[AC-AISVC-RES-01~15] feat(retrieval): 实现检索策略Pipeline模块

- 新增策略配置模型 (config.py) - GrayscaleConfig: 灰度发布配置 - ModeRouterConfig: 模式路由配置 - MetadataInferenceConfig: 元数据推断配置 - 新增 Pipeline 实现 - DefaultPipeline: 复用现有 OptimizedRetriever 逻辑 - EnhancedPipeline: Dense + Keyword + RRF 组合检索 - 新增路由器 - StrategyRouter: 策略路由器（default/enhanced） - ModeRouter: 模式路由器（direct/react/auto） - 新增 RollbackManager: 回退与审计管理器 - 新增 MetadataInferenceService: 元数据推断统一入口 - 新增单元测试 (51 passed)
2026-03-10 20:50:16 +08:00 · 2026-03-10 20:50:16 +08:00 · 7027097513
parent 9f28498b97
commit 7027097513
12 changed files with 2568 additions and 0 deletions
--- a/ai-service/app/services/retrieval/strategy/init.py
+++ b/ai-service/app/services/retrieval/strategy/init.py
@ -0,0 +1,102 @@
+"""
+Retrieval Strategy Module for AI Service.
+[AC-AISVC-RES-01~15] 策略化检索与嵌入模块。
+
+核心组件：
+- RetrievalStrategyConfig: 策略配置模型
+- BasePipeline: Pipeline 抽象基类
+- DefaultPipeline: 默认策略（复用现有逻辑）
+- EnhancedPipeline: 增强策略（新端到端流程）
+- MetadataInferenceService: 元数据推断统一入口
+- StrategyRouter: 策略路由器
+- ModeRouter: 模式路由器（direct/react/auto）
+- RollbackManager: 回退管理器
+"""
+
+from app.services.retrieval.strategy.config import (
+    FilterMode,
+    GrayscaleConfig,
+    HybridRetrievalConfig,
+    MetadataInferenceConfig,
+    ModeRouterConfig,
+    PipelineConfig,
+    RerankerConfig,
+    RetrievalStrategyConfig,
+    RuntimeMode,
+    StrategyType,
+    get_strategy_config,
+    set_strategy_config,
+)
+from app.services.retrieval.strategy.default_pipeline import (
+    DefaultPipeline,
+    get_default_pipeline,
+)
+from app.services.retrieval.strategy.enhanced_pipeline import (
+    EnhancedPipeline,
+    get_enhanced_pipeline,
+)
+from app.services.retrieval.strategy.metadata_inference import (
+    InferenceContext,
+    InferenceResult,
+    MetadataInferenceService,
+)
+from app.services.retrieval.strategy.mode_router import (
+    ModeDecision,
+    ModeRouter,
+    get_mode_router,
+)
+from app.services.retrieval.strategy.pipeline_base import (
+    BasePipeline,
+    MetadataFilterResult,
+    PipelineContext,
+    PipelineResult,
+)
+from app.services.retrieval.strategy.rollback_manager import (
+    AuditLog,
+    RollbackManager,
+    RollbackResult,
+    RollbackTrigger,
+    get_rollback_manager,
+)
+from app.services.retrieval.strategy.strategy_router import (
+    RoutingDecision,
+    StrategyRouter,
+    get_strategy_router,
+)
+
+__all__ = [
+    "BasePipeline",
+    "PipelineContext",
+    "PipelineResult",
+    "MetadataFilterResult",
+    "DefaultPipeline",
+    "get_default_pipeline",
+    "EnhancedPipeline",
+    "get_enhanced_pipeline",
+    "RetrievalStrategyConfig",
+    "GrayscaleConfig",
+    "PipelineConfig",
+    "RerankerConfig",
+    "ModeRouterConfig",
+    "HybridRetrievalConfig",
+    "MetadataInferenceConfig",
+    "StrategyType",
+    "FilterMode",
+    "RuntimeMode",
+    "get_strategy_config",
+    "set_strategy_config",
+    "MetadataInferenceService",
+    "InferenceContext",
+    "InferenceResult",
+    "StrategyRouter",
+    "RoutingDecision",
+    "get_strategy_router",
+    "ModeRouter",
+    "ModeDecision",
+    "get_mode_router",
+    "RollbackManager",
+    "RollbackResult",
+    "RollbackTrigger",
+    "AuditLog",
+    "get_rollback_manager",
+]
--- a/ai-service/app/services/retrieval/strategy/config.py
+++ b/ai-service/app/services/retrieval/strategy/config.py
@ -0,0 +1,201 @@
+"""
+Retrieval Strategy Configuration.
+[AC-AISVC-RES-01~15] 检索策略配置模型。
+"""
+
+import random
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import Any
+
+
+class StrategyType(str, Enum):
+    """策略类型。"""
+    DEFAULT = "default"
+    ENHANCED = "enhanced"
+
+
+class RuntimeMode(str, Enum):
+    """运行时模式。"""
+    DIRECT = "direct"
+    REACT = "react"
+    AUTO = "auto"
+
+
+class FilterMode(str, Enum):
+    """过滤模式。"""
+    HARD = "hard"
+    SOFT = "soft"
+    NONE = "none"
+
+
+@dataclass
+class GrayscaleConfig:
+    """灰度发布配置。【AC-AISVC-RES-03】"""
+    enabled: bool = False
+    percentage: float = 0.0
+    allowlist: list[str] = field(default_factory=list)
+    
+    def should_use_enhanced(self, tenant_id: str, user_id: str | None = None) -> bool:
+        """判断是否应该使用增强策略。"""
+        if not self.enabled:
+            return False
+        
+        if tenant_id in self.allowlist or (user_id and user_id in self.allowlist):
+            return True
+        
+        return random.random() * 100 < self.percentage
+
+
+@dataclass
+class HybridRetrievalConfig:
+    """混合检索配置。"""
+    dense_weight: float = 0.7
+    keyword_weight: float = 0.3
+    rrf_k: int = 60
+    enable_keyword: bool = True
+    keyword_top_k_multiplier: int = 2
+
+
+@dataclass
+class RerankerConfig:
+    """重排器配置。【AC-AISVC-RES-08】"""
+    enabled: bool = False
+    model: str = "cross-encoder"
+    top_k_after_rerank: int = 5
+    min_score_threshold: float = 0.3
+
+
+@dataclass
+class ModeRouterConfig:
+    """模式路由配置。【AC-AISVC-RES-09~15】"""
+    runtime_mode: RuntimeMode = RuntimeMode.DIRECT
+    react_trigger_confidence_threshold: float = 0.6
+    react_trigger_complexity_score: float = 0.5
+    react_max_steps: int = 5
+    direct_fallback_on_low_confidence: bool = True
+    short_query_threshold: int = 20
+    
+    def should_use_react(
+        self,
+        query: str,
+        confidence: float | None = None,
+        complexity_score: float | None = None,
+    ) -> bool:
+        """判断是否应该使用 ReAct 模式。【AC-AISVC-RES-11~13】"""
+        if self.runtime_mode == RuntimeMode.REACT:
+            return True
+        if self.runtime_mode == RuntimeMode.DIRECT:
+            return False
+        
+        if len(query) <= self.short_query_threshold and confidence and confidence >= self.react_trigger_confidence_threshold:
+            return False
+        
+        if complexity_score and complexity_score >= self.react_trigger_complexity_score:
+            return True
+        
+        if confidence and confidence < self.react_trigger_confidence_threshold:
+            return True
+        
+        return False
+
+
+@dataclass
+class MetadataInferenceConfig:
+    """元数据推断配置。"""
+    enabled: bool = True
+    confidence_high_threshold: float = 0.8
+    confidence_low_threshold: float = 0.5
+    default_filter_mode: FilterMode = FilterMode.SOFT
+    cache_ttl_seconds: int = 300
+    
+    def determine_filter_mode(self, confidence: float | None) -> FilterMode:
+        """根据置信度确定过滤模式。"""
+        if confidence is None:
+            return FilterMode.NONE
+        if confidence >= self.confidence_high_threshold:
+            return FilterMode.HARD
+        if confidence >= self.confidence_low_threshold:
+            return FilterMode.SOFT
+        return FilterMode.NONE
+
+
+@dataclass
+class PipelineConfig:
+    """Pipeline 配置。"""
+    top_k: int = 5
+    score_threshold: float = 0.01
+    min_hits: int = 1
+    two_stage_enabled: bool = True
+    two_stage_expand_factor: int = 10
+    hybrid: HybridRetrievalConfig = field(default_factory=HybridRetrievalConfig)
+
+
+@dataclass
+class RetrievalStrategyConfig:
+    """检索策略顶层配置。【AC-AISVC-RES-01~15】"""
+    active_strategy: StrategyType = StrategyType.DEFAULT
+    grayscale: GrayscaleConfig = field(default_factory=GrayscaleConfig)
+    pipeline: PipelineConfig = field(default_factory=PipelineConfig)
+    reranker: RerankerConfig = field(default_factory=RerankerConfig)
+    mode_router: ModeRouterConfig = field(default_factory=ModeRouterConfig)
+    metadata_inference: MetadataInferenceConfig = field(default_factory=MetadataInferenceConfig)
+    performance_thresholds: dict[str, float] = field(default_factory=lambda: {
+        "max_latency_ms": 2000.0,
+        "min_success_rate": 0.95,
+        "max_error_rate": 0.05,
+    })
+    
+    def is_enhanced_enabled(self, tenant_id: str, user_id: str | None = None) -> bool:
+        """判断是否启用增强策略。"""
+        if self.active_strategy == StrategyType.ENHANCED:
+            return True
+        return self.grayscale.should_use_enhanced(tenant_id, user_id)
+    
+    def to_dict(self) -> dict[str, Any]:
+        """转换为字典。"""
+        return {
+            "active_strategy": self.active_strategy.value,
+            "grayscale": {
+                "enabled": self.grayscale.enabled,
+                "percentage": self.grayscale.percentage,
+                "allowlist": self.grayscale.allowlist,
+            },
+            "pipeline": {
+                "top_k": self.pipeline.top_k,
+                "score_threshold": self.pipeline.score_threshold,
+                "min_hits": self.pipeline.min_hits,
+                "two_stage_enabled": self.pipeline.two_stage_enabled,
+            },
+            "reranker": {
+                "enabled": self.reranker.enabled,
+                "model": self.reranker.model,
+                "top_k_after_rerank": self.reranker.top_k_after_rerank,
+            },
+            "mode_router": {
+                "runtime_mode": self.mode_router.runtime_mode.value,
+                "react_trigger_confidence_threshold": self.mode_router.react_trigger_confidence_threshold,
+            },
+            "metadata_inference": {
+                "enabled": self.metadata_inference.enabled,
+                "confidence_high_threshold": self.metadata_inference.confidence_high_threshold,
+            },
+            "performance_thresholds": self.performance_thresholds,
+        }
+
+
+_global_config: RetrievalStrategyConfig | None = None
+
+
+def get_strategy_config() -> RetrievalStrategyConfig:
+    """获取全局策略配置。"""
+    global _global_config
+    if _global_config is None:
+        _global_config = RetrievalStrategyConfig()
+    return _global_config
+
+
+def set_strategy_config(config: RetrievalStrategyConfig) -> None:
+    """设置全局策略配置。"""
+    global _global_config
+    _global_config = config
--- a/ai-service/app/services/retrieval/strategy/default_pipeline.py
+++ b/ai-service/app/services/retrieval/strategy/default_pipeline.py
@ -0,0 +1,117 @@
+"""
+Default Pipeline.
+[AC-AISVC-RES-01] 默认策略 Pipeline，复用现有逻辑。
+"""
+
+import logging
+import time
+from typing import Any
+
+from app.services.retrieval.base import RetrievalContext, RetrievalResult
+from app.services.retrieval.optimized_retriever import OptimizedRetriever, get_optimized_retriever
+from app.services.retrieval.strategy.config import PipelineConfig
+from app.services.retrieval.strategy.pipeline_base import (
+    BasePipeline,
+    PipelineContext,
+    PipelineResult,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class DefaultPipeline(BasePipeline):
+    """
+    默认策略 Pipeline。【AC-AISVC-RES-01】
+    
+    复用现有 OptimizedRetriever 逻辑，保持线上行为不变。
+    """
+    
+    def __init__(
+        self,
+        config: PipelineConfig | None = None,
+        optimized_retriever: OptimizedRetriever | None = None,
+    ):
+        self._config = config or PipelineConfig()
+        self._optimized_retriever = optimized_retriever
+    
+    @property
+    def name(self) -> str:
+        return "default_pipeline"
+    
+    @property
+    def description(self) -> str:
+        return "默认检索策略，复用现有 OptimizedRetriever 逻辑。"
+    
+    async def _get_retriever(self) -> OptimizedRetriever:
+        if self._optimized_retriever is None:
+            self._optimized_retriever = await get_optimized_retriever()
+        return self._optimized_retriever
+    
+    async def retrieve(self, ctx: PipelineContext) -> PipelineResult:
+        """执行默认检索流程。【AC-AISVC-RES-01】"""
+        start_time = time.time()
+        
+        logger.info(
+            f"[DefaultPipeline] Starting retrieval: tenant={ctx.tenant_id}, "
+            f"query={ctx.query[:50]}..."
+        )
+        
+        try:
+            retriever = await self._get_retriever()
+            
+            metadata_filter = None
+            if ctx.metadata_filter:
+                metadata_filter = ctx.metadata_filter.filter_dict
+            
+            retrieval_ctx = RetrievalContext(
+                tenant_id=ctx.tenant_id,
+                query=ctx.query,
+                session_id=ctx.session_id,
+                metadata_filter=metadata_filter,
+                kb_ids=ctx.kb_ids,
+            )
+            
+            result = await retriever.retrieve(retrieval_ctx)
+            
+            latency_ms = (time.time() - start_time) * 1000
+            
+            logger.info(
+                f"[DefaultPipeline] Retrieval completed: hits={len(result.hits)}, "
+                f"latency_ms={latency_ms:.2f}"
+            )
+            
+            return PipelineResult(
+                retrieval_result=result,
+                pipeline_name=self.name,
+                metadata_filter_applied=metadata_filter is not None,
+                latency_ms=latency_ms,
+                diagnostics={
+                    "retriever": "OptimizedRetriever",
+                    **(result.diagnostics or {}),
+                },
+            )
+            
+        except Exception as e:
+            latency_ms = (time.time() - start_time) * 1000
+            logger.error(f"[DefaultPipeline] Retrieval error: {e}", exc_info=True)
+            return self._create_empty_result(ctx, error=str(e), latency_ms=latency_ms)
+    
+    async def health_check(self) -> bool:
+        """健康检查。"""
+        try:
+            retriever = await self._get_retriever()
+            return await retriever.health_check()
+        except Exception as e:
+            logger.error(f"[DefaultPipeline] Health check failed: {e}")
+            return False
+
+
+_default_pipeline: DefaultPipeline | None = None
+
+
+async def get_default_pipeline() -> DefaultPipeline:
+    """获取 DefaultPipeline 单例。"""
+    global _default_pipeline
+    if _default_pipeline is None:
+        _default_pipeline = DefaultPipeline()
+    return _default_pipeline
--- a/ai-service/app/services/retrieval/strategy/enhanced_pipeline.py
+++ b/ai-service/app/services/retrieval/strategy/enhanced_pipeline.py
@ -0,0 +1,364 @@
+"""
+Enhanced Pipeline.
+[AC-AISVC-RES-02] 增强策略 Pipeline，新端到端流程。
+"""
+
+import asyncio
+import logging
+import re
+import time
+from dataclasses import dataclass
+from typing import Any
+
+import numpy as np
+
+from app.core.config import get_settings
+from app.core.qdrant_client import QdrantClient, get_qdrant_client
+from app.services.embedding.nomic_provider import NomicEmbeddingProvider
+from app.services.retrieval.base import RetrievalHit, RetrievalResult
+from app.services.retrieval.optimized_retriever import RRFCombiner
+from app.services.retrieval.strategy.config import (
+    HybridRetrievalConfig,
+    PipelineConfig,
+    RerankerConfig,
+)
+from app.services.retrieval.strategy.pipeline_base import (
+    BasePipeline,
+    PipelineContext,
+    PipelineResult,
+)
+
+logger = logging.getLogger(__name__)
+settings = get_settings()
+
+
+@dataclass
+class RetrievalCandidate:
+    """检索候选结果。"""
+    id: str
+    text: str
+    score: float
+    vector_score: float = 0.0
+    keyword_score: float = 0.0
+    metadata: dict[str, Any] = None
+    
+    def __post_init__(self):
+        if self.metadata is None:
+            self.metadata = {}
+
+
+class EnhancedPipeline(BasePipeline):
+    """
+    增强策略 Pipeline。【AC-AISVC-RES-02】
+    
+    新端到端流程：
+    1. Dense 向量检索
+    2. Keyword 关键词检索
+    3. RRF 融合排序
+    4. 可选重排
+    """
+    
+    def __init__(
+        self,
+        config: PipelineConfig | None = None,
+        reranker_config: RerankerConfig | None = None,
+        qdrant_client: QdrantClient | None = None,
+    ):
+        self._config = config or PipelineConfig()
+        self._reranker_config = reranker_config or RerankerConfig()
+        self._qdrant_client = qdrant_client
+        self._rrf_combiner = RRFCombiner(k=self._config.hybrid.rrf_k)
+        self._embedding_provider: NomicEmbeddingProvider | None = None
+    
+    @property
+    def name(self) -> str:
+        return "enhanced_pipeline"
+    
+    @property
+    def description(self) -> str:
+        return "增强检索策略，支持 Dense + Keyword + RRF 组合检索。"
+    
+    async def _get_client(self) -> QdrantClient:
+        if self._qdrant_client is None:
+            self._qdrant_client = await get_qdrant_client()
+        return self._qdrant_client
+    
+    async def _get_embedding_provider(self) -> NomicEmbeddingProvider:
+        if self._embedding_provider is None:
+            from app.services.embedding.factory import get_embedding_config_manager
+            manager = get_embedding_config_manager()
+            provider = await manager.get_provider()
+            if isinstance(provider, NomicEmbeddingProvider):
+                self._embedding_provider = provider
+            else:
+                self._embedding_provider = NomicEmbeddingProvider(
+                    base_url=settings.ollama_base_url,
+                    model=settings.ollama_embedding_model,
+                    dimension=settings.qdrant_vector_size,
+                )
+        return self._embedding_provider
+    
+    async def retrieve(self, ctx: PipelineContext) -> PipelineResult:
+        """执行增强检索流程。【AC-AISVC-RES-02】"""
+        start_time = time.time()
+        
+        logger.info(
+            f"[EnhancedPipeline] Starting retrieval: tenant={ctx.tenant_id}, "
+            f"query={ctx.query[:50]}..."
+        )
+        
+        try:
+            provider = await self._get_embedding_provider()
+            embedding_result = await provider.embed_query(ctx.query)
+            
+            candidates = await self._hybrid_retrieve(
+                tenant_id=ctx.tenant_id,
+                query=ctx.query,
+                embedding_result=embedding_result,
+                metadata_filter=ctx.metadata_filter.filter_dict if ctx.metadata_filter else None,
+                kb_ids=ctx.kb_ids,
+            )
+            
+            if self._reranker_config.enabled and ctx.use_reranker:
+                candidates = await self._rerank(
+                    candidates=candidates,
+                    query=ctx.query,
+                )
+            
+            top_k = self._config.top_k
+            final_candidates = candidates[:top_k]
+            
+            hits = [
+                RetrievalHit(
+                    text=c.text,
+                    score=c.score,
+                    source=self.name,
+                    metadata=c.metadata,
+                )
+                for c in final_candidates
+                if c.score >= self._config.score_threshold
+            ]
+            
+            latency_ms = (time.time() - start_time) * 1000
+            
+            logger.info(
+                f"[EnhancedPipeline] Retrieval completed: hits={len(hits)}, "
+                f"latency_ms={latency_ms:.2f}"
+            )
+            
+            result = RetrievalResult(
+                hits=hits,
+                diagnostics={
+                    "total_candidates": len(candidates),
+                    "after_rerank": self._reranker_config.enabled and ctx.use_reranker,
+                },
+            )
+            
+            return PipelineResult(
+                retrieval_result=result,
+                pipeline_name=self.name,
+                used_reranker=self._reranker_config.enabled and ctx.use_reranker,
+                metadata_filter_applied=ctx.metadata_filter is not None,
+                latency_ms=latency_ms,
+                diagnostics={
+                    "dense_weight": self._config.hybrid.dense_weight,
+                    "keyword_weight": self._config.hybrid.keyword_weight,
+                },
+            )
+            
+        except Exception as e:
+            latency_ms = (time.time() - start_time) * 1000
+            logger.error(f"[EnhancedPipeline] Retrieval error: {e}", exc_info=True)
+            return self._create_empty_result(ctx, error=str(e), latency_ms=latency_ms)
+    
+    async def _hybrid_retrieve(
+        self,
+        tenant_id: str,
+        query: str,
+        embedding_result: Any,
+        metadata_filter: dict[str, Any] | None = None,
+        kb_ids: list[str] | None = None,
+    ) -> list[RetrievalCandidate]:
+        """混合检索：Dense + Keyword + RRF。"""
+        client = await self._get_client()
+        top_k = self._config.top_k
+        expand_factor = self._config.hybrid.keyword_top_k_multiplier
+        
+        vector_task = self._dense_search(
+            client=client,
+            tenant_id=tenant_id,
+            embedding=embedding_result.embedding_full,
+            top_k=top_k * expand_factor,
+            metadata_filter=metadata_filter,
+            kb_ids=kb_ids,
+        )
+        
+        keyword_task = self._keyword_search(
+            client=client,
+            tenant_id=tenant_id,
+            query=query,
+            top_k=top_k * expand_factor,
+            metadata_filter=metadata_filter,
+            kb_ids=kb_ids,
+        ) if self._config.hybrid.enable_keyword else asyncio.sleep(0, result=[])
+        
+        vector_results, keyword_results = await asyncio.gather(
+            vector_task, keyword_task, return_exceptions=True
+        )
+        
+        if isinstance(vector_results, Exception):
+            logger.warning(f"[EnhancedPipeline] Dense search failed: {vector_results}")
+            vector_results = []
+        
+        if isinstance(keyword_results, Exception):
+            logger.warning(f"[EnhancedPipeline] Keyword search failed: {keyword_results}")
+            keyword_results = []
+        
+        combined = self._rrf_combiner.combine(
+            vector_results=vector_results,
+            bm25_results=keyword_results,
+            vector_weight=self._config.hybrid.dense_weight,
+            bm25_weight=self._config.hybrid.keyword_weight,
+        )
+        
+        candidates = []
+        for item in combined:
+            candidates.append(RetrievalCandidate(
+                id=item.get("id", ""),
+                text=item.get("payload", {}).get("text", ""),
+                score=item.get("score", 0.0),
+                vector_score=item.get("vector_score", 0.0),
+                keyword_score=item.get("bm25_score", 0.0),
+                metadata=item.get("payload", {}),
+            ))
+        
+        return candidates
+    
+    async def _dense_search(
+        self,
+        client: QdrantClient,
+        tenant_id: str,
+        embedding: list[float],
+        top_k: int,
+        metadata_filter: dict[str, Any] | None = None,
+        kb_ids: list[str] | None = None,
+    ) -> list[dict[str, Any]]:
+        """Dense 向量检索。"""
+        try:
+            results = await client.search(
+                tenant_id=tenant_id,
+                query_vector=embedding,
+                limit=top_k,
+                vector_name="full",
+                metadata_filter=metadata_filter,
+                kb_ids=kb_ids,
+            )
+            return results
+        except Exception as e:
+            logger.error(f"[EnhancedPipeline] Dense search error: {e}")
+            return []
+    
+    async def _keyword_search(
+        self,
+        client: QdrantClient,
+        tenant_id: str,
+        query: str,
+        top_k: int,
+        metadata_filter: dict[str, Any] | None = None,
+        kb_ids: list[str] | None = None,
+    ) -> list[dict[str, Any]]:
+        """Keyword 关键词检索。"""
+        try:
+            qdrant = await client.get_client()
+            collection_name = client.get_collection_name(tenant_id)
+            
+            query_terms = set(re.findall(r'\w+', query.lower()))
+            
+            results = await qdrant.scroll(
+                collection_name=collection_name,
+                limit=top_k * 3,
+                with_payload=True,
+            )
+            
+            scored_results = []
+            for point in results[0]:
+                text = point.payload.get("text", "").lower()
+                text_terms = set(re.findall(r'\w+', text))
+                overlap = len(query_terms & text_terms)
+                
+                if overlap > 0:
+                    score = overlap / (len(query_terms) + len(text_terms) - overlap)
+                    scored_results.append({
+                        "id": str(point.id),
+                        "score": score,
+                        "payload": point.payload or {},
+                    })
+            
+            scored_results.sort(key=lambda x: x["score"], reverse=True)
+            return scored_results[:top_k]
+            
+        except Exception as e:
+            logger.debug(f"[EnhancedPipeline] Keyword search failed: {e}")
+            return []
+    
+    async def _rerank(
+        self,
+        candidates: list[RetrievalCandidate],
+        query: str,
+    ) -> list[RetrievalCandidate]:
+        """可选重排。"""
+        if not candidates:
+            return candidates
+        
+        try:
+            provider = await self._get_embedding_provider()
+            query_embedding = await provider.embed_query(query)
+            
+            reranked = []
+            for candidate in candidates:
+                candidate_text = candidate.text[:500]
+                if candidate_text:
+                    candidate_embedding = await provider.embed(candidate_text)
+                    similarity = self._cosine_similarity(
+                        query_embedding.embedding_full,
+                        candidate_embedding,
+                    )
+                    candidate.score = similarity
+                
+                if candidate.score >= self._reranker_config.min_score_threshold:
+                    reranked.append(candidate)
+            
+            reranked.sort(key=lambda x: x.score, reverse=True)
+            return reranked[:self._reranker_config.top_k_after_rerank]
+            
+        except Exception as e:
+            logger.warning(f"[EnhancedPipeline] Rerank failed: {e}")
+            return candidates
+    
+    def _cosine_similarity(self, vec1: list[float], vec2: list[float]) -> float:
+        """计算余弦相似度。"""
+        a = np.array(vec1)
+        b = np.array(vec2)
+        return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b)))
+    
+    async def health_check(self) -> bool:
+        """健康检查。"""
+        try:
+            client = await self._get_client()
+            qdrant = await client.get_client()
+            await qdrant.get_collections()
+            return True
+        except Exception as e:
+            logger.error(f"[EnhancedPipeline] Health check failed: {e}")
+            return False
+
+
+_enhanced_pipeline: EnhancedPipeline | None = None
+
+
+async def get_enhanced_pipeline() -> EnhancedPipeline:
+    """获取 EnhancedPipeline 单例。"""
+    global _enhanced_pipeline
+    if _enhanced_pipeline is None:
+        _enhanced_pipeline = EnhancedPipeline()
+    return _enhanced_pipeline
--- a/ai-service/app/services/retrieval/strategy/metadata_inference.py
+++ b/ai-service/app/services/retrieval/strategy/metadata_inference.py
@ -0,0 +1,136 @@
+"""
+Metadata Inference Service.
+[AC-AISVC-RES-04] 元数据推断统一入口。
+"""
+
+import logging
+from dataclasses import dataclass, field
+from typing import Any
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.services.mid.metadata_filter_builder import (
+    FilterBuildResult,
+    MetadataFilterBuilder,
+)
+from app.services.retrieval.strategy.config import (
+    FilterMode,
+    MetadataInferenceConfig,
+)
+from app.services.retrieval.strategy.pipeline_base import MetadataFilterResult
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class InferenceContext:
+    """元数据推断上下文。"""
+    tenant_id: str
+    query: str
+    session_id: str | None = None
+    user_id: str | None = None
+    channel_type: str | None = None
+    existing_context: dict[str, Any] = field(default_factory=dict)
+    slot_state: Any = None
+
+
+@dataclass
+class InferenceResult:
+    """元数据推断结果。"""
+    filter_result: MetadataFilterResult
+    inferred_fields: dict[str, Any] = field(default_factory=dict)
+    confidence_scores: dict[str, float] = field(default_factory=dict)
+    overall_confidence: float | None = None
+    inference_source: str = "unknown"
+
+
+class MetadataInferenceService:
+    """
+    元数据推断统一入口。【AC-AISVC-RES-04】
+    
+    职责：
+    1. 统一的元数据推断入口（策略无关）
+    2. 根据置信度决定 hard/soft filter 模式
+    3. 与现有 MetadataFilterBuilder 保持一致
+    """
+    
+    def __init__(
+        self,
+        session: AsyncSession,
+        config: MetadataInferenceConfig | None = None,
+    ):
+        self._session = session
+        self._config = config or MetadataInferenceConfig()
+        self._filter_builder: MetadataFilterBuilder | None = None
+    
+    async def infer(self, ctx: InferenceContext) -> InferenceResult:
+        """执行元数据推断。【AC-AISVC-RES-04】"""
+        logger.info(
+            f"[MetadataInference] Starting inference: tenant={ctx.tenant_id}, "
+            f"query={ctx.query[:50]}..."
+        )
+        
+        if self._filter_builder is None:
+            self._filter_builder = MetadataFilterBuilder(self._session)
+        
+        effective_context = dict(ctx.existing_context)
+        
+        if ctx.slot_state:
+            effective_context = await self._merge_slot_state(
+                effective_context, ctx.slot_state
+            )
+        
+        build_result = await self._filter_builder.build_filter(
+            tenant_id=ctx.tenant_id,
+            context=effective_context,
+        )
+        
+        confidence = self._calculate_confidence(build_result, effective_context)
+        filter_mode = self._config.determine_filter_mode(confidence)
+        
+        filter_result = MetadataFilterResult(
+            filter_dict=build_result.applied_filter,
+            filter_mode=filter_mode,
+            confidence=confidence,
+            missing_required_slots=build_result.missing_required_slots,
+            debug_info=build_result.debug_info,
+        )
+        
+        logger.info(
+            f"[MetadataInference] Inference completed: filter_mode={filter_mode.value}, "
+            f"confidence={confidence}"
+        )
+        
+        return InferenceResult(
+            filter_result=filter_result,
+            inferred_fields=build_result.applied_filter,
+            overall_confidence=confidence,
+            inference_source="metadata_filter_builder",
+        )
+    
+    async def _merge_slot_state(
+        self, context: dict[str, Any], slot_state: Any
+    ) -> dict[str, Any]:
+        """合并槽位状态到上下文。"""
+        if hasattr(slot_state, 'filled_slots'):
+            for slot_key, slot_value in slot_state.filled_slots.items():
+                if slot_key not in context:
+                    context[slot_key] = slot_value
+        return context
+    
+    def _calculate_confidence(
+        self, build_result: FilterBuildResult, context: dict[str, Any]
+    ) -> float | None:
+        """计算推断置信度。"""
+        if build_result.missing_required_slots:
+            return 0.3
+        if not build_result.applied_filter:
+            return None
+        if not context:
+            return 0.5
+        applied_ratio = len(build_result.applied_filter) / max(len(context), 1)
+        if applied_ratio >= 0.8:
+            return 0.9
+        elif applied_ratio >= 0.5:
+            return 0.7
+        return 0.5
--- a/ai-service/app/services/retrieval/strategy/mode_router.py
+++ b/ai-service/app/services/retrieval/strategy/mode_router.py
@ -0,0 +1,118 @@
+"""
+Mode Router.
+[AC-AISVC-RES-09~15] 模式路由器。
+"""
+
+import logging
+from dataclasses import dataclass
+
+from app.services.retrieval.strategy.config import ModeRouterConfig, RuntimeMode
+from app.services.retrieval.strategy.pipeline_base import PipelineResult
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModeDecision:
+    """模式决策结果。"""
+    mode: RuntimeMode
+    reason: str
+    confidence: float | None = None
+    complexity_score: float | None = None
+
+
+class ModeRouter:
+    """
+    模式路由器。【AC-AISVC-RES-09~15】
+    
+    职责：
+    1. 根据 rag_runtime_mode 选择 direct/react/auto 模式
+    2. auto 模式下根据复杂度与置信度自动选择路由
+    3. direct 低置信度时触发 react 回退
+    """
+    
+    def __init__(self, config: ModeRouterConfig | None = None):
+        self._config = config or ModeRouterConfig()
+    
+    def decide(
+        self,
+        query: str,
+        confidence: float | None = None,
+        complexity_score: float | None = None,
+    ) -> ModeDecision:
+        """决定使用哪种模式。【AC-AISVC-RES-09~13】"""
+        if self._config.runtime_mode == RuntimeMode.REACT:
+            return ModeDecision(mode=RuntimeMode.REACT, reason="runtime_mode=react")
+        
+        if self._config.runtime_mode == RuntimeMode.DIRECT:
+            return ModeDecision(mode=RuntimeMode.DIRECT, reason="runtime_mode=direct")
+        
+        calculated_complexity = complexity_score or self._calculate_complexity(query)
+        
+        if self._should_use_direct(query, confidence, calculated_complexity):
+            return ModeDecision(
+                mode=RuntimeMode.DIRECT,
+                reason="auto: short_query_high_confidence",
+                confidence=confidence,
+                complexity_score=calculated_complexity,
+            )
+        
+        return ModeDecision(
+            mode=RuntimeMode.REACT,
+            reason="auto: complex_or_low_confidence",
+            confidence=confidence,
+            complexity_score=calculated_complexity,
+        )
+    
+    def should_fallback_to_react(self, direct_result: PipelineResult) -> bool:
+        """判断是否应该从 direct 回退到 react。【AC-AISVC-RES-14】"""
+        if not self._config.direct_fallback_on_low_confidence:
+            return False
+        if direct_result.is_empty:
+            return True
+        max_score = direct_result.retrieval_result.max_score
+        if max_score < 0.3:
+            return True
+        if direct_result.retrieval_result.hit_count < 2:
+            return True
+        return False
+    
+    def _should_use_direct(
+        self, query: str, confidence: float | None, complexity_score: float
+    ) -> bool:
+        if len(query) <= self._config.short_query_threshold:
+            if confidence and confidence >= self._config.react_trigger_confidence_threshold:
+                return True
+        if confidence and confidence < self._config.react_trigger_confidence_threshold:
+            return False
+        if complexity_score >= self._config.react_trigger_complexity_score:
+            return False
+        return True
+    
+    def _calculate_complexity(self, query: str) -> float:
+        score = 0.0
+        if len(query) > 50:
+            score += 0.2
+        if len(query) > 100:
+            score += 0.2
+        condition_words = ["和", "或", "但是", "如果", "同时", "并且", "或者", "以及"]
+        for word in condition_words:
+            if word in query:
+                score += 0.1
+        return min(score, 1.0)
+    
+    def get_config(self) -> ModeRouterConfig:
+        return self._config
+    
+    def update_config(self, config: ModeRouterConfig) -> None:
+        self._config = config
+
+
+_mode_router: ModeRouter | None = None
+
+
+def get_mode_router() -> ModeRouter:
+    global _mode_router
+    if _mode_router is None:
+        _mode_router = ModeRouter()
+    return _mode_router
--- a/ai-service/app/services/retrieval/strategy/pipeline_base.py
+++ b/ai-service/app/services/retrieval/strategy/pipeline_base.py
@ -0,0 +1,116 @@
+"""
+Pipeline Base Classes.
+[AC-AISVC-RES-01~15] Pipeline 抽象基类。
+"""
+
+import logging
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from typing import Any
+
+from app.services.retrieval.base import RetrievalContext, RetrievalHit, RetrievalResult
+from app.services.retrieval.strategy.config import FilterMode
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class MetadataFilterResult:
+    """元数据过滤结果。"""
+    filter_dict: dict[str, Any] = field(default_factory=dict)
+    filter_mode: FilterMode = FilterMode.NONE
+    confidence: float | None = None
+    missing_required_slots: list[dict[str, str]] = field(default_factory=list)
+    debug_info: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class PipelineContext:
+    """Pipeline 执行上下文。"""
+    retrieval_ctx: RetrievalContext
+    metadata_filter: MetadataFilterResult | None = None
+    use_reranker: bool = False
+    use_react: bool = False
+    react_iteration: int = 0
+    previous_results: list[RetrievalHit] = field(default_factory=list)
+    extra: dict[str, Any] = field(default_factory=dict)
+    
+    @property
+    def tenant_id(self) -> str:
+        return self.retrieval_ctx.tenant_id
+    
+    @property
+    def query(self) -> str:
+        return self.retrieval_ctx.query
+    
+    @property
+    def session_id(self) -> str | None:
+        return self.retrieval_ctx.session_id
+    
+    @property
+    def kb_ids(self) -> list[str] | None:
+        return self.retrieval_ctx.kb_ids
+
+
+@dataclass
+class PipelineResult:
+    """Pipeline 执行结果。"""
+    retrieval_result: RetrievalResult
+    pipeline_name: str = ""
+    used_reranker: bool = False
+    used_react: bool = False
+    react_iterations: int = 0
+    metadata_filter_applied: bool = False
+    fallback_triggered: bool = False
+    fallback_reason: str | None = None
+    latency_ms: float = 0.0
+    diagnostics: dict[str, Any] = field(default_factory=dict)
+    
+    @property
+    def hits(self) -> list[RetrievalHit]:
+        return self.retrieval_result.hits
+    
+    @property
+    def is_empty(self) -> bool:
+        return self.retrieval_result.is_empty
+
+
+class BasePipeline(ABC):
+    """Pipeline 抽象基类。【AC-AISVC-RES-01, AC-AISVC-RES-02】"""
+    
+    @property
+    @abstractmethod
+    def name(self) -> str:
+        """Pipeline 名称。"""
+        pass
+    
+    @property
+    @abstractmethod
+    def description(self) -> str:
+        """Pipeline 描述。"""
+        pass
+    
+    @abstractmethod
+    async def retrieve(self, ctx: PipelineContext) -> PipelineResult:
+        """执行检索。"""
+        pass
+    
+    @abstractmethod
+    async def health_check(self) -> bool:
+        """健康检查。"""
+        pass
+    
+    def _create_empty_result(
+        self,
+        ctx: PipelineContext,
+        error: str | None = None,
+        latency_ms: float = 0.0,
+    ) -> PipelineResult:
+        """创建空结果。"""
+        diagnostics = {"error": error} if error else {}
+        return PipelineResult(
+            retrieval_result=RetrievalResult(hits=[], diagnostics=diagnostics),
+            pipeline_name=self.name,
+            latency_ms=latency_ms,
+            diagnostics=diagnostics,
+        )
--- a/ai-service/app/services/retrieval/strategy/retrieval_strategy.py
+++ b/ai-service/app/services/retrieval/strategy/retrieval_strategy.py
@ -0,0 +1,301 @@
+"""
+Retrieval Strategy - Unified Entry Point.
+[AC-AISVC-RES-01~15] 检索策略统一入口。
+
+整合：
+- StrategyRouter: 策略路由（default/enhanced）
+- ModeRouter: 模式路由（direct/react/auto）
+- MetadataInferenceService: 元数据推断
+- RollbackManager: 回退管理
+"""
+
+import logging
+import time
+from dataclasses import dataclass
+from typing import Any
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.services.retrieval.base import RetrievalContext, RetrievalResult
+from app.services.retrieval.strategy.config import (
+    RetrievalStrategyConfig,
+    RuntimeMode,
+    StrategyType,
+)
+from app.services.retrieval.strategy.default_pipeline import DefaultPipeline
+from app.services.retrieval.strategy.enhanced_pipeline import EnhancedPipeline
+from app.services.retrieval.strategy.metadata_inference import (
+    InferenceContext,
+    MetadataInferenceService,
+)
+from app.services.retrieval.strategy.mode_router import ModeDecision, ModeRouter
+from app.services.retrieval.strategy.pipeline_base import (
+    MetadataFilterResult,
+    PipelineContext,
+    PipelineResult,
+)
+from app.services.retrieval.strategy.rollback_manager import (
+    RollbackManager,
+    RollbackTrigger,
+)
+from app.services.retrieval.strategy.strategy_router import (
+    RoutingDecision,
+    StrategyRouter,
+)
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RetrievalStrategyResult:
+    """检索策略执行结果。"""
+    
+    retrieval_result: RetrievalResult
+    strategy_used: StrategyType
+    mode_used: RuntimeMode
+    metadata_filter: MetadataFilterResult | None
+    latency_ms: float
+    diagnostics: dict[str, Any]
+
+
+class RetrievalStrategy:
+    """
+    检索策略统一入口。【AC-AISVC-RES-01~15】
+    
+    整合所有策略组件：
+    1. 元数据推断（MetadataInferenceService）
+    2. 策略路由（StrategyRouter）
+    3. 模式路由（ModeRouter）
+    4. 回退管理（RollbackManager）
+    
+    使用方式：
+    ```python
+    strategy = RetrievalStrategy(session)
+    result = await strategy.retrieve(
+        tenant_id="tenant_1",
+        query="用户问题",
+        context={"user_id": "user_1"},
+    )
+    ```
+    """
+    
+    def __init__(
+        self,
+        session: AsyncSession,
+        config: RetrievalStrategyConfig | None = None,
+    ):
+        self._session = session
+        self._config = config or RetrievalStrategyConfig()
+        
+        self._strategy_router = StrategyRouter(self._config)
+        self._mode_router = ModeRouter(self._config.mode_router)
+        self._rollback_manager = RollbackManager(self._config)
+        self._metadata_inference: MetadataInferenceService | None = None
+    
+    async def retrieve(
+        self,
+        tenant_id: str,
+        query: str,
+        context: dict[str, Any] | None = None,
+        kb_ids: list[str] | None = None,
+        session_id: str | None = None,
+        user_id: str | None = None,
+        use_reranker: bool = False,
+        use_react: bool = False,
+    ) -> RetrievalStrategyResult:
+        """
+        执行检索策略。【AC-AISVC-RES-01~15】
+        
+        流程：
+        1. 元数据推断
+        2. 策略路由
+        3. 模式路由
+        4. 执行检索
+        5. 检查是否需要回退
+        
+        Args:
+            tenant_id: 租户 ID
+            query: 查询文本
+            context: 上下文信息
+            kb_ids: 知识库 ID 列表
+            session_id: 会话 ID
+            user_id: 用户 ID
+            use_reranker: 是否使用重排
+            use_react: 是否使用 ReAct 模式
+        
+        Returns:
+            RetrievalStrategyResult 包含检索结果和诊断信息
+        """
+        start_time = time.time()
+        context = context or {}
+        
+        logger.info(
+            f"[RetrievalStrategy] Starting retrieval: tenant={tenant_id}, "
+            f"query={query[:50]}..."
+        )
+        
+        try:
+            metadata_filter = await self._infer_metadata(
+                tenant_id=tenant_id,
+                query=query,
+                context=context,
+                session_id=session_id,
+                user_id=user_id,
+            )
+            
+            routing_decision = await self._strategy_router.route(tenant_id, user_id)
+            
+            mode_decision = self._mode_router.decide(
+                query=query,
+                confidence=metadata_filter.confidence if metadata_filter else None,
+            )
+            
+            retrieval_ctx = RetrievalContext(
+                tenant_id=tenant_id,
+                query=query,
+                session_id=session_id,
+                metadata_filter=metadata_filter.filter_dict if metadata_filter else None,
+                kb_ids=kb_ids,
+            )
+            
+            pipeline_ctx = PipelineContext(
+                retrieval_ctx=retrieval_ctx,
+                metadata_filter=metadata_filter,
+                use_reranker=use_reranker or self._config.reranker.enabled,
+                use_react=use_react or mode_decision.mode == RuntimeMode.REACT,
+            )
+            
+            pipeline_result = await routing_decision.pipeline.retrieve(pipeline_ctx)
+            
+            if mode_decision.mode == RuntimeMode.DIRECT and self._mode_router.should_fallback_to_react(pipeline_result):
+                logger.info("[RetrievalStrategy] Falling back to react mode")
+                pipeline_ctx.use_react = True
+                pipeline_result = await routing_decision.pipeline.retrieve(pipeline_ctx)
+                mode_decision = ModeDecision(
+                    mode=RuntimeMode.REACT,
+                    reason="fallback_from_direct",
+                )
+            
+            latency_ms = (time.time() - start_time) * 1000
+            
+            self._check_performance(latency_ms, tenant_id)
+            
+            logger.info(
+                f"[RetrievalStrategy] Retrieval completed: strategy={routing_decision.strategy.value}, "
+                f"mode={mode_decision.mode.value}, hits={len(pipeline_result.hits)}, "
+                f"latency_ms={latency_ms:.2f}"
+            )
+            
+            return RetrievalStrategyResult(
+                retrieval_result=pipeline_result.retrieval_result,
+                strategy_used=routing_decision.strategy,
+                mode_used=mode_decision.mode,
+                metadata_filter=metadata_filter,
+                latency_ms=latency_ms,
+                diagnostics={
+                    "routing_reason": routing_decision.reason,
+                    "mode_reason": mode_decision.reason,
+                    "grayscale_hit": routing_decision.grayscale_hit,
+                    **pipeline_result.diagnostics,
+                },
+            )
+            
+        except Exception as e:
+            latency_ms = (time.time() - start_time) * 1000
+            logger.error(
+                f"[RetrievalStrategy] Retrieval error: {e}",
+                exc_info=True
+            )
+            
+            self._rollback_manager.rollback(
+                trigger=RollbackTrigger.ERROR,
+                reason=str(e),
+                tenant_id=tenant_id,
+            )
+            
+            return RetrievalStrategyResult(
+                retrieval_result=RetrievalResult(
+                    hits=[],
+                    diagnostics={"error": str(e)},
+                ),
+                strategy_used=StrategyType.DEFAULT,
+                mode_used=RuntimeMode.DIRECT,
+                metadata_filter=None,
+                latency_ms=latency_ms,
+                diagnostics={"error": str(e)},
+            )
+    
+    async def _infer_metadata(
+        self,
+        tenant_id: str,
+        query: str,
+        context: dict[str, Any],
+        session_id: str | None = None,
+        user_id: str | None = None,
+    ) -> MetadataFilterResult | None:
+        """执行元数据推断。"""
+        try:
+            if self._metadata_inference is None:
+                self._metadata_inference = MetadataInferenceService(
+                    self._session,
+                    self._config.metadata_inference,
+                )
+            
+            inference_ctx = InferenceContext(
+                tenant_id=tenant_id,
+                query=query,
+                session_id=session_id,
+                user_id=user_id,
+                existing_context=context,
+            )
+            
+            result = await self._metadata_inference.infer(inference_ctx)
+            return result.filter_result
+            
+        except Exception as e:
+            logger.warning(f"[RetrievalStrategy] Metadata inference failed: {e}")
+            return None
+    
+    def _check_performance(self, latency_ms: float, tenant_id: str | None) -> None:
+        """检查性能指标，必要时触发回退。"""
+        self._rollback_manager.check_and_rollback(
+            metrics={"latency_ms": latency_ms},
+            tenant_id=tenant_id,
+        )
+    
+    def get_config(self) -> RetrievalStrategyConfig:
+        """获取当前配置。"""
+        return self._config
+    
+    def update_config(self, config: RetrievalStrategyConfig) -> None:
+        """更新配置。"""
+        self._config = config
+        self._strategy_router.update_config(config)
+        self._mode_router.update_config(config.mode_router)
+        self._rollback_manager.update_config(config)
+    
+    async def health_check(self) -> dict[str, bool]:
+        """健康检查。"""
+        results = {}
+        
+        try:
+            default_pipeline = await self._strategy_router._get_default_pipeline()
+            results["default_pipeline"] = await default_pipeline.health_check()
+        except Exception:
+            results["default_pipeline"] = False
+        
+        try:
+            enhanced_pipeline = await self._strategy_router._get_enhanced_pipeline()
+            results["enhanced_pipeline"] = await enhanced_pipeline.health_check()
+        except Exception:
+            results["enhanced_pipeline"] = False
+        
+        return results
+
+
+async def create_retrieval_strategy(
+    session: AsyncSession,
+    config: RetrievalStrategyConfig | None = None,
+) -> RetrievalStrategy:
+    """创建 RetrievalStrategy 实例。"""
+    return RetrievalStrategy(session, config)
--- a/ai-service/app/services/retrieval/strategy/rollback_manager.py
+++ b/ai-service/app/services/retrieval/strategy/rollback_manager.py
@ -0,0 +1,192 @@
+"""
+Rollback Manager.
+[AC-AISVC-RES-07] 策略回退与审计管理器。
+"""
+
+import logging
+from dataclasses import dataclass, field
+from datetime import datetime
+from enum import Enum
+from typing import Any
+
+from app.services.retrieval.strategy.config import RetrievalStrategyConfig, StrategyType
+
+logger = logging.getLogger(__name__)
+
+
+class RollbackTrigger(str, Enum):
+    """回退触发原因。"""
+    MANUAL = "manual"
+    ERROR = "error"
+    PERFORMANCE = "performance"
+    TIMEOUT = "timeout"
+
+
+@dataclass
+class AuditLog:
+    """审计日志记录。"""
+    timestamp: str
+    action: str
+    from_strategy: str
+    to_strategy: str
+    trigger: str
+    reason: str
+    tenant_id: str | None = None
+    details: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class RollbackResult:
+    """回退结果。"""
+    success: bool
+    previous_strategy: StrategyType
+    current_strategy: StrategyType
+    trigger: RollbackTrigger
+    reason: str
+    audit_log: AuditLog | None = None
+
+
+class RollbackManager:
+    """
+    策略回退管理器。【AC-AISVC-RES-07】
+    
+    职责：
+    1. 策略异常时回退到默认策略
+    2. 记录审计日志
+    3. 支持手动触发回退
+    """
+    
+    def __init__(
+        self,
+        config: RetrievalStrategyConfig | None = None,
+        max_audit_logs: int = 1000,
+    ):
+        self._config = config or RetrievalStrategyConfig()
+        self._max_audit_logs = max_audit_logs
+        self._audit_logs: list[AuditLog] = []
+        self._previous_strategy: StrategyType = StrategyType.DEFAULT
+    
+    def rollback(
+        self,
+        trigger: RollbackTrigger,
+        reason: str,
+        tenant_id: str | None = None,
+        details: dict[str, Any] | None = None,
+    ) -> RollbackResult:
+        """执行策略回退。【AC-AISVC-RES-07】"""
+        previous = self._config.active_strategy
+        current = StrategyType.DEFAULT
+        
+        if previous == StrategyType.DEFAULT:
+            return RollbackResult(
+                success=False,
+                previous_strategy=previous,
+                current_strategy=current,
+                trigger=trigger,
+                reason="Already on default strategy",
+            )
+        
+        self._previous_strategy = previous
+        self._config.active_strategy = current
+        
+        audit_log = AuditLog(
+            timestamp=datetime.utcnow().isoformat(),
+            action="rollback",
+            from_strategy=previous.value,
+            to_strategy=current.value,
+            trigger=trigger.value,
+            reason=reason,
+            tenant_id=tenant_id,
+            details=details or {},
+        )
+        
+        self._add_audit_log(audit_log)
+        
+        logger.info(
+            f"[RollbackManager] Rollback executed: from={previous.value}, "
+            f"to={current.value}, trigger={trigger.value}"
+        )
+        
+        return RollbackResult(
+            success=True,
+            previous_strategy=previous,
+            current_strategy=current,
+            trigger=trigger,
+            reason=reason,
+            audit_log=audit_log,
+        )
+    
+    def check_and_rollback(
+        self, metrics: dict[str, float], tenant_id: str | None = None
+    ) -> RollbackResult | None:
+        """检查性能指标并自动回退。【AC-AISVC-RES-08】"""
+        thresholds = self._config.performance_thresholds
+        
+        latency = metrics.get("latency_ms", 0)
+        if latency > thresholds.get("max_latency_ms", 2000):
+            return self.rollback(
+                trigger=RollbackTrigger.PERFORMANCE,
+                reason=f"Latency {latency}ms exceeds threshold",
+                tenant_id=tenant_id,
+            )
+        
+        error_rate = metrics.get("error_rate", 0)
+        if error_rate > thresholds.get("max_error_rate", 0.05):
+            return self.rollback(
+                trigger=RollbackTrigger.ERROR,
+                reason=f"Error rate {error_rate} exceeds threshold",
+                tenant_id=tenant_id,
+            )
+        
+        return None
+    
+    def _add_audit_log(self, log: AuditLog) -> None:
+        self._audit_logs.append(log)
+        if len(self._audit_logs) > self._max_audit_logs:
+            self._audit_logs = self._audit_logs[-self._max_audit_logs:]
+    
+    def record_audit(
+        self,
+        action: str,
+        details: dict[str, Any],
+        tenant_id: str | None = None,
+    ) -> AuditLog:
+        """记录审计日志。"""
+        audit_log = AuditLog(
+            timestamp=datetime.utcnow().isoformat(),
+            action=action,
+            from_strategy=self._config.active_strategy.value,
+            to_strategy=self._config.active_strategy.value,
+            trigger="n/a",
+            reason=details.get("reason", ""),
+            tenant_id=tenant_id,
+            details=details,
+        )
+        
+        self._add_audit_log(audit_log)
+        
+        logger.info(
+            f"[RollbackManager] Audit recorded: action={action}, "
+            f"strategy={self._config.active_strategy.value}"
+        )
+        
+        return audit_log
+    
+    def get_audit_logs(self, limit: int = 100) -> list[AuditLog]:
+        return self._audit_logs[-limit:]
+    
+    def get_config(self) -> RetrievalStrategyConfig:
+        return self._config
+    
+    def update_config(self, config: RetrievalStrategyConfig) -> None:
+        self._config = config
+
+
+_rollback_manager: RollbackManager | None = None
+
+
+def get_rollback_manager() -> RollbackManager:
+    global _rollback_manager
+    if _rollback_manager is None:
+        _rollback_manager = RollbackManager()
+    return _rollback_manager
--- a/ai-service/app/services/retrieval/strategy/strategy_router.py
+++ b/ai-service/app/services/retrieval/strategy/strategy_router.py
@ -0,0 +1,109 @@
+"""
+Strategy Router.
+[AC-AISVC-RES-01~03] 策略路由器。
+"""
+
+import logging
+from dataclasses import dataclass
+from typing import Any
+
+from app.services.retrieval.strategy.config import (
+    RetrievalStrategyConfig,
+    StrategyType,
+)
+from app.services.retrieval.strategy.default_pipeline import DefaultPipeline, get_default_pipeline
+from app.services.retrieval.strategy.enhanced_pipeline import EnhancedPipeline, get_enhanced_pipeline
+from app.services.retrieval.strategy.pipeline_base import BasePipeline
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RoutingDecision:
+    """路由决策结果。"""
+    strategy: StrategyType
+    pipeline: BasePipeline
+    reason: str
+    grayscale_hit: bool = False
+
+
+class StrategyRouter:
+    """
+    策略路由器。【AC-AISVC-RES-01~03】
+    
+    职责：
+    1. 根据配置选择默认策略或增强策略
+    2. 支持灰度发布（percentage/allowlist）
+    3. 不影响正在运行的默认策略请求
+    """
+    
+    def __init__(
+        self,
+        config: RetrievalStrategyConfig | None = None,
+        default_pipeline: DefaultPipeline | None = None,
+        enhanced_pipeline: EnhancedPipeline | None = None,
+    ):
+        self._config = config or RetrievalStrategyConfig()
+        self._default_pipeline = default_pipeline
+        self._enhanced_pipeline = enhanced_pipeline
+    
+    async def route(
+        self, tenant_id: str, user_id: str | None = None
+    ) -> RoutingDecision:
+        """路由到合适的策略。【AC-AISVC-RES-01~03】"""
+        if self._config.active_strategy == StrategyType.ENHANCED:
+            pipeline = await self._get_enhanced_pipeline()
+            return RoutingDecision(
+                strategy=StrategyType.ENHANCED,
+                pipeline=pipeline,
+                reason="active_strategy=enhanced",
+            )
+        
+        if self._config.grayscale.should_use_enhanced(tenant_id, user_id):
+            pipeline = await self._get_enhanced_pipeline()
+            return RoutingDecision(
+                strategy=StrategyType.ENHANCED,
+                pipeline=pipeline,
+                reason="grayscale_hit",
+                grayscale_hit=True,
+            )
+        
+        pipeline = await self._get_default_pipeline()
+        return RoutingDecision(
+            strategy=StrategyType.DEFAULT,
+            pipeline=pipeline,
+            reason="default_strategy",
+        )
+    
+    async def _get_default_pipeline(self) -> DefaultPipeline:
+        if self._default_pipeline is None:
+            self._default_pipeline = await get_default_pipeline()
+        return self._default_pipeline
+    
+    async def _get_enhanced_pipeline(self) -> EnhancedPipeline:
+        if self._enhanced_pipeline is None:
+            self._enhanced_pipeline = await get_enhanced_pipeline()
+        return self._enhanced_pipeline
+    
+    def get_config(self) -> RetrievalStrategyConfig:
+        return self._config
+    
+    def update_config(self, config: RetrievalStrategyConfig) -> None:
+        self._config = config
+        logger.info(f"[StrategyRouter] Config updated: strategy={config.active_strategy.value}")
+
+
+_strategy_router: StrategyRouter | None = None
+
+
+def get_strategy_router() -> StrategyRouter:
+    global _strategy_router
+    if _strategy_router is None:
+        _strategy_router = StrategyRouter()
+    return _strategy_router
+
+
+def set_strategy_router(router: StrategyRouter) -> None:
+    """Set the global strategy router instance."""
+    global _strategy_router
+    _strategy_router = router
--- a/ai-service/tests/test_retrieval_strategy_v2.py
+++ b/ai-service/tests/test_retrieval_strategy_v2.py
@ -0,0 +1,645 @@
+"""
+Unit tests for Retrieval Strategy Module.
+[AC-AISVC-RES-01~15] Tests for strategy config, pipelines, routers, and rollback.
+"""
+
+import pytest
+from unittest.mock import AsyncMock, MagicMock, patch
+from dataclasses import asdict
+
+from app.services.retrieval.strategy.config import (
+    FilterMode,
+    GrayscaleConfig,
+    HybridRetrievalConfig,
+    MetadataInferenceConfig,
+    ModeRouterConfig,
+    PipelineConfig,
+    RerankerConfig,
+    RetrievalStrategyConfig,
+    RuntimeMode,
+    StrategyType,
+    get_strategy_config,
+    set_strategy_config,
+)
+from app.services.retrieval.strategy.pipeline_base import (
+    BasePipeline,
+    MetadataFilterResult,
+    PipelineContext,
+    PipelineResult,
+)
+from app.services.retrieval.strategy.default_pipeline import DefaultPipeline
+from app.services.retrieval.strategy.enhanced_pipeline import EnhancedPipeline
+from app.services.retrieval.strategy.strategy_router import (
+    RoutingDecision,
+    StrategyRouter,
+    get_strategy_router,
+)
+from app.services.retrieval.strategy.mode_router import (
+    ModeDecision,
+    ModeRouter,
+    get_mode_router,
+)
+from app.services.retrieval.strategy.rollback_manager import (
+    AuditLog,
+    RollbackManager,
+    RollbackResult,
+    RollbackTrigger,
+    get_rollback_manager,
+)
+from app.services.retrieval.base import RetrievalContext, RetrievalHit, RetrievalResult
+
+
+class TestStrategyConfig:
+    """[AC-AISVC-RES-01~15] Tests for strategy configuration models."""
+
+    def test_strategy_type_enum(self):
+        """[AC-AISVC-RES-01] Strategy type should have default and enhanced values."""
+        assert StrategyType.DEFAULT.value == "default"
+        assert StrategyType.ENHANCED.value == "enhanced"
+
+    def test_runtime_mode_enum(self):
+        """[AC-AISVC-RES-09] Runtime mode should have direct, react, and auto values."""
+        assert RuntimeMode.DIRECT.value == "direct"
+        assert RuntimeMode.REACT.value == "react"
+        assert RuntimeMode.AUTO.value == "auto"
+
+    def test_filter_mode_enum(self):
+        """[AC-AISVC-RES-04] Filter mode should have hard, soft, and none values."""
+        assert FilterMode.HARD.value == "hard"
+        assert FilterMode.SOFT.value == "soft"
+        assert FilterMode.NONE.value == "none"
+
+    def test_grayscale_config_default(self):
+        """[AC-AISVC-RES-03] Default grayscale config should be disabled."""
+        config = GrayscaleConfig()
+        assert config.enabled is False
+        assert config.percentage == 0.0
+        assert config.allowlist == []
+
+    def test_grayscale_config_should_use_enhanced_disabled(self):
+        """[AC-AISVC-RES-03] Should not use enhanced when grayscale disabled."""
+        config = GrayscaleConfig(enabled=False, percentage=50.0)
+        assert config.should_use_enhanced("tenant_a") is False
+
+    def test_grayscale_config_should_use_enhanced_allowlist(self):
+        """[AC-AISVC-RES-03] Should use enhanced for tenants in allowlist."""
+        config = GrayscaleConfig(enabled=True, allowlist=["tenant_a", "tenant_b"])
+        assert config.should_use_enhanced("tenant_a") is True
+        assert config.should_use_enhanced("tenant_b") is True
+        assert config.should_use_enhanced("tenant_c") is False
+
+    def test_grayscale_config_should_use_enhanced_percentage(self):
+        """[AC-AISVC-RES-03] Should use enhanced based on percentage."""
+        config = GrayscaleConfig(enabled=True, percentage=100.0)
+        assert config.should_use_enhanced("any_tenant") is True
+
+        config = GrayscaleConfig(enabled=True, percentage=0.0)
+        assert config.should_use_enhanced("any_tenant") is False
+
+    def test_reranker_config_default(self):
+        """[AC-AISVC-RES-08] Default reranker config should be disabled."""
+        config = RerankerConfig()
+        assert config.enabled is False
+        assert config.model == "cross-encoder"
+        assert config.top_k_after_rerank == 5
+
+    def test_mode_router_config_default(self):
+        """[AC-AISVC-RES-09] Default mode router config should be direct."""
+        config = ModeRouterConfig()
+        assert config.runtime_mode == RuntimeMode.DIRECT
+        assert config.react_trigger_confidence_threshold == 0.6
+        assert config.react_max_steps == 5
+
+    def test_mode_router_config_should_use_react_always(self):
+        """[AC-AISVC-RES-10] React mode should always use react."""
+        config = ModeRouterConfig(runtime_mode=RuntimeMode.REACT)
+        assert config.should_use_react("any query") is True
+
+    def test_mode_router_config_should_use_react_never(self):
+        """[AC-AISVC-RES-09] Direct mode should never use react."""
+        config = ModeRouterConfig(runtime_mode=RuntimeMode.DIRECT)
+        assert config.should_use_react("any query") is False
+
+    def test_mode_router_config_auto_short_query_high_confidence(self):
+        """[AC-AISVC-RES-12] Auto mode with short query and high confidence should use direct."""
+        config = ModeRouterConfig(runtime_mode=RuntimeMode.AUTO)
+        assert config.should_use_react("短问题", confidence=0.8) is False
+
+    def test_mode_router_config_auto_low_confidence(self):
+        """[AC-AISVC-RES-13] Auto mode with low confidence should use react."""
+        config = ModeRouterConfig(runtime_mode=RuntimeMode.AUTO)
+        assert config.should_use_react("any query", confidence=0.3) is True
+
+    def test_metadata_inference_config_determine_filter_mode(self):
+        """[AC-AISVC-RES-04] Should determine filter mode based on confidence."""
+        config = MetadataInferenceConfig()
+        
+        assert config.determine_filter_mode(0.9) == FilterMode.HARD
+        assert config.determine_filter_mode(0.6) == FilterMode.SOFT
+        assert config.determine_filter_mode(0.3) == FilterMode.NONE
+        assert config.determine_filter_mode(None) == FilterMode.NONE
+
+    def test_pipeline_config_default(self):
+        """[AC-AISVC-RES-01] Default pipeline config should have sensible defaults."""
+        config = PipelineConfig()
+        assert config.top_k == 5
+        assert config.score_threshold == 0.01
+        assert config.two_stage_enabled is True
+
+    def test_retrieval_strategy_config_default(self):
+        """[AC-AISVC-RES-01] Default strategy config should use default strategy."""
+        config = RetrievalStrategyConfig()
+        assert config.active_strategy == StrategyType.DEFAULT
+        assert config.grayscale.enabled is False
+        assert config.mode_router.runtime_mode == RuntimeMode.DIRECT
+
+    def test_retrieval_strategy_config_is_enhanced_enabled(self):
+        """[AC-AISVC-RES-02] Should check if enhanced is enabled."""
+        config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        assert config.is_enhanced_enabled("tenant_a") is True
+
+        config = RetrievalStrategyConfig(
+            active_strategy=StrategyType.DEFAULT,
+            grayscale=GrayscaleConfig(enabled=True, allowlist=["tenant_a"]),
+        )
+        assert config.is_enhanced_enabled("tenant_a") is True
+        assert config.is_enhanced_enabled("tenant_b") is False
+
+    def test_retrieval_strategy_config_to_dict(self):
+        """[AC-AISVC-RES-01] Should convert config to dictionary."""
+        config = RetrievalStrategyConfig()
+        d = config.to_dict()
+        
+        assert d["active_strategy"] == "default"
+        assert "grayscale" in d
+        assert "pipeline" in d
+        assert "reranker" in d
+        assert "mode_router" in d
+
+    def test_global_config_functions(self):
+        """[AC-AISVC-RES-01] Should get and set global config."""
+        config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        set_strategy_config(config)
+        
+        retrieved = get_strategy_config()
+        assert retrieved.active_strategy == StrategyType.ENHANCED
+        
+        set_strategy_config(RetrievalStrategyConfig())
+
+
+class TestPipelineBase:
+    """[AC-AISVC-RES-01~02] Tests for pipeline base classes."""
+
+    def test_metadata_filter_result_default(self):
+        """[AC-AISVC-RES-04] Default metadata filter result should be empty."""
+        result = MetadataFilterResult()
+        assert result.filter_dict == {}
+        assert result.filter_mode == FilterMode.NONE
+        assert result.confidence is None
+
+    def test_pipeline_context_properties(self):
+        """[AC-AISVC-RES-01] Pipeline context should expose retrieval context properties."""
+        retrieval_ctx = RetrievalContext(
+            tenant_id="tenant_1",
+            query="test query",
+            session_id="session_1",
+            kb_ids=["kb_1"],
+        )
+        pipeline_ctx = PipelineContext(retrieval_ctx=retrieval_ctx)
+        
+        assert pipeline_ctx.tenant_id == "tenant_1"
+        assert pipeline_ctx.query == "test query"
+        assert pipeline_ctx.session_id == "session_1"
+        assert pipeline_ctx.kb_ids == ["kb_1"]
+
+    def test_pipeline_result_properties(self):
+        """[AC-AISVC-RES-01] Pipeline result should expose retrieval result properties."""
+        hits = [
+            RetrievalHit(text="hit 1", score=0.9, source="test", metadata={}),
+            RetrievalHit(text="hit 2", score=0.8, source="test", metadata={}),
+        ]
+        retrieval_result = RetrievalResult(hits=hits)
+        pipeline_result = PipelineResult(
+            retrieval_result=retrieval_result,
+            pipeline_name="test_pipeline",
+        )
+        
+        assert pipeline_result.hits == hits
+        assert pipeline_result.is_empty is False
+        assert pipeline_result.pipeline_name == "test_pipeline"
+
+    def test_pipeline_result_is_empty(self):
+        """[AC-AISVC-RES-01] Pipeline result should detect empty results."""
+        pipeline_result = PipelineResult(
+            retrieval_result=RetrievalResult(hits=[]),
+        )
+        assert pipeline_result.is_empty is True
+
+
+class TestDefaultPipeline:
+    """[AC-AISVC-RES-01] Tests for default pipeline."""
+
+    @pytest.fixture
+    def mock_retriever(self):
+        """Create a mock optimized retriever."""
+        retriever = AsyncMock()
+        retriever.retrieve = AsyncMock(return_value=RetrievalResult(
+            hits=[
+                RetrievalHit(text="result 1", score=0.9, source="default", metadata={}),
+            ],
+            diagnostics={"test": True},
+        ))
+        retriever.health_check = AsyncMock(return_value=True)
+        retriever._two_stage_enabled = True
+        retriever._hybrid_enabled = True
+        return retriever
+
+    @pytest.fixture
+    def pipeline(self, mock_retriever):
+        """Create a default pipeline with mock retriever."""
+        return DefaultPipeline(optimized_retriever=mock_retriever)
+
+    def test_pipeline_name(self, pipeline):
+        """[AC-AISVC-RES-01] Pipeline should have correct name."""
+        assert pipeline.name == "default_pipeline"
+
+    def test_pipeline_description(self, pipeline):
+        """[AC-AISVC-RES-01] Pipeline should have description."""
+        assert "默认" in pipeline.description
+
+    @pytest.mark.asyncio
+    async def test_retrieve(self, pipeline, mock_retriever):
+        """[AC-AISVC-RES-01] Should retrieve results using optimized retriever."""
+        retrieval_ctx = RetrievalContext(
+            tenant_id="tenant_1",
+            query="test query",
+        )
+        pipeline_ctx = PipelineContext(retrieval_ctx=retrieval_ctx)
+        
+        result = await pipeline.retrieve(pipeline_ctx)
+        
+        assert result.pipeline_name == "default_pipeline"
+        assert len(result.hits) == 1
+        assert result.diagnostics["retriever"] == "OptimizedRetriever"
+        mock_retriever.retrieve.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_retrieve_with_metadata_filter(self, pipeline, mock_retriever):
+        """[AC-AISVC-RES-04] Should apply metadata filter."""
+        retrieval_ctx = RetrievalContext(
+            tenant_id="tenant_1",
+            query="test query",
+        )
+        metadata_filter = MetadataFilterResult(
+            filter_dict={"grade": "初一"},
+            filter_mode=FilterMode.HARD,
+        )
+        pipeline_ctx = PipelineContext(
+            retrieval_ctx=retrieval_ctx,
+            metadata_filter=metadata_filter,
+        )
+        
+        result = await pipeline.retrieve(pipeline_ctx)
+        
+        assert result.metadata_filter_applied is True
+        call_args = mock_retriever.retrieve.call_args[0][0]
+        assert call_args.metadata_filter == {"grade": "初一"}
+
+    @pytest.mark.asyncio
+    async def test_health_check(self, pipeline, mock_retriever):
+        """[AC-AISVC-RES-01] Should check health."""
+        result = await pipeline.health_check()
+        assert result is True
+        mock_retriever.health_check.assert_called_once()
+
+
+class TestEnhancedPipeline:
+    """[AC-AISVC-RES-02] Tests for enhanced pipeline."""
+
+    @pytest.fixture
+    def mock_qdrant_client(self):
+        """Create a mock Qdrant client."""
+        client = AsyncMock()
+        client.search = AsyncMock(return_value=[
+            {"id": "1", "score": 0.9, "payload": {"text": "result 1"}},
+        ])
+        client.get_client = AsyncMock()
+        return client
+
+    @pytest.fixture
+    def mock_embedding_provider(self):
+        """Create a mock embedding provider."""
+        provider = AsyncMock()
+        provider.embed_query = AsyncMock()
+        provider.embed_query.return_value = MagicMock(
+            embedding_full=[0.1] * 768,
+        )
+        provider.embed = AsyncMock(return_value=[0.1] * 768)
+        return provider
+
+    @pytest.fixture
+    def pipeline(self, mock_qdrant_client, mock_embedding_provider):
+        """Create an enhanced pipeline with mocks."""
+        pipeline = EnhancedPipeline(qdrant_client=mock_qdrant_client)
+        pipeline._embedding_provider = mock_embedding_provider
+        return pipeline
+
+    def test_pipeline_name(self, pipeline):
+        """[AC-AISVC-RES-02] Pipeline should have correct name."""
+        assert pipeline.name == "enhanced_pipeline"
+
+    def test_pipeline_description(self, pipeline):
+        """[AC-AISVC-RES-02] Pipeline should have description."""
+        assert "增强" in pipeline.description
+
+    @pytest.mark.asyncio
+    async def test_retrieve_basic(self, pipeline):
+        """[AC-AISVC-RES-02] Should retrieve results using hybrid search."""
+        retrieval_ctx = RetrievalContext(
+            tenant_id="tenant_1",
+            query="test query",
+        )
+        pipeline_ctx = PipelineContext(retrieval_ctx=retrieval_ctx)
+        
+        result = await pipeline.retrieve(pipeline_ctx)
+        
+        assert result.pipeline_name == "enhanced_pipeline"
+        assert result.diagnostics is not None
+
+
+class TestStrategyRouter:
+    """[AC-AISVC-RES-01~03] Tests for strategy router."""
+
+    @pytest.fixture
+    def mock_default_pipeline(self):
+        """Create a mock default pipeline."""
+        pipeline = AsyncMock(spec=DefaultPipeline)
+        pipeline.name = "default_pipeline"
+        pipeline.retrieve = AsyncMock(return_value=PipelineResult(
+            retrieval_result=RetrievalResult(hits=[]),
+            pipeline_name="default_pipeline",
+        ))
+        return pipeline
+
+    @pytest.fixture
+    def mock_enhanced_pipeline(self):
+        """Create a mock enhanced pipeline."""
+        pipeline = AsyncMock(spec=EnhancedPipeline)
+        pipeline.name = "enhanced_pipeline"
+        pipeline.retrieve = AsyncMock(return_value=PipelineResult(
+            retrieval_result=RetrievalResult(hits=[]),
+            pipeline_name="enhanced_pipeline",
+        ))
+        return pipeline
+
+    @pytest.fixture
+    def router(self, mock_default_pipeline, mock_enhanced_pipeline):
+        """Create a strategy router with mock pipelines."""
+        config = RetrievalStrategyConfig()
+        return StrategyRouter(
+            config=config,
+            default_pipeline=mock_default_pipeline,
+            enhanced_pipeline=mock_enhanced_pipeline,
+        )
+
+    def test_route_default_strategy(self, router):
+        """[AC-AISVC-RES-01] Should route to default strategy by default."""
+        import asyncio
+        decision = asyncio.get_event_loop().run_until_complete(router.route("tenant_1"))
+        
+        assert decision.strategy == StrategyType.DEFAULT
+        assert decision.reason == "default_strategy"
+
+    def test_route_enhanced_strategy(self, mock_default_pipeline, mock_enhanced_pipeline):
+        """[AC-AISVC-RES-02] Should route to enhanced strategy when configured."""
+        config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        router = StrategyRouter(
+            config=config,
+            default_pipeline=mock_default_pipeline,
+            enhanced_pipeline=mock_enhanced_pipeline,
+        )
+        
+        import asyncio
+        decision = asyncio.get_event_loop().run_until_complete(router.route("tenant_1"))
+        
+        assert decision.strategy == StrategyType.ENHANCED
+        assert decision.reason == "active_strategy=enhanced"
+
+    def test_route_grayscale_allowlist(self, mock_default_pipeline, mock_enhanced_pipeline):
+        """[AC-AISVC-RES-03] Should route to enhanced for allowlist tenants."""
+        config = RetrievalStrategyConfig(
+            active_strategy=StrategyType.DEFAULT,
+            grayscale=GrayscaleConfig(enabled=True, allowlist=["tenant_a"]),
+        )
+        router = StrategyRouter(
+            config=config,
+            default_pipeline=mock_default_pipeline,
+            enhanced_pipeline=mock_enhanced_pipeline,
+        )
+        
+        import asyncio
+        
+        decision = asyncio.get_event_loop().run_until_complete(router.route("tenant_a"))
+        assert decision.strategy == StrategyType.ENHANCED
+        assert decision.grayscale_hit is True
+        
+        decision = asyncio.get_event_loop().run_until_complete(router.route("tenant_b"))
+        assert decision.strategy == StrategyType.DEFAULT
+
+    def test_update_config(self, router):
+        """[AC-AISVC-RES-02] Should update config."""
+        new_config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        router.update_config(new_config)
+        
+        assert router.get_config().active_strategy == StrategyType.ENHANCED
+
+
+class TestModeRouter:
+    """[AC-AISVC-RES-09~15] Tests for mode router."""
+
+    @pytest.fixture
+    def router(self):
+        """Create a mode router."""
+        return ModeRouter()
+
+    def test_decide_react_mode(self):
+        """[AC-AISVC-RES-10] Should decide react when configured."""
+        config = ModeRouterConfig(runtime_mode=RuntimeMode.REACT)
+        router = ModeRouter(config)
+        
+        decision = router.decide("any query")
+        
+        assert decision.mode == RuntimeMode.REACT
+        assert decision.reason == "runtime_mode=react"
+
+    def test_decide_direct_mode(self, router):
+        """[AC-AISVC-RES-09] Should decide direct when configured."""
+        decision = router.decide("any query")
+        
+        assert decision.mode == RuntimeMode.DIRECT
+        assert decision.reason == "runtime_mode=direct"
+
+    def test_decide_auto_short_query_high_confidence(self):
+        """[AC-AISVC-RES-12] Auto with short query and high confidence should use direct."""
+        config = ModeRouterConfig(runtime_mode=RuntimeMode.AUTO)
+        router = ModeRouter(config)
+        
+        decision = router.decide("短问题", confidence=0.8)
+        
+        assert decision.mode == RuntimeMode.DIRECT
+
+    def test_decide_auto_low_confidence(self):
+        """[AC-AISVC-RES-13] Auto with low confidence should use react."""
+        config = ModeRouterConfig(runtime_mode=RuntimeMode.AUTO)
+        router = ModeRouter(config)
+        
+        decision = router.decide("any query", confidence=0.3)
+        
+        assert decision.mode == RuntimeMode.REACT
+
+    def test_should_fallback_to_react_empty_results(self, router):
+        """[AC-AISVC-RES-14] Should fallback to react on empty results."""
+        result = PipelineResult(retrieval_result=RetrievalResult(hits=[]))
+        
+        assert router.should_fallback_to_react(result) is True
+
+    def test_should_fallback_to_react_low_score(self, router):
+        """[AC-AISVC-RES-14] Should fallback to react on low score."""
+        result = PipelineResult(
+            retrieval_result=RetrievalResult(
+                hits=[RetrievalHit(text="test", score=0.1, source="test", metadata={})],
+            ),
+        )
+        
+        assert router.should_fallback_to_react(result) is True
+
+    def test_should_not_fallback_to_react_disabled(self):
+        """[AC-AISVC-RES-14] Should not fallback when disabled."""
+        config = ModeRouterConfig(direct_fallback_on_low_confidence=False)
+        router = ModeRouter(config)
+        
+        result = PipelineResult(retrieval_result=RetrievalResult(hits=[]))
+        
+        assert router.should_fallback_to_react(result) is False
+
+
+class TestRollbackManager:
+    """[AC-AISVC-RES-07] Tests for rollback manager."""
+
+    @pytest.fixture
+    def manager(self):
+        """Create a rollback manager."""
+        return RollbackManager()
+
+    def test_rollback_from_enhanced(self, manager):
+        """[AC-AISVC-RES-07] Should rollback from enhanced to default."""
+        config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        manager.update_config(config)
+        
+        result = manager.rollback(
+            trigger=RollbackTrigger.MANUAL,
+            reason="Testing rollback",
+        )
+        
+        assert result.success is True
+        assert result.previous_strategy == StrategyType.ENHANCED
+        assert result.current_strategy == StrategyType.DEFAULT
+        assert result.audit_log is not None
+
+    def test_rollback_already_default(self, manager):
+        """[AC-AISVC-RES-07] Should not rollback when already on default."""
+        result = manager.rollback(
+            trigger=RollbackTrigger.MANUAL,
+            reason="Testing rollback",
+        )
+        
+        assert result.success is False
+        assert result.reason == "Already on default strategy"
+
+    def test_check_and_rollback_latency(self, manager):
+        """[AC-AISVC-RES-08] Should rollback on high latency."""
+        config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        manager.update_config(config)
+        
+        result = manager.check_and_rollback(
+            metrics={"latency_ms": 3000.0},
+            tenant_id="tenant_1",
+        )
+        
+        assert result is not None
+        assert result.trigger == RollbackTrigger.PERFORMANCE
+
+    def test_check_and_rollback_error_rate(self, manager):
+        """[AC-AISVC-RES-08] Should rollback on high error rate."""
+        config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        manager.update_config(config)
+        
+        result = manager.check_and_rollback(
+            metrics={"error_rate": 0.1},
+            tenant_id="tenant_1",
+        )
+        
+        assert result is not None
+        assert result.trigger == RollbackTrigger.ERROR
+
+    def test_check_and_rollback_ok(self, manager):
+        """[AC-AISVC-RES-08] Should not rollback when metrics are ok."""
+        config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        manager.update_config(config)
+        
+        result = manager.check_and_rollback(
+            metrics={"latency_ms": 100.0, "error_rate": 0.01},
+            tenant_id="tenant_1",
+        )
+        
+        assert result is None
+
+    def test_get_audit_logs(self, manager):
+        """[AC-AISVC-RES-07] Should get audit logs."""
+        config = RetrievalStrategyConfig(active_strategy=StrategyType.ENHANCED)
+        manager.update_config(config)
+        manager.rollback(trigger=RollbackTrigger.MANUAL, reason="Test")
+        
+        logs = manager.get_audit_logs()
+        
+        assert len(logs) == 1
+        assert logs[0].action == "rollback"
+
+    def test_record_audit(self, manager):
+        """[AC-AISVC-RES-07] Should record audit log."""
+        log = manager.record_audit(
+            action="test_action",
+            details={"reason": "Testing"},
+            tenant_id="tenant_1",
+        )
+        
+        assert log.action == "test_action"
+        assert log.tenant_id == "tenant_1"
+
+
+class TestSingletonInstances:
+    """Tests for singleton instance getters."""
+
+    def test_get_mode_router_singleton(self):
+        """Should return same mode router instance."""
+        from app.services.retrieval.strategy.mode_router import _mode_router
+        
+        import app.services.retrieval.strategy.mode_router as module
+        module._mode_router = None
+        
+        router1 = get_mode_router()
+        router2 = get_mode_router()
+        
+        assert router1 is router2
+
+    def test_get_rollback_manager_singleton(self):
+        """Should return same rollback manager instance."""
+        from app.services.retrieval.strategy.rollback_manager import _rollback_manager
+        
+        import app.services.retrieval.strategy.rollback_manager as module
+        module._rollback_manager = None
+        
+        manager1 = get_rollback_manager()
+        manager2 = get_rollback_manager()
+        
+        assert manager1 is manager2
--- a/docs/progress/ai-service-AISVC-RES-progress.md
+++ b/docs/progress/ai-service-AISVC-RES-progress.md
@ -0,0 +1,167 @@
+---
+context:
+  module: "ai-service"
+  feature: "AISVC-RES"
+  status: "✅已完成"
+  version: "0.9.0"
+  active_ac_range: "AC-AISVC-RES-01~15"
+
+spec_references:
+  requirements: "spec/ai-service/iterations/v0.9.0-retrieval-embedding-strategy/requirements.md"
+  design: "spec/ai-service/iterations/v0.9.0-retrieval-embedding-strategy/design.md"
+  tasks: "spec/ai-service/iterations/v0.9.0-retrieval-embedding-strategy/tasks.md"
+  openapi_provider: "spec/ai-service/iterations/v0.9.0-retrieval-embedding-strategy/openapi.provider.yaml"
+  active_version: "0.1.0"
+
+overall_progress:
+  - "[x] Phase 1: Schema与数据模型定义 (100%) [API与校验]"
+  - "[x] Phase 2: 策略服务层实现 (100%) [策略层与配置]"
+  - "[x] Phase 3: 审计日志与指标埋点 (100%) [观测与灰度验证]"
+  - "[x] Phase 4: API端点实现 (100%) [API与校验]"
+  - "[x] Phase 5: 单元测试与验证 (100%) [验收]"
+  - "[x] Phase 6: 检索策略Pipeline实现 (100%) [检索与嵌入策略]"
+
+current_phase:
+  goal: "检索与嵌入策略化改造已完成"
+  sub_tasks:
+    - "[x] 创建 Schema 模型 (app/schemas/retrieval_strategy.py)"
+    - "[x] 创建策略服务层 (app/services/retrieval/strategy_service.py)"
+    - "[x] 创建审计日志服务 (app/services/retrieval/strategy_audit.py)"
+    - "[x] 创建指标埋点服务 (app/services/retrieval/strategy_metrics.py)"
+    - "[x] 创建 API 端点 (app/api/admin/retrieval_strategy.py)"
+    - "[x] 创建策略配置模型 (app/services/retrieval/strategy/config.py)"
+    - "[x] 实现 DefaultPipeline（复用现有逻辑）"
+    - "[x] 实现 EnhancedPipeline（新端到端流程）"
+    - "[x] 实现元数据推断统一入口 (MetadataInferenceService)"
+    - "[x] 实现 StrategyRouter 和 ModeRouter"
+    - "[x] 实现 Dense + Keyword + RRF 组合检索"
+    - "[x] 实现可选重排与降级开关"
+    - "[x] 实现 RollbackManager（回退与审计）"
+    - "[x] 实现策略 API 接口"
+    - "[x] 创建单元测试 (tests/test_retrieval_strategy_v2.py)"
+    - "[x] 运行单元测试验证 (51 passed)"
+  
+  next_action:
+    immediate: "任务已完成，可进行集成测试"
+    details:
+      file: "ai-service/app/services/retrieval/strategy/__init__.py:1"
+      action: "模块已完整实现，可通过 API 接口测试策略切换功能"
+      reference: "http://localhost:8000/docs"
+    constraints: "新策略可配置启用，不影响默认策略"
+
+technical_context:
+  module_structure: |
+    ai-service/app/
+    ├── api/admin/retrieval_strategy.py      (API端点)
+    ├── schemas/retrieval_strategy.py        (Schema模型)
+    └── services/retrieval/
+        ├── strategy_service.py              (策略服务)
+        ├── strategy_audit.py                (审计日志)
+        ├── strategy_metrics.py              (指标埋点)
+        └── strategy/                        (新增 - 策略模块)
+            ├── __init__.py                  (模块导出)
+            ├── config.py                    (策略配置模型)
+            ├── pipeline_base.py             (Pipeline基类)
+            ├── default_pipeline.py          (默认策略Pipeline)
+            ├── enhanced_pipeline.py         (增强策略Pipeline)
+            ├── metadata_inference.py        (元数据推断统一入口)
+            ├── strategy_router.py           (策略路由器)
+            ├── mode_router.py               (模式路由器)
+            └── rollback_manager.py          (回退管理器)
+    └── tests/
+        ├── test_retrieval_strategy.py       (单元测试 - 原有)
+        └── test_retrieval_strategy_v2.py    (单元测试 - 新增)
+  key_decisions:
+    - decision: "使用内存存储策略状态，后续可扩展为持久化"
+      reason: "快速实现，满足灰度验证需求"
+      impact: "服务重启后策略状态重置为默认值"
+    - decision: "审计日志使用结构化日志记录"
+      reason: "与现有日志体系一致，便于检索"
+      impact: "需要配置日志聚合系统收集审计日志"
+    - decision: "API与现有strategy_router.py互补"
+      reason: "strategy_router.py负责检索路由逻辑，新增的API负责策略管理"
+      impact: "两者协同工作，API提供管理界面"
+    - decision: "DefaultPipeline 复用现有 OptimizedRetriever 逻辑"
+      reason: "保持线上行为不变，最小化改动风险"
+      impact: "新策略与旧策略完全隔离，可独立灰度"
+    - decision: "EnhancedPipeline 实现新端到端流程"
+      reason: "支持 Dense + Keyword + RRF 组合检索，可选重排"
+      impact: "需要配置启用，不影响默认策略"
+    - decision: "元数据推断统一入口处理 hard/soft filter"
+      reason: "新旧策略共享同一推断逻辑，确保一致性"
+      impact: "置信度高用硬过滤，置信度低用软过滤/加权"
+  code_snippets: |
+    # 使用示例
+    from app.services.retrieval.strategy import (
+        get_strategy_router,
+        get_mode_router,
+        get_rollback_manager,
+    )
+    
+    # 策略路由
+    router = get_strategy_router()
+    decision = await router.route(tenant_id, user_id)
+    result = await decision.pipeline.retrieve(ctx)
+    
+    # 模式路由
+    mode_router = get_mode_router()
+    mode_decision = mode_router.decide(query, confidence=0.8)
+    
+    # 回退管理
+    rollback = get_rollback_manager()
+    rollback.rollback(trigger="manual", reason="测试回退")
+
+session_history:
+  - session: "Session #1 (2026-03-10)"
+    completed:
+      - "创建 Schema 模型 (app/schemas/retrieval_strategy.py)"
+      - "创建策略服务层 (app/services/retrieval/strategy_service.py)"
+      - "创建审计日志服务 (app/services/retrieval/strategy_audit.py)"
+      - "创建指标埋点服务 (app/services/retrieval/strategy_metrics.py)"
+      - "创建 API 端点 (app/api/admin/retrieval_strategy.py)"
+      - "更新 admin __init__.py 注册新路由"
+      - "更新 main.py 注册新路由"
+      - "创建单元测试 (tests/test_retrieval_strategy.py)"
+      - "运行单元测试验证 (46 passed)"
+    changes:
+      - "新增: ai-service/app/schemas/retrieval_strategy.py"
+      - "新增: ai-service/app/services/retrieval/strategy_service.py"
+      - "新增: ai-service/app/services/retrieval/strategy_audit.py"
+      - "新增: ai-service/app/services/retrieval/strategy_metrics.py"
+      - "新增: ai-service/app/api/admin/retrieval_strategy.py"
+      - "修改: ai-service/app/api/admin/__init__.py"
+      - "修改: ai-service/app/main.py"
+      - "新增: ai-service/tests/test_retrieval_strategy.py"
+    status: "✅ 任务完成"
+  - session: "Session #2 (2026-03-10) - 检索策略Pipeline实现"
+    completed:
+      - "实现策略配置模型 (config.py)"
+      - "实现 Pipeline 基类 (pipeline_base.py)"
+      - "实现 DefaultPipeline（复用现有逻辑）"
+      - "实现 EnhancedPipeline（新端到端流程）"
+      - "实现 MetadataInferenceService"
+      - "实现 StrategyRouter"
+      - "实现 ModeRouter"
+      - "实现 RollbackManager"
+      - "更新模块导出 (__init__.py)"
+      - "创建单元测试 (tests/test_retrieval_strategy_v2.py)"
+      - "运行单元测试验证 (51 passed)"
+    changes:
+      - "新增: ai-service/app/services/retrieval/strategy/__init__.py"
+      - "新增: ai-service/app/services/retrieval/strategy/config.py"
+      - "新增: ai-service/app/services/retrieval/strategy/pipeline_base.py"
+      - "新增: ai-service/app/services/retrieval/strategy/default_pipeline.py"
+      - "新增: ai-service/app/services/retrieval/strategy/enhanced_pipeline.py"
+      - "新增: ai-service/app/services/retrieval/strategy/metadata_inference.py"
+      - "新增: ai-service/app/services/retrieval/strategy/strategy_router.py"
+      - "新增: ai-service/app/services/retrieval/strategy/mode_router.py"
+      - "新增: ai-service/app/services/retrieval/strategy/rollback_manager.py"
+      - "新增: ai-service/tests/test_retrieval_strategy_v2.py"
+    status: "✅ 任务完成"
+
+startup_guide:
+  - "Step 1: 读取本进度文档（了解当前位置与下一步）"
+  - "Step 2: 读取 spec_references 中定义的模块规范（了解业务与接口约束）"
+  - "Step 3: 通过 API 接口测试策略切换功能"
+  - "Step 4: 运行单元测试验证: pytest tests/test_retrieval_strategy_v2.py -v"
+---