feat: make chat model routing configurable

ai_knowledge_base_v2/apps/backend/app/services/agent_debug_service.py

@@ -136,6 +136,7 @@ class AgentDebugService:
             chat_route = ModelRoutingService.resolve_chat(
                 db,
                 [chunk.knowledge_type for chunk in rag_result.chunks],
+                rag_result.question,
             )
             automatic_route = default_model is not None and requested_model.id == default_model.id
             if automatic_route:

ai_knowledge_base_v2/apps/backend/app/services/chat_service.py

@@ -146,6 +146,7 @@ class ChatService:
             failed_route = ModelRoutingService.resolve_chat(
                 db,
                 [chunk.knowledge_type for chunk in rag_result.chunks] if rag_result is not None else [],
+                rag_result.question if rag_result is not None else normalized_question,
             )
             AiRequestLogService.write_failed(
                 db,

ai_knowledge_base_v2/apps/backend/app/services/chat_stream_service.py

@@ -441,6 +441,7 @@ def _model_log_context(
     route = ModelRoutingService.resolve_chat(
         db,
         [chunk.knowledge_type for chunk in rag_result.chunks] if rag_result is not None else [],
+        rag_result.question if rag_result is not None else "",
     )
     return (
         route.reason,

ai_knowledge_base_v2/apps/backend/app/services/model_routing_service.py

@@ -1,19 +1,21 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import Literal
+from typing import Literal, cast
 
 from sqlalchemy import desc, select
 from sqlalchemy.orm import Session
 
 from app.models.ai_config import ModelConfig, SystemConfig
 
-ModelScenario = Literal["report", "summary", "fixed_info", "deep_chat"]
+ModelScenario = Literal["report", "summary", "fixed_info", "simple_knowledge", "deep_chat"]
+ChatRoutingMode = Literal["off", "fixed_only", "conservative"]
 
 SCENARIO_LABELS: dict[ModelScenario, str] = {
     "report": "周期报告",
     "summary": "摘要沉淀",
     "fixed_info": "固定信息",
+    "simple_knowledge": "简单知识问答",
     "deep_chat": "深度对话",
 }
 
@@ -21,9 +23,59 @@ _SCENARIO_FIELDS = {
     "report": ModelConfig.allow_report,
     "summary": ModelConfig.allow_summary,
     "fixed_info": ModelConfig.allow_fixed_info,
+    "simple_knowledge": ModelConfig.allow_simple_knowledge,
     "deep_chat": ModelConfig.allow_deep_chat,
 }
 
+_SIMPLE_KNOWLEDGE_TYPES = {"course", "qa", "general"}
+_SIMPLE_QUERY_MARKERS = (
+    "是什么",
+    "有哪些",
+    "有什么",
+    "包括什么",
+    "定义",
+    "含义",
+    "区别",
+    "时间",
+    "几点",
+    "在哪里",
+    "多少",
+    "多久",
+    "步骤",
+    "怎么操作",
+    "怎么练",
+    "注意事项",
+    "作业",
+    "功课",
+)
+_DEEP_QUERY_MARKERS = (
+    "我",
+    "自己",
+    "感觉",
+    "感受",
+    "情绪",
+    "身体",
+    "害怕",
+    "担心",
+    "焦虑",
+    "难受",
+    "痛苦",
+    "卡住",
+    "怎么办",
+    "为什么",
+    "关系",
+    "孩子",
+    "父母",
+    "伴侣",
+    "创伤",
+    "建议",
+    "分析",
+    "帮我",
+    "适合我",
+    "对不对",
+    "确认一下",
+)
+
 
 @dataclass(frozen=True)
 class ModelRoute:
@@ -90,14 +142,20 @@ class ModelRoutingService:
         )
 
     @classmethod
-    def resolve_chat(cls, db: Session, knowledge_types: list[str]) -> ModelRoute:
+    def resolve_chat(
+        cls,
+        db: Session,
+        knowledge_types: list[str],
+        question: str = "",
+    ) -> ModelRoute:
         normalized_types = {item.strip().lower() for item in knowledge_types if item and item.strip()}
+        mode = cls.chat_routing_mode(db)
         if normalized_types == {"fixed"}:
-            if not cls._config_bool(db, "fixed_info_model_routing_enabled", True):
+            if mode == "off":
                 return ModelRoute(
                     model=cls.default_model(db),
                     scenario="deep_chat",
-                    reason="正式聊天：仅召回固定信息类知识库，但固定信息模型分流开关已关闭；使用默认主模型",
+                    reason="正式聊天：仅召回固定信息类知识库，但正式问答模型分流开关已关闭；使用默认主模型",
                     fallback_used=False,
                     question_type="fixed_info",
                 )
@@ -110,6 +168,16 @@ class ModelRoutingService:
                 question_type="fixed_info",
             )
 
+        if mode == "conservative" and cls._is_simple_knowledge_query(question, normalized_types):
+            route = cls.resolve(db, "simple_knowledge")
+            return ModelRoute(
+                model=route.model,
+                scenario=route.scenario,
+                reason=f"正式聊天：命中知识库且符合保守简单知识规则；{route.reason}",
+                fallback_used=route.fallback_used,
+                question_type="simple_knowledge",
+            )
+
         default = cls.default_model(db)
         if not normalized_types:
             reason = "正式聊天：未召回知识库；使用默认主模型"
@@ -128,9 +196,35 @@ class ModelRoutingService:
             question_type=question_type,
         )
 
+    @classmethod
+    def chat_routing_mode(cls, db: Session) -> ChatRoutingMode:
+        value = cls._config_value(db, "chat_model_routing_mode")
+        if value in {"off", "fixed_only", "conservative"}:
+            return cast(ChatRoutingMode, value)
+        # Compatibility with the previous emergency switch.
+        return "fixed_only" if cls._config_bool(db, "fixed_info_model_routing_enabled", True) else "off"
+
+    @staticmethod
+    def _is_simple_knowledge_query(question: str, knowledge_types: set[str]) -> bool:
+        normalized_question = "".join(question.split()).lower()
+        if not normalized_question or len(normalized_question) > 80:
+            return False
+        if not knowledge_types or not knowledge_types.issubset(_SIMPLE_KNOWLEDGE_TYPES):
+            return False
+        if any(marker in normalized_question for marker in _DEEP_QUERY_MARKERS):
+            return False
+        return any(marker in normalized_question for marker in _SIMPLE_QUERY_MARKERS)
+
     @staticmethod
     def _config_bool(db: Session, key: str, default: bool) -> bool:
+        value = ModelRoutingService._config_value(db, key)
+        if value is None:
+            return default
+        return value.lower() in {"1", "true", "yes", "on", "启用"}
+
+    @staticmethod
+    def _config_value(db: Session, key: str) -> str | None:
         config = db.scalar(select(SystemConfig).where(SystemConfig.config_key == key))
         if config is None or not config.config_value.strip():
-            return default
-        return config.config_value.strip().lower() in {"1", "true", "yes", "on", "启用"}
+            return None
+        return config.config_value.strip()

ai_knowledge_base_v2/apps/backend/app/services/model_service.py

@@ -35,6 +35,7 @@ class ModelClientService:
         route = ModelRoutingService.resolve_chat(
             db,
             [chunk.knowledge_type for chunk in rag_result.chunks],
+            rag_result.question,
         )
         model = route.model
         mock_model_enabled = _system_config_bool(db, "mock_model_enabled", get_settings().mock_model_enabled)

ai_knowledge_base_v2/apps/backend/app/services/model_stream_service.py

@@ -154,6 +154,7 @@ def _chat_route(db: Session, rag_result: RagResult) -> ModelRoute:
     return ModelRoutingService.resolve_chat(
         db,
         [chunk.knowledge_type for chunk in rag_result.chunks],
+        rag_result.question,
     )