默认返回本地 lq_qwen3_8b_emd 模型
        """
        embedding_model_type = self.config.get("model", "EMBEDDING_MODEL_TYPE", "lq_qwen3_8b_emd")
        logger.info(f"正在初始化Embedding模型，模型类型: {embedding_model_type}")

        # 检查缓存
        cache_key = f"embed_{embedding_model_type}"
        if cache_key in self._model_cache:
            logger.info(f"使用缓存的Embedding模型: {embedding_model_type}")
            return self._model_cache[cache_key]

        model = None

        try:
            if embedding_model_type == "siliconflow_embed":
                model = self._get_siliconflow_embedding_model()
            elif embedding_model_type == "lq_qwen3_8b_emd":
                model = self._get_lq_qwen3_8b_emd()
            else:
                # 默认返回本地模型
                logger.warning(f"未知的Embedding模型类型 '{embedding_model_type}'，使用默认本地模型")
                model = self._get_lq_qwen3_8b_emd()

            if model:
                self._model_cache[cache_key] = model
                logger.info(f"Embedding模型初始化完成: {embedding_model_type}")
                return model
            else:
                raise ModelAPIError(f"Embedding模型初始化返回None: {embedding_model_type}")

        except Exception as e:
            logger.error(f"获取Embedding模型失败 [{embedding_model_type}]: {e}")
            raise ModelConnectionError(f"无法初始化Embedding模型服务: {e}")

    def _get_doubao_model(self):
        """
        获取豆包模型

        Returns:
            ChatOpenAI: 配置好的豆包模型实例
        """
        try:
            doubao_url = self.config.get("doubao", "DOUBAO_SERVER_URL")
            doubao_model_id = self.config.get("doubao", "DOUBAO_MODEL_ID")
            doubao_api_key = self.config.get("doubao", "DOUBAO_API_KEY")

            # 验证配置完整性
            if not all([doubao_url, doubao_model_id, doubao_api_key]):
                missing = []
                if not doubao_url:
                    missing.append("DOUBAO_SERVER_URL")
                if not doubao_model_id:
                    missing.append("DOUBAO_MODEL_ID")
                if not doubao_api_key:
                    missing.append("DOUBAO_API_KEY")
                raise ModelConfigError(f"豆包模型配置不完整，缺少: {', '.join(missing)}")

            # 检查连接
            if not self._check_connection(doubao_url, doubao_api_key):
                logger.warning(f"豆包模型服务连接失败: {doubao_url}")
                raise ModelConnectionError(f"无法连接到豆包模型服务: {doubao_url}")

            llm = ChatOpenAI(
                base_url=doubao_url,
                model=doubao_model_id,
                api_key=doubao_api_key,
                temperature=0.7,
                timeout=self.REQUEST_TIMEOUT,
                extra_body={
                    "enable_thinking": False,
                })

            logger.info(f"豆包模型初始化成功: {doubao_model_id}")
            return llm

        except ModelConfigError:
            raise
        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"豆包模型初始化异常: {e}")
            return self._handle_model_error("doubao", error)

    def _get_qwen_model(self):
        """
        获取通义千问模型

        Returns:
            ChatOpenAI: 配置好的通义千问模型实例
        """
        try:
            qwen_url = self.config.get("qwen", "QWEN_SERVER_URL")
            qwen_model_id = self.config.get("qwen", "QWEN_MODEL_ID")
            qwen_api_key = self.config.get("qwen", "QWEN_API_KEY")

            # 验证配置完整性
            if not all([qwen_url, qwen_model_id, qwen_api_key]):
                missing = []
                if not qwen_url:
                    missing.append("QWEN_SERVER_URL")
                if not qwen_model_id:
                    missing.append("QWEN_MODEL_ID")
                if not qwen_api_key:
                    missing.append("QWEN_API_KEY")
                raise ModelConfigError(f"通义千问模型配置不完整，缺少: {', '.join(missing)}")

            # 检查连接
            if not self._check_connection(qwen_url, qwen_api_key):
                logger.warning(f"通义千问模型服务连接失败: {qwen_url}")
                raise ModelConnectionError(f"无法连接到通义千问模型服务: {qwen_url}")

            llm = ChatOpenAI(
                base_url=qwen_url,
                model=qwen_model_id,
                api_key=qwen_api_key,
                temperature=0.7,
                timeout=self.REQUEST_TIMEOUT,
                extra_body={
                    "enable_thinking": False,
                })

            logger.info(f"通义千问模型初始化成功: {qwen_model_id}")
            return llm

        except ModelConfigError:
            raise
        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"通义千问模型初始化异常: {e}")
            return self._handle_model_error("qwen", error)

    def _get_deepseek_model(self):
        """
        获取DeepSeek模型

        Returns:
            ChatOpenAI: 配置好的DeepSeek模型实例
        """
        try:
            deepseek_url = self.config.get("deepseek", "DEEPSEEK_SERVER_URL")
            deepseek_model_id = self.config.get("deepseek", "DEEPSEEK_MODEL_ID")
            deepseek_api_key = self.config.get("deepseek", "DEEPSEEK_API_KEY")

            # 验证配置完整性
            if not all([deepseek_url, deepseek_model_id, deepseek_api_key]):
                missing = []
                if not deepseek_url:
                    missing.append("DEEPSEEK_SERVER_URL")
                if not deepseek_model_id:
                    missing.append("DEEPSEEK_MODEL_ID")
                if not deepseek_api_key:
                    missing.append("DEEPSEEK_API_KEY")
                raise ModelConfigError(f"DeepSeek模型配置不完整，缺少: {', '.join(missing)}")

            # 检查连接
            if not self._check_connection(deepseek_url, deepseek_api_key):
                logger.warning(f"DeepSeek模型服务连接失败: {deepseek_url}")
                raise ModelConnectionError(f"无法连接到DeepSeek模型服务: {deepseek_url}")

            llm = ChatOpenAI(
                base_url=deepseek_url,
                model=deepseek_model_id,
                api_key=deepseek_api_key,
                temperature=0.7,
                timeout=self.REQUEST_TIMEOUT,
                extra_body={
                    "enable_thinking": False,
                })

            logger.info(f"DeepSeek模型初始化成功: {deepseek_model_id}")
            return llm

        except ModelConfigError:
            raise
        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"DeepSeek模型初始化异常: {e}")
            return self._handle_model_error("deepseek", error)

    def _get_gemini_model(self):
        """
        获取Gemini模型

        Returns:
            ChatOpenAI: 配置好的Gemini模型实例
        """
        try:
            gemini_url = self.config.get("gemini", "GEMINI_SERVER_URL")
            gemini_model_id = self.config.get("gemini", "GEMINI_MODEL_ID")
            gemini_api_key = self.config.get("gemini", "GEMINI_API_KEY")

            # 验证配置完整性
            if not all([gemini_url, gemini_model_id, gemini_api_key]):
                missing = []
                if not gemini_url:
                    missing.append("GEMINI_SERVER_URL")
                if not gemini_model_id:
                    missing.append("GEMINI_MODEL_ID")
                if not gemini_api_key:
                    missing.append("GEMINI_API_KEY")
                raise ModelConfigError(f"Gemini模型配置不完整，缺少: {', '.join(missing)}")

            # 检查连接
            if not self._check_connection(gemini_url, gemini_api_key):
                logger.warning(f"Gemini模型服务连接失败: {gemini_url}")
                raise ModelConnectionError(f"无法连接到Gemini模型服务: {gemini_url}")

            llm = ChatOpenAI(
                base_url=gemini_url,
                model=gemini_model_id,
                api_key=gemini_api_key,
                temperature=0.7,
                timeout=self.REQUEST_TIMEOUT,
            )

            logger.info(f"Gemini模型初始化成功: {gemini_model_id}")
            return llm

        except ModelConfigError:
            raise
        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"Gemini模型初始化异常: {e}")
            return self._handle_model_error("gemini", error)

    def _get_lq_qwen3_8b_model(self):
        """
        获取本地Qwen3-8B-Instruct模型

        Returns:
            ChatOpenAI: 配置好的本地Qwen3-8B模型实例
        """
        try:
            server_url = "http://192.168.91.253:9002/v1"
            model_id = "Qwen3-8B"

            # 检查本地服务连接
            if not self._check_connection(server_url, "dummy", timeout=3):
                logger.warning(f"本地Qwen3-8B模型服务连接失败: {server_url}")
                raise ModelConnectionError(f"无法连接到本地Qwen3-8B模型服务: {server_url}")

            llm = ChatOpenAI(
                base_url=server_url,
                model=model_id,
                api_key="dummy",
                temperature=0.7,
                timeout=self.REQUEST_TIMEOUT,
            )

            logger.info(f"本地Qwen3-8B模型初始化成功: {model_id}")
            return llm

        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"本地Qwen3-8B模型初始化异常: {e}")
            return self._handle_model_error("lq_qwen3_8b", error)

    def _get_lq_qwen3_8b_lora_model(self):
        """
        获取本地Qwen3-8B-lq-lora模型

        Returns:
            ChatOpenAI: 配置好的本地Qwen3-8B-lq-lora模型实例
        """
        try:
            server_url = self.config.get("lq_qwen3_8B_lora", "LQ_QWEN3_8B_LQ_LORA_SERVER_URL")
            model_id = self.config.get("lq_qwen3_8B_lora", "LQ_QWEN3_8B_LQ_LORA_MODEL_ID")
            api_key = self.config.get("lq_qwen3_8B_lora", "LQ_QWEN3_8B_LQ_LORA_API_KEY", "dummy")

            # 验证配置完整性
            if not all([server_url, model_id]):
                missing = []
                if not server_url:
                    missing.append("LQ_QWEN3_8B_LQ_LORA_SERVER_URL")
                if not model_id:
                    missing.append("LQ_QWEN3_8B_LQ_LORA_MODEL_ID")
                raise ModelConfigError(f"本地Qwen3-8B-lq-lora模型配置不完整，缺少: {', '.join(missing)}")

            # 检查本地服务连接
            if not self._check_connection(server_url, api_key, timeout=3):
                logger.warning(f"本地Qwen3-8B-lq-lora模型服务连接失败: {server_url}")
                raise ModelConnectionError(f"无法连接到本地Qwen3-8B-lq-lora模型服务: {server_url}")

            llm = ChatOpenAI(
                base_url=server_url,
                model=model_id,
                api_key=api_key,
                temperature=0.7,
                timeout=self.REQUEST_TIMEOUT,
            )

            logger.info(f"本地Qwen3-8B-lq-lora模型初始化成功: {model_id}")
            return llm

        except ModelConfigError:
            raise
        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"本地Qwen3-8B-lq-lora模型初始化异常: {e}")
            return self._handle_model_error("lq_qwen3_8b_lora", error)

    def _get_lq_qwen3_4b_model(self):
        """
        获取本地Qwen3-4B-Instruct模型

        Returns:
            ChatOpenAI: 配置好的本地Qwen3-4B模型实例
        """
        try:
            server_url = "http://192.168.91.253:9001/v1"
            model_id = "Qwen3-4B"

            # 检查本地服务连接
            if not self._check_connection(server_url, "dummy", timeout=3):
                logger.warning(f"本地Qwen3-4B模型服务连接失败: {server_url}")
                raise ModelConnectionError(f"无法连接到本地Qwen3-4B模型服务: {server_url}")

            llm = ChatOpenAI(
                base_url=server_url,
                model=model_id,
                api_key="dummy",
                temperature=0.7,
                timeout=self.REQUEST_TIMEOUT,
            )

            logger.info(f"本地Qwen3-4B模型初始化成功: {model_id}")
            return llm

        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"本地Qwen3-4B模型初始化异常: {e}")
            return self._handle_model_error("lq_qwen3_4b", error)

    def _get_qwen_local_14b_model(self):
        """
        获取本地Qwen3-14B-Instruct模型

        Returns:
            ChatOpenAI: 配置好的本地Qwen3-14B模型实例
        """
        try:
            server_url = "http://192.168.91.253:9003/v1"
            model_id = "Qwen3-14B"

            # 检查本地服务连接
            if not self._check_connection(server_url, "dummy", timeout=3):
                logger.warning(f"本地Qwen3-14B模型服务连接失败: {server_url}")
                raise ModelConnectionError(f"无法连接到本地Qwen3-14B模型服务: {server_url}")

            llm = ChatOpenAI(
                base_url=server_url,
                model=model_id,
                api_key="dummy",
                temperature=0.7,
                timeout=self.REQUEST_TIMEOUT,
            )

            logger.info(f"本地Qwen3-14B模型初始化成功: {model_id}")
            return llm

        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"本地Qwen3-14B模型初始化异常: {e}")
            return self._handle_model_error("qwen_local_14b", error)

    def _get_lq_qwen3_8b_emd(self):
        """
        获取本地Qwen3-Embedding-8B嵌入模型

        Returns:
            OpenAIEmbeddings: 配置好的本地Qwen3-Embedding-8B嵌入模型实例
        """
        try:
            server_url = "http://192.168.91.253:9003/v1"
            model_id = "Qwen3-Embedding-8B"

            # 检查本地服务连接
            if not self._check_connection(server_url, "dummy", timeout=3):
                logger.warning(f"本地Qwen3-Embedding-8B模型服务连接失败: {server_url}")
                raise ModelConnectionError(f"无法连接到本地Qwen3-Embedding-8B模型服务: {server_url}")

            # 使用 langchain_openai 的 OpenAIEmbeddings
            embeddings = OpenAIEmbeddings(
                base_url=server_url,
                model=model_id,
                api_key="dummy",  # 本地模型使用虚拟API key
                timeout=self.REQUEST_TIMEOUT,
            )

            logger.info(f"本地Qwen3-Embedding-8B模型初始化成功: {model_id}")
            return embeddings

        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"本地Qwen3-Embedding-8B模型初始化异常: {e}")
            return self._handle_model_error("lq_qwen3_8b_emd", error)

    def _get_siliconflow_embedding_model(self):
        """
        获取硅基流动Qwen3-Embedding-8B嵌入模型

        Returns:
            OpenAIEmbeddings: 配置好的硅基流动Qwen3-Embedding-8B嵌入模型实例
        """
        try:
            server_url = self.config.get("siliconflow_embed", "SLCF_EMBED_SERVER_URL")
            api_key = self.config.get("siliconflow_embed", "SLCF_EMBED_API_KEY")
            model_id = self.config.get("siliconflow_embed", "SLCF_EMBED_MODEL_ID", "Qwen/Qwen3-Embedding-8B")
            dimensions = self.config.get("siliconflow_embed", "SLCF_EMBED_DIMENSIONS", "4096")

            # 验证配置完整性
            if not all([server_url, api_key, model_id]):
                missing = []
                if not server_url:
                    missing.append("SLCF_EMBED_SERVER_URL")
                if not api_key:
                    missing.append("SLCF_EMBED_API_KEY")
                if not model_id:
                    missing.append("SLCF_EMBED_MODEL_ID")
                raise ModelConfigError(f"硅基流动Embedding模型配置不完整，缺少: {', '.join(missing)}")

            # 检查连接
            if not self._check_connection(server_url, api_key):
                logger.warning(f"硅基流动Embedding模型服务连接失败: {server_url}")
                raise ModelConnectionError(f"无法连接到硅基流动Embedding模型服务: {server_url}")

            # 使用 langchain_openai 的 OpenAIEmbeddings
            embeddings = OpenAIEmbeddings(
                base_url=server_url,
                model=model_id,
                api_key=api_key,
                timeout=self.REQUEST_TIMEOUT,
            )

            logger.info(f"硅基流动Embedding模型初始化成功: {model_id} (dimensions: {dimensions})")
            return embeddings

        except ModelConfigError:
            raise
        except ModelConnectionError:
            raise
        except Exception as e:
            error = ModelAPIError(f"硅基流动Embedding模型初始化异常: {e}")
            return self._handle_model_error("siliconflow_embed", error)
    



# 创建全局实例
model_handler = ModelHandler()

def get_models():
    """
    获取模型的全局函数

    Returns:
        tuple: (llm, chat, embed) - LLM模型、聊天模型和嵌入模型实例
               注意：当前llm和chat使用相同模型实例，embed暂时返回None

    Note:
        这是一个便捷函数，直接使用全局model_handler实例获取模型
    """
    try:
        llm = model_handler.get_models()
        # 暂时返回相同的模型作为chat和embed
        return llm, llm, None
    except Exception as e:
        logger.error(f"获取模型失败: {e}")
        raise ModelConnectionError(f"无法获取模型服务: {e}")