feat(llm): add json_output support for structured LLM responses

api/app/core/models/base.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 import os
-from typing import Any, Dict, Optional, TypeVar
+from typing import Any, Dict, List, Optional, TypeVar
 
 from langchain_aws import ChatBedrock
 from langchain_community.chat_models import ChatTongyi
@@ -9,7 +9,7 @@ from langchain_core.embeddings import Embeddings
 from langchain_core.language_models import BaseLLM
 from langchain_ollama import OllamaLLM
 from langchain_openai import ChatOpenAI, OpenAI
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, model_validator
 
 from app.core.error_codes import BizCode
 from app.core.exceptions import BusinessException
@@ -25,10 +25,11 @@ class RedBearModelConfig(BaseModel):
     provider: str
     api_key: str
     base_url: Optional[str] = None
+    capability: List[str] = Field(default_factory=list)  # 模型能力列表，驱动所有能力开关
     is_omni: bool = False  # 是否为 Omni 模型
     deep_thinking: bool = False  # 是否启用深度思考模式
     thinking_budget_tokens: Optional[int] = None  # 深度思考 token 预算
-    support_thinking: bool = False  # 模型是否支持 enable_thinking 参数（capability 含 thinking）
+    json_output: bool = False  # 是否强制 JSON 输出
     # 请求超时时间（秒）- 默认120秒以支持复杂的LLM调用，可通过环境变量 LLM_TIMEOUT 配置
     timeout: float = Field(default_factory=lambda: float(os.getenv("LLM_TIMEOUT", "120.0")))
     # 最大重试次数 - 默认2次以避免过长等待，可通过环境变量 LLM_MAX_RETRIES 配置
@@ -36,6 +37,29 @@ class RedBearModelConfig(BaseModel):
     concurrency: int = 5  # 并发限流
     extra_params: Dict[str, Any] = {}
 
+    @model_validator(mode="after")
+    def _resolve_capabilities(self) -> "RedBearModelConfig":
+        from app.core.logging_config import get_business_logger
+        logger = get_business_logger()
+        if self.deep_thinking and "thinking" not in self.capability:
+            logger.warning(
+                f"模型 {self.model_name} 不支持深度思考（capability 中无 'thinking'），已自动关闭 deep_thinking"
+            )
+            self.deep_thinking = False
+            self.thinking_budget_tokens = None
+        if self.json_output and "json_output" not in self.capability:
+            logger.warning(
+                f"模型 {self.model_name} 不支持 JSON 输出（capability 中无 'json_output'），已自动关闭 json_output"
+            )
+            self.json_output = False
+        if self.json_output and self.deep_thinking:
+            logger.warning(
+                f"模型 {self.model_name} json_output 与 deep_thinking 互斥，已自动关闭 deep_thinking"
+            )
+            self.deep_thinking = False
+            self.thinking_budget_tokens = None
+        return self
+
 
 class RedBearModelFactory:
     """模型工厂类"""
@@ -74,17 +98,20 @@ class RedBearModelFactory:
             is_streaming = bool(config.extra_params.get("streaming"))
             if is_streaming:
                 params["stream_usage"] = True
-            # 只有支持 thinking 的模型才传 enable_thinking
-            if config.support_thinking:
-                model_kwargs: Dict[str, Any] = config.extra_params.get("model_kwargs", {})
-                if is_streaming:
-                    model_kwargs["enable_thinking"] = config.deep_thinking
-                    if config.deep_thinking:
-                        model_kwargs["incremental_output"] = True
-                        if config.thinking_budget_tokens:
-                            model_kwargs["thinking_budget"] = config.thinking_budget_tokens
-                else:
-                    model_kwargs["enable_thinking"] = False
+            # 支持 thinking 的模型始终传 enable_thinking，关闭时显式传 False 避免模型默认开启思考
+            if "thinking" in config.capability:
+                extra_body = params.setdefault("extra_body", {})
+                if config.deep_thinking:
+                    extra_body["enable_thinking"] = False
+                    if is_streaming:
+                        extra_body["enable_thinking"] = True
+                    if config.thinking_budget_tokens:
+                        extra_body["thinking_budget"] = config.thinking_budget_tokens
+                params["extra_body"] = extra_body
+            # JSON 输出模式
+            if config.json_output:
+                model_kwargs = params.setdefault("model_kwargs", {})
+                model_kwargs["response_format"] = {"type": "json_object"}
                 params["model_kwargs"] = model_kwargs
             return params
 
@@ -108,27 +135,30 @@ class RedBearModelFactory:
                 **config.extra_params
             }
             # 流式模式下启用 stream_usage 以获取 token 统计
-            if config.extra_params.get("streaming"):
-                params["stream_usage"] = True
-            # 深度思考模式
             is_streaming = bool(config.extra_params.get("streaming"))
-            if config.support_thinking:
-                if is_streaming and not config.is_omni:
-                    if provider == ModelProvider.VOLCANO:
-                        # 火山引擎深度思考仅流式调用支持，非流式时不传 thinking 参数
-                        thinking_config: Dict[str, Any] = {
-                            "type": "enabled" if config.deep_thinking else "disabled"
-                        }
-                        if config.deep_thinking and config.thinking_budget_tokens:
-                            thinking_config["budget_tokens"] = config.thinking_budget_tokens
-                        params["extra_body"] = {"thinking": thinking_config}
-                    else:
-                        # 始终显式传递 enable_thinking，不支持该参数的模型（如 DeepSeek-R1）会直接忽略
-                        model_kwargs: Dict[str, Any] = config.extra_params.get("model_kwargs", {})
-                        model_kwargs["enable_thinking"] = config.deep_thinking
-                        if config.deep_thinking and config.thinking_budget_tokens:
-                            model_kwargs["thinking_budget"] = config.thinking_budget_tokens
-                        params["model_kwargs"] = model_kwargs
+            if is_streaming:
+                params["stream_usage"] = True
+            # 支持 thinking 的模型始终传 enable_thinking，关闭时显式传 False 避免模型默认开启思考
+            if "thinking" in config.capability:
+                # VOLCANO 深度思考仅流式支持
+                if provider == ModelProvider.VOLCANO:
+                    thinking_config: Dict[str, Any] = {"type": "enabled" if config.deep_thinking else "disabled"}
+                    if config.deep_thinking and config.thinking_budget_tokens:
+                        thinking_config["budget_tokens"] = config.thinking_budget_tokens
+                    params["extra_body"] = {"thinking": thinking_config}
+                else:
+                    extra_body = params.setdefault("extra_body", {})
+                    if config.deep_thinking:
+                        extra_body["enable_thinking"] = False
+                        if is_streaming:
+                            extra_body["enable_thinking"] = True
+                        if config.thinking_budget_tokens:
+                            extra_body["thinking_budget"] = config.thinking_budget_tokens
+                    params["extra_body"] = extra_body
+            # JSON 输出模式
+            if config.json_output:
+                params.setdefault("model_kwargs", {})
+                params["model_kwargs"]["response_format"] = {"type": "json_object"}
             return params
         elif provider == ModelProvider.DASHSCOPE:
             params = {
@@ -137,18 +167,21 @@ class RedBearModelFactory:
                 "max_retries": config.max_retries,
                 **config.extra_params
             }
-            # 只有支持 thinking 的模型才传 enable_thinking
-            if config.support_thinking:
+            # 支持 thinking 的模型始终传 enable_thinking，关闭时显式传 False 避免模型默认开启思考
+            if "thinking" in config.capability:
                 is_streaming = bool(config.extra_params.get("streaming"))
-                model_kwargs: Dict[str, Any] = config.extra_params.get("model_kwargs", {})
-                if is_streaming:
-                    model_kwargs["enable_thinking"] = config.deep_thinking
-                    if config.deep_thinking:
-                        model_kwargs["incremental_output"] = True
-                        if config.thinking_budget_tokens:
-                            model_kwargs["thinking_budget"] = config.thinking_budget_tokens
-                else:
+                model_kwargs = params.setdefault("model_kwargs", {})
+                if config.deep_thinking:
                     model_kwargs["enable_thinking"] = False
+                    if is_streaming:
+                        model_kwargs["enable_thinking"] = True
+                        model_kwargs["incremental_output"] = True
+                    if config.thinking_budget_tokens:
+                        model_kwargs["thinking_budget"] = config.thinking_budget_tokens
+                params["model_kwargs"] = model_kwargs
+            if config.json_output:
+                model_kwargs = params.setdefault("model_kwargs", {})
+                model_kwargs["response_format"] = {"type": "json_object"}
                 params["model_kwargs"] = model_kwargs
             return params
         elif provider == ModelProvider.BEDROCK:
@@ -196,6 +229,10 @@ class RedBearModelFactory:
                 params["additional_model_request_fields"] = {
                     "thinking": {"type": "enabled", "budget_tokens": budget}
                 }
+            # JSON 输出模式
+            if config.json_output:
+                params.setdefault("model_kwargs", {})
+                params["model_kwargs"]["response_format"] = {"type": "json_object"}
             return params
         else:
             raise BusinessException(f"不支持的提供商: {provider}", code=BizCode.PROVIDER_NOT_SUPPORTED)