merge: T8 pgvector 混合检索与缺口埋点（T-P2-03/04/06，TD-08）

packages/retrieval/README.md

@@ -1,14 +1,23 @@
-# `packages/retrieval/` — 混合检索（占位）
+# `packages/retrieval/` — 向量与混合检索
 
-架构 §8.3。进程内包，CPU + PostgreSQL。
+架构 §8.3。进程内包，CPU + PostgreSQL，无 HTTP 路由（被 agent / report 的
+service 调用）。
 
-## 待实现
+## 模块
 
-一条 SQL 完成硬过滤 + 向量召回（这正是选 pgvector 的收益），再做 RRF 融合：
+- `embedding.py`：`EmbeddingProvider` Protocol + `DeterministicEmbedder`
+  （测试/dev 的确定性假实现，sha256 种子 → 归一化 1024 维向量）+ 工厂。
+  真实 BGE-M3 ONNX 加载是后续任务（`kb_embedding_device` 已预留）。
+- `repo.py`：`kb_chunk` / `kb_gap` 数据访问。一条 CTE 完成硬过滤 +
+  向量召回（`<=>` 余弦）+ 关键词召回（`ts_rank`）。
+- `rerank.py`：RRF 融合重排的纯数学（单测逐项断言，不连库）。
+- `service.py`：编排 embed → 检索 →（无结果）gap 埋点。
+
+管线：
 
 ```
-硬过滤(SQL WHERE) → 规则命中 → 向量召回(pgvector) → 关键词召回(tsvector)
-→ 融合重排 → 输出知识片段 + 命中规则 ID + 图谱路径
+硬过滤(SQL WHERE) → 向量召回(pgvector) → 关键词召回(tsvector)
+→ RRF 融合 + 权重重排 → 输出 ScoredChunk；空结果落 kb_gap
 ```
 
 重排权重（可配置，需用黄金集调参）：
@@ -16,8 +25,10 @@
 
 ## 两条不能妥协的要求
 
-- **硬过滤零漏**（FR-KB-06）：禁忌项不得出现在候选中。这不是排序问题，
-  是过滤问题 —— 不能靠权重压低，必须在 SQL 里排除。
-- **规则命中要有明显加权**：专业规则的可信度高于语义相似度，这是自建知识库
-  的意义所在。
+- **硬过滤零漏**（FR-KB-06 / T-P2-04）：draft/禁忌实体必须在 SQL 的
+  WHERE 里排除，不能靠权重压低 —— 这是过滤问题，不是排序问题。
+  集成测试 `tests/integration/test_hybrid_retrieval.py` 断言"禁忌实体
+  即使向量最近也绝不出现"。
+- **规则命中要有明显加权**：专业规则的可信度高于语义相似度（`rule_boost`
+  接口已留，规则引擎联动属后续任务）。
 - p95 < 200ms（NFR-PERF-06）。

packages/retrieval/__init__.py

@@ -1 +1,34 @@
-"""混合检索：硬过滤 → 规则命中 → 向量召回 → BM25 → 重排 → 解释路径。占位。"""
+"""混合检索：硬过滤 → 向量召回 → 关键词召回 → RRF 重排 + 缺口埋点（架构 §8.3）.
+
+进程内包，被 agent / report 的 service 调用，没有 HTTP 路由。
+硬过滤零漏（FR-KB-06）与向量化可重跑（T-P2-03）的保证都在
+``repo.py`` 顶部注释里，改代码前先读。
+"""
+
+from __future__ import annotations
+
+from packages.retrieval.embedding import (
+    DETERMINISTIC_MODEL_NAME,
+    DeterministicEmbedder,
+    EmbeddingProvider,
+    build_embedding_provider,
+)
+from packages.retrieval.repo import ChunkInput, hybrid_search, record_gap, upsert_chunks
+from packages.retrieval.rerank import RecallRow, ScoredChunk
+from packages.retrieval.service import RetrievalService, SearchOutcome, UserContext
+
+__all__ = [
+    "DETERMINISTIC_MODEL_NAME",
+    "ChunkInput",
+    "DeterministicEmbedder",
+    "EmbeddingProvider",
+    "RecallRow",
+    "RetrievalService",
+    "ScoredChunk",
+    "SearchOutcome",
+    "UserContext",
+    "build_embedding_provider",
+    "hybrid_search",
+    "record_gap",
+    "upsert_chunks",
+]

packages/retrieval/embedding.py

@@ -0,0 +1,99 @@
+"""嵌入提供方：接口位 + 确定性假实现（架构 §8.3，TD-08）.
+
+当前只有 ``deterministic-v1``：以文本 sha256 为种子生成伪随机向量后 L2 归一化。
+它是**测试与 dev 的假实现** —— 真实 BGE-M3 ONNX 加载是后续任务，配置里的
+``kb_embedding_device`` 就是为它预留的。接口先定死，届时换模型只改工厂，
+检索管线（repo/service）不动。
+"""
+
+from __future__ import annotations
+
+import hashlib
+from typing import Protocol, runtime_checkable
+
+import numpy as np
+
+from packages.shared.config import Settings
+from packages.shared.errors import AppError, ErrorCode
+
+#: 唯一被承认的假实现模型名（与 Settings.kb_embedding_model 的默认值对应）。
+DETERMINISTIC_MODEL_NAME = "deterministic-v1"
+
+
+@runtime_checkable
+class EmbeddingProvider(Protocol):
+    """嵌入模型的最小接口。
+
+    ``model_name`` + ``version`` 会随向量逐行写进 kb_chunk（T-P2-03）：
+    没有版本号的向量不可重跑 —— 换模型即换版本，这是向量化可重跑的判据。
+    """
+
+    @property
+    def model_name(self) -> str: ...
+
+    @property
+    def version(self) -> str: ...
+
+    @property
+    def dimensions(self) -> int: ...
+
+    async def embed(self, texts: list[str]) -> list[list[float]]:
+        """批量嵌入。返回顺序与输入一一对应。"""
+        ...
+
+
+class DeterministicEmbedder:
+    """确定性假嵌入：同文本同向量，不同文本近似正交。
+
+    sha256(text) 做种子的标准高斯伪随机向量再 L2 归一化。1024 维空间里
+    随机向量两两夹角趋于 90°（|dot| ≈ 1/√1024 ≈ 0.03），这让"按余弦距离
+    排序"的管线在没有真模型时也能被测试驱动：同文本距离 0，异文本距离 ~1。
+    """
+
+    def __init__(self, *, dimensions: int = 1024, version: str = "v1") -> None:
+        if dimensions <= 0:
+            raise ValueError(f"dimensions 必须为正数, 得到 {dimensions}")
+        self._dimensions = dimensions
+        self._version = version
+
+    @property
+    def model_name(self) -> str:
+        return DETERMINISTIC_MODEL_NAME
+
+    @property
+    def version(self) -> str:
+        return self._version
+
+    @property
+    def dimensions(self) -> int:
+        return self._dimensions
+
+    def _vector(self, text: str) -> np.ndarray:
+        # 只取 sha256 前 16 字节做种子：PCG64 的种子空间足够，且避免跨平台
+        # int 大小差异。同一 numpy 版本内逐位确定（假实现不承诺跨大版本稳定，
+        # 真正的可重跑保证来自 kb_chunk.embedding_model/version 落库）。
+        seed = int.from_bytes(hashlib.sha256(text.encode("utf-8")).digest()[:16], "big")
+        vec = np.random.default_rng(seed).standard_normal(self._dimensions)
+        norm = float(np.linalg.norm(vec))
+        # 标准正态向量的范数恒 > 0，这里只防御浮点极端值，不做业务分支。
+        return vec / norm if norm > 0.0 else vec
+
+    async def embed(self, texts: list[str]) -> list[list[float]]:
+        return [self._vector(t).tolist() for t in texts]
+
+
+def build_embedding_provider(settings: Settings) -> EmbeddingProvider:
+    """按 ``settings.kb_embedding_model`` 选择实现。
+
+    当前只认 ``deterministic-v1``：出现其它名字说明配置与代码版本不匹配
+    （例如提前把 KB_EMBEDDING_MODEL 改成 BGE-M3 但 ONNX 加载任务还没合入），
+    直接拒绝比静默回退到假向量更能暴露问题。
+    """
+    if settings.kb_embedding_model == DETERMINISTIC_MODEL_NAME:
+        return DeterministicEmbedder(dimensions=settings.kb_vector_dimensions)
+    raise AppError(
+        ErrorCode.CONFIG_ERROR,
+        message=f"未知的嵌入模型: {settings.kb_embedding_model}",
+        hint="当前仅支持 deterministic-v1，真实 BGE-M3 属后续任务",
+        data={"model": settings.kb_embedding_model},
+    )

packages/retrieval/repo.py

@@ -0,0 +1,263 @@
+"""kb_chunk / kb_gap 数据访问。只有本模块写这两张表（架构 §8.3）.
+
+两条红线写在这里，改代码前先读：
+
+1. **硬过滤零漏**（T-P2-04 / FR-KB-06）：draft/禁忌实体必须在 SQL 的
+   WHERE 里排除，不能靠重排权重压低 —— 权重是排序问题，过滤是安全问题。
+2. **没有版本号的向量不可重跑**（T-P2-03）：写入必须携带
+   embedding_model/version；重跑按实体整组替换而不是追加。
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from collections.abc import Mapping, Sequence
+from dataclasses import dataclass
+from datetime import UTC, datetime
+from typing import Any
+
+from sqlalchemy import RowMapping, delete, select, text
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from packages.domain.ids import next_id
+from packages.retrieval.embedding import EmbeddingProvider
+from packages.retrieval.rerank import (
+    DEFAULT_RERANK_WEIGHTS,
+    RecallRow,
+    ScoredChunk,
+    rerank,
+)
+from packages.shared.errors import AppError, ErrorCode
+from packages.shared.models import KbChunk, KbGap
+
+# 硬过滤 + 双路召回 + RRF 的 CTE 骨架（架构 §8.3）。
+# 用原生 SQL 而不是 ORM 表达式：``<=>`` 余弦距离、ts_rank、jsonb ``?`` 操作符
+# 与窗口函数在 ORM 层拼写出来可读性更差，CTE 结构才是这份查询的可维护形态。
+# 参数一律显式 CAST，让 asyncpg 在 NULL/数组参数上也能推断类型。
+#
+# jsonb NULL 语义是两条容易被写错的规则，明确如下：
+# - occasion：没声明 occasions 的实体在按场景过滤时**不命中**（它没有声明
+#   适配该场景，宁可少给不能错给）；
+# - avoid_body_shapes：没声明禁忌的实体**必须保留**（NULL ? x 是 NULL，直接
+#   NOT 会把好实体一起丢掉），所以用 COALESCE 空数组兜底 —— 没写禁忌 ≠ 全都禁忌。
+_SEARCH_SQL = text(
+    """
+    WITH filtered AS (
+        SELECT c.chunk_id,
+               c.entity_id,
+               c.chunk_type,
+               c.text,
+               c.tsv,
+               e.updated_at,
+               c.embedding <=> CAST(:query_vec AS vector) AS vec_distance,
+               websearch_to_tsquery('simple', CAST(:query_text AS text)) AS kw_query,
+               ts_rank(c.tsv, websearch_to_tsquery('simple', CAST(:query_text AS text)))
+                   AS kw_tsrank
+        FROM kb_chunk c
+        JOIN kb_entity e ON e.entity_id = c.entity_id
+        WHERE e.status = 'published'
+          AND (CAST(:entity_types AS text[]) IS NULL
+               OR e.entity_type = ANY(CAST(:entity_types AS text[])))
+          AND (CAST(:occasion AS text) IS NULL
+               OR e.attrs -> 'occasions' ? CAST(:occasion AS text))
+          AND (CAST(:body_shape AS text) IS NULL
+               OR NOT (COALESCE(e.attrs -> 'avoid_body_shapes', '[]'::jsonb)
+                       ? CAST(:body_shape AS text)))
+    ),
+    vec_ranked AS (
+        SELECT chunk_id,
+               ROW_NUMBER() OVER (ORDER BY vec_distance, chunk_id) AS vec_rank,
+               vec_distance
+        FROM filtered
+        ORDER BY vec_distance, chunk_id
+        LIMIT CAST(:recall_k AS bigint)
+    ),
+    kw_ranked AS (
+        SELECT chunk_id,
+               ROW_NUMBER() OVER (ORDER BY kw_tsrank DESC, chunk_id) AS kw_rank,
+               kw_tsrank
+        FROM filtered
+        -- 过滤必须用 @@ 而不是 ts_rank > 0：ts_rank 对不匹配文本也返回
+        -- 极小的正 epsilon（~1e-20），靠分数阈值拦截会把无关 chunk 全放进来。
+        WHERE kw_tsrank > 0 AND tsv @@ kw_query
+        ORDER BY kw_tsrank DESC, chunk_id
+        LIMIT CAST(:recall_k AS bigint)
+    )
+    SELECT f.chunk_id,
+           f.entity_id,
+           f.chunk_type,
+           f.text,
+           f.updated_at,
+           v.vec_rank,
+           v.vec_distance,
+           k.kw_rank,
+           k.kw_tsrank
+    FROM vec_ranked v
+    FULL OUTER JOIN kw_ranked k ON k.chunk_id = v.chunk_id
+    JOIN filtered f ON f.chunk_id = COALESCE(v.chunk_id, k.chunk_id)
+    """
+)
+
+
+@dataclass(frozen=True, slots=True)
+class ChunkInput:
+    """待向量化写入的一个知识片段。tsv 是生成列，不在这里出现。"""
+
+    text: str
+    chunk_type: str = "text"
+
+
+def to_pgvector_literal(vec: Sequence[float]) -> str:
+    """把浮点向量序列化成 PG 的 vector 字面量 ``'[0.1,0.2,...]'``.
+
+    asyncpg 对扩展类型没有二进制编解码器，文本格式是官方支持的传参方式；
+    SQL 侧再显式 ``::vector`` 转型，PG 负责解析。
+    """
+    return "[" + ",".join(repr(float(x)) for x in vec) + "]"
+
+
+async def upsert_chunks(
+    session: AsyncSession,
+    entity_id: int,
+    chunks: Sequence[ChunkInput],
+    embedding: EmbeddingProvider,
+) -> list[KbChunk]:
+    """整实体替换式写入 kb_chunk（可重跑判据 T-P2-03）.
+
+    kb_chunk 没有业务自然键（text 改一个字就是新 chunk），所以 upsert 的
+    粒度是**实体**：先删该实体全部旧向量再插入新组 —— 重跑（无论换没换
+    模型）不留脏数据，embedding_model/version 逐行携带，换模型名即视为
+    不同版本。
+    """
+    for chunk in chunks:
+        if not chunk.text.strip():
+            raise AppError(
+                ErrorCode.VALIDATION_FAILED,
+                message="知识片段文本不能为空",
+                data={"entity_id": entity_id},
+            )
+
+    vectors = await embedding.embed([c.text for c in chunks])
+    if len(vectors) != len(chunks):
+        raise AppError(
+            ErrorCode.INTERNAL_ERROR,
+            message="嵌入实现返回的向量数与输入文本数不一致",
+            data={"model": embedding.model_name},
+        )
+    for vec in vectors:
+        if len(vec) != embedding.dimensions:
+            raise AppError(
+                ErrorCode.INTERNAL_ERROR,
+                message="嵌入实现返回的维度与声明不一致",
+                data={
+                    "model": embedding.model_name,
+                    "declared": embedding.dimensions,
+                    "actual": len(vec),
+                },
+            )
+
+    # 同一实体的旧向量在事务内先删后插：失败整体回滚，不会出现半新半旧。
+    await session.execute(delete(KbChunk).where(KbChunk.entity_id == entity_id))
+    rows = [
+        KbChunk(
+            chunk_id=next_id(),
+            entity_id=entity_id,
+            chunk_type=chunk.chunk_type,
+            text=chunk.text,
+            embedding=vec,
+            embedding_model=embedding.model_name,
+            embedding_version=embedding.version,
+        )
+        for chunk, vec in zip(chunks, vectors, strict=True)
+    ]
+    session.add_all(rows)
+    return rows
+
+
+async def hybrid_search(
+    session: AsyncSession,
+    query_vec: Sequence[float],
+    query_text: str,
+    *,
+    entity_types: Sequence[str] | None = None,
+    occasion: str | None = None,
+    body_shape: str | None = None,
+    rule_boost: Mapping[int, float] | None = None,
+    weights: Mapping[str, float] | None = None,
+    top_k: int = 50,
+) -> list[ScoredChunk]:
+    """硬过滤 → 向量召回 → 关键词召回 → RRF 融合重排（架构 §8.3）.
+
+    硬过滤（published / entity_types / occasion / avoid_body_shapes 禁忌）
+    全部发生在 ``filtered`` CTE 里，两路召回只从过滤后的集合里取 top N ——
+    这是"零漏"的结构保证，而非概率保证。空 ``entity_types`` 视为不限类型。
+    """
+    if top_k <= 0:
+        return []
+    result = await session.execute(
+        _SEARCH_SQL,
+        {
+            "query_vec": to_pgvector_literal(query_vec),
+            "query_text": query_text,
+            "entity_types": list(entity_types) if entity_types else None,
+            "occasion": occasion,
+            "body_shape": body_shape,
+            "recall_k": top_k,
+        },
+    )
+    rows: list[RecallRow] = [_to_recall_row(row) for row in result.mappings().all()]
+    return rerank(
+        rows,
+        weights=weights if weights is not None else DEFAULT_RERANK_WEIGHTS,
+        rule_boost=rule_boost,
+        now=datetime.now(UTC),
+        top_k=top_k,
+    )
+
+
+def _to_recall_row(row: RowMapping) -> RecallRow:
+    """把裸行收敛成 RecallRow：SQL 返回的 Any 在边界处一次性转干净。"""
+    return RecallRow(
+        chunk_id=int(row["chunk_id"]),
+        entity_id=int(row["entity_id"]),
+        chunk_type=str(row["chunk_type"]),
+        text=str(row["text"]),
+        updated_at=row["updated_at"],
+        vec_rank=int(row["vec_rank"]) if row["vec_rank"] is not None else None,
+        vec_distance=float(row["vec_distance"]) if row["vec_distance"] is not None else None,
+        kw_rank=int(row["kw_rank"]) if row["kw_rank"] is not None else None,
+        kw_tsrank=float(row["kw_tsrank"]) if row["kw_tsrank"] is not None else None,
+    )
+
+
+async def record_gap(
+    session: AsyncSession,
+    query_text: str,
+    filter_ctx: Mapping[str, Any],
+) -> None:
+    """无结果 query 落 kb_gap：同 query_text + filter_ctx 则 hit_count += 1（T-P2-06）.
+
+    kb_gap 只有普通索引没有唯一约束，``ON CONFLICT`` 用不了；用事务级
+    咨询锁把「查-改-插」串行化，保证同 key 并发时仍是一行、计数不丢。
+    锁在事务提交/回滚时自动释放，由调用方的 session_scope 兜底。
+    """
+    ctx = dict(filter_ctx)
+    payload = json.dumps(
+        {"query_text": query_text, "filter_ctx": ctx},
+        sort_keys=True,
+        ensure_ascii=False,
+    )
+    digest = hashlib.sha256(payload.encode("utf-8")).digest()[:8]
+    lock_key = int.from_bytes(digest, "big", signed=True)
+    await session.execute(text("SELECT pg_advisory_xact_lock(:lock_key)"), {"lock_key": lock_key})
+
+    found = await session.execute(
+        select(KbGap).where(KbGap.query_text == query_text, KbGap.filter_ctx == ctx)
+    )
+    gap = found.scalar_one_or_none()
+    if gap is None:
+        session.add(KbGap(gap_id=next_id(), query_text=query_text, filter_ctx=ctx, hit_count=1))
+    else:
+        gap.hit_count += 1
+        gap.last_seen_at = datetime.now(UTC)

packages/retrieval/rerank.py

@@ -0,0 +1,145 @@
+"""RRF 融合重排的纯数学部分（架构 §8.3 唯一口径）.
+
+刻意与 SQL 分离：重排公式必须能被单测逐项断言（T-P2-04 验收 3），
+混进 repo 就会把数学正确性测试拖进数据库。repo 只负责召回与名次，
+这里把名次变成最终分。
+"""
+
+from __future__ import annotations
+
+from collections.abc import Mapping, Sequence
+from dataclasses import dataclass
+from datetime import UTC, datetime
+
+#: RRF 常数。60 是业界默认值：名次差异对分数的影响被压成平滑倒数，
+#: 单路召回头部的偏差不至于淹没另一路。
+RRF_K = 60
+
+#: 新鲜度半衰期（天）：30 天前的内容得 0.5，60 天前 0.25……
+#: 知识库是慢更新的，半衰期必须显著大于内容评审周期，否则新版本刚
+#: 发布就会压制尚未重审的旧版本。
+FRESHNESS_HALF_LIFE_DAYS = 30.0
+
+#: 架构 §8.3 唯一口径：0.45 向量 + 0.2 关键词 + 0.25 规则 + 0.1 新鲜度。
+#: 与 Settings.retrieval_rerank_weights 的默认值保持一致（可被 env 覆盖）。
+DEFAULT_RERANK_WEIGHTS: dict[str, float] = {
+    "vector": 0.45,
+    "keyword": 0.2,
+    "rule": 0.25,
+    "freshness": 0.1,
+}
+
+
+@dataclass(frozen=True, slots=True)
+class RecallRow:
+    """混合检索的一行召回结果：两路召回各自的名次，未被召回的一路为 ``None``."""
+
+    chunk_id: int
+    entity_id: int
+    chunk_type: str
+    text: str
+    updated_at: datetime
+    vec_rank: int | None = None  # 向量路名次（1 起）
+    vec_distance: float | None = None  # 向量余弦距离，观测用
+    kw_rank: int | None = None  # 关键词路名次（1 起）
+    kw_tsrank: float | None = None  # ts_rank 原始分，观测用
+
+
+@dataclass(frozen=True, slots=True)
+class ScoredChunk:
+    """重排后的最终结果。各分量都归一到 [0, 1]，final_score 是加权和。"""
+
+    chunk_id: int
+    entity_id: int
+    chunk_type: str
+    text: str
+    vec_score: float
+    kw_score: float
+    rule_score: float
+    freshness_score: float
+    final_score: float
+
+
+def rrf_component(rank: int | None, k: int = RRF_K) -> float:
+    """单路 RRF 分量 ``1/(k+rank)``，再除以 ``1/(k+1)`` 归一到 (0, 1].
+
+    rank=1 → 1.0，rank=k+1 → 0.5。若直接用原始倒数，0.45×(1/61) ≈ 0.007
+    与规则项 0.25×1.0 完全不可比，架构 §8.3 的权重就失去"相对重要性"的
+    含义 —— 归一化只是线性缩放，不改变单路内的排序。
+    """
+    if rank is None or rank <= 0:
+        return 0.0
+    return (k + 1) / (k + rank)
+
+
+def rrf_score(vec_rank: int | None, kw_rank: int | None, k: int = RRF_K) -> float:
+    """经典 RRF 原始分：``1/(k+rank_vec) + 1/(k+rank_kw)``（观测/调试用）.
+
+    未被某路召回的 chunk 只累计被召回那一路的倒数。
+    """
+    total = 0.0
+    for rank in (vec_rank, kw_rank):
+        if rank is not None and rank > 0:
+            total += 1.0 / (k + rank)
+    return total
+
+
+def freshness_score(updated_at: datetime, now: datetime | None = None) -> float:
+    """新鲜度：``0.5 ** (age_days / 半衰期)``，越新越接近 1.
+
+    ``now`` 缺省取当前 UTC 时刻；显式传入是为了让单测能构造已知年龄。
+    未来时间戳（时钟漂移）按 0 天处理，得 1.0。
+    """
+    current = now if now is not None else datetime.now(UTC)
+    age_seconds = (current - updated_at).total_seconds()
+    age_days = max(age_seconds, 0.0) / 86400.0
+    # float ** float 在 typeshed 里是 Any（负底数分数幂会出 complex），显式收窄。
+    return float(0.5 ** (age_days / FRESHNESS_HALF_LIFE_DAYS))
+
+
+def rerank(
+    rows: Sequence[RecallRow],
+    *,
+    weights: Mapping[str, float] | None = None,
+    rule_boost: Mapping[int, float] | None = None,
+    now: datetime | None = None,
+    top_k: int | None = None,
+) -> list[ScoredChunk]:
+    """融合重排：``final = w_v×vec + w_k×kw + w_r×rule + w_f×fresh``（架构 §8.3）.
+
+    - ``vec``/``kw`` 是归一化的 RRF 分量（见 :func:`rrf_component`）；
+    - ``rule`` 直接取 ``rule_boost[chunk_id]``（调用方传入的规则权重，
+      规则引擎联动属后续任务，管线只认接口）；
+    - ``fresh`` 是半衰期衰减分。
+
+    排序稳定：final_score 降序，同分按 chunk_id 升序 —— 同一数据集重跑
+    结果可复现。``top_k`` 为 ``None`` 时不截断。
+    """
+    resolved_weights = weights if weights is not None else DEFAULT_RERANK_WEIGHTS
+    w_vec = float(resolved_weights.get("vector", 0.0))
+    w_kw = float(resolved_weights.get("keyword", 0.0))
+    w_rule = float(resolved_weights.get("rule", 0.0))
+    w_fresh = float(resolved_weights.get("freshness", 0.0))
+    boosts = rule_boost or {}
+
+    out = []
+    for row in rows:
+        vec = rrf_component(row.vec_rank)
+        kw = rrf_component(row.kw_rank)
+        rule = float(boosts.get(row.chunk_id, 0.0))
+        fresh = freshness_score(row.updated_at, now)
+        out.append(
+            ScoredChunk(
+                chunk_id=row.chunk_id,
+                entity_id=row.entity_id,
+                chunk_type=row.chunk_type,
+                text=row.text,
+                vec_score=vec,
+                kw_score=kw,
+                rule_score=rule,
+                freshness_score=fresh,
+                final_score=w_vec * vec + w_kw * kw + w_rule * rule + w_fresh * fresh,
+            )
+        )
+    out.sort(key=lambda c: (-c.final_score, c.chunk_id))
+    return out[:top_k] if top_k is not None else out

packages/retrieval/service.py

@@ -0,0 +1,114 @@
+"""检索编排：embed → 混合检索 →（无结果时）缺口埋点（架构 §8.3）.
+
+只做编排不做策略：模型、权重、top_k 全部来自 Settings；``rule_boost`` 的
+生产者（规则引擎联动）属后续任务，这里只认接口。检索是内部能力，
+没有 HTTP 路由 —— 被 agent / report 的 service 调用。
+"""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from dataclasses import dataclass, field
+from typing import Any
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from packages.retrieval import repo
+from packages.retrieval.embedding import EmbeddingProvider, build_embedding_provider
+from packages.retrieval.rerank import DEFAULT_RERANK_WEIGHTS, ScoredChunk
+from packages.shared.config import Settings
+from packages.shared.logging import get_logger
+
+logger = get_logger(__name__)
+
+
+@dataclass(frozen=True, slots=True)
+class UserContext:
+    """调用方给的检索上下文（用户画像/场景），驱动硬过滤与规则加权。
+
+    ``body_shape`` 是禁忌过滤的输入：命中 ``avoid_body_shapes`` 的实体
+    绝不出现（FR-KB-06），而不是排得靠后。
+    """
+
+    body_shape: str | None = None
+    occasion: str | None = None
+    entity_types: tuple[str, ...] = ()
+    rule_boost: Mapping[int, float] = field(default_factory=dict)
+
+    def filter_ctx(self) -> dict[str, Any]:
+        """落进 kb_gap 的过滤条件快照：缺口是在什么条件下出现的。"""
+        return {
+            "body_shape": self.body_shape,
+            "occasion": self.occasion,
+            "entity_types": list(self.entity_types),
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class SearchOutcome:
+    """一次检索的结果。``degraded`` 表示无结果并已记入 kb_gap（T-P2-06）。"""
+
+    results: list[ScoredChunk]
+    degraded: bool
+
+
+class RetrievalService:
+    """混合检索的编排入口（嵌入 + 检索 + 缺口埋点）."""
+
+    def __init__(
+        self,
+        session: AsyncSession,
+        *,
+        embedding: EmbeddingProvider,
+        weights: Mapping[str, float] | None = None,
+        top_k: int = 50,
+    ) -> None:
+        self._session = session
+        self._embedding = embedding
+        self._weights: dict[str, float] = (
+            dict(weights) if weights is not None else dict(DEFAULT_RERANK_WEIGHTS)
+        )
+        self._top_k = top_k
+
+    @classmethod
+    def from_settings(cls, session: AsyncSession, settings: Settings) -> RetrievalService:
+        """从 Settings 装配（模型、权重、top_k 一处配置，换模型不发版）."""
+        return cls(
+            session,
+            embedding=build_embedding_provider(settings),
+            weights=settings.retrieval_rerank_weights,
+            top_k=settings.retrieval_top_k,
+        )
+
+    async def search(
+        self, query_text: str, *, user_ctx: UserContext | None = None
+    ) -> SearchOutcome:
+        ctx = user_ctx or UserContext()
+        # embed 的输入就是 query 原文（DeterministicEmbedder/未来的 BGE-M3 都
+        # 不需要拼提示词），保持「同 query 同向量」，缓存与复跑才有意义。
+        query_vec = (await self._embedding.embed([query_text]))[0]
+        results = await repo.hybrid_search(
+            self._session,
+            query_vec,
+            query_text,
+            entity_types=ctx.entity_types or None,
+            occasion=ctx.occasion,
+            body_shape=ctx.body_shape,
+            rule_boost=ctx.rule_boost,
+            weights=self._weights,
+            top_k=self._top_k,
+        )
+        if results:
+            return SearchOutcome(results=results, degraded=False)
+
+        # 无结果也是信号：落 kb_gap 驱动内容生产（T-P2-06）。埋点失败不该
+        # 让调用方拿到异常 —— 检索空结果本身是合法结局，这里吞掉埋点异常
+        # 只记日志。query 原文进表不进日志（日志里只留长度与过滤条件）。
+        try:
+            await repo.record_gap(self._session, query_text, ctx.filter_ctx())
+            await self._session.commit()
+        except Exception:
+            await self._session.rollback()
+            logger.error("retrieval_gap_record_failed", query_len=len(query_text))
+        logger.info("retrieval_gap", query_len=len(query_text), **ctx.filter_ctx())
+        return SearchOutcome(results=[], degraded=True)

tests/integration/test_hybrid_retrieval.py

@@ -0,0 +1,297 @@
+"""混合检索集成：真实 PostgreSQL + pgvector，嵌入用 DeterministicEmbedder（不联网）.
+
+证明三件单测证明不了的事（任务文档 §4）：
+1. 硬过滤零漏 —— 禁忌/draft 实体即使向量最近也绝不出现（T-P2-04 红线）；
+2. 关键词路径在 tsv 'simple' 分词下真实命中；
+3. 无结果 query 落 kb_gap 且 hit_count 递增（T-P2-06）。
+外加：向量化重跑整组替换且逐行带版本（T-P2-03）、kb_asset partial 索引可用。
+
+Run: MEIDA_RUN_INTEGRATION=1 .venv/bin/pytest -m integration -q --no-cov
+"""
+
+from __future__ import annotations
+
+import os
+import uuid
+from collections.abc import AsyncIterator
+from typing import Any
+
+import pytest
+from sqlalchemy import delete, select, text
+
+from packages.domain.ids import next_id
+from packages.retrieval.embedding import DeterministicEmbedder
+from packages.retrieval.repo import ChunkInput, hybrid_search, upsert_chunks
+from packages.retrieval.service import RetrievalService, UserContext
+from packages.shared.db import dispose_engine, session_scope
+from packages.shared.models import KbAsset, KbChunk, KbEntity, KbGap
+
+
+def _infra_available() -> bool:
+    return bool(os.getenv("MEIDA_RUN_INTEGRATION"))
+
+
+pytestmark = [
+    pytest.mark.integration,
+    pytest.mark.skipif(
+        not _infra_available(),
+        reason="set MEIDA_RUN_INTEGRATION=1 with PostgreSQL and Redis running",
+    ),
+]
+
+
+def _unique(name: str) -> str:
+    return f"it-t8-{name}-{uuid.uuid4().hex[:10]}"
+
+
+@pytest.fixture(autouse=True)
+async def _hermetic_engine() -> AsyncIterator[None]:
+    """每个用例前后重建全局 engine：asyncpg 连接绑定创建它的事件循环，
+    pytest-asyncio 给每个用例一个新 loop，复用池里旧连接会报
+    "attached to a different loop"。app 系用例由 lifespan 的 dispose_engine
+    兜底；直接用 session_scope 的本文件必须自己管。"""
+    await dispose_engine()
+    yield
+    await dispose_engine()
+
+
+@pytest.fixture
+async def clean_kb() -> AsyncIterator[dict[str, list[Any]]]:
+    """登记本用例造的脏数据并在结束时清掉：集成库长活，不能留假实体."""
+    registry: dict[str, list[Any]] = {"entities": [], "gaps": []}
+    yield registry
+    async with session_scope() as session:
+        if registry["entities"]:
+            ids = registry["entities"]
+            await session.execute(delete(KbChunk).where(KbChunk.entity_id.in_(ids)))
+            await session.execute(delete(KbAsset).where(KbAsset.entity_id.in_(ids)))
+            await session.execute(delete(KbEntity).where(KbEntity.entity_id.in_(ids)))
+        for query_text in registry["gaps"]:
+            await session.execute(delete(KbGap).where(KbGap.query_text == query_text))
+
+
+async def _add_entity(
+    registry: dict[str, list[Any]],
+    *,
+    entity_type: str = "item",
+    status: str = "published",
+    attrs: dict[str, Any] | None = None,
+) -> int:
+    entity_id = next_id()
+    registry["entities"].append(entity_id)
+    async with session_scope() as session:
+        session.add(
+            KbEntity(
+                entity_id=entity_id,
+                entity_type=entity_type,
+                name=_unique(entity_type),
+                attrs=attrs or {},
+                status=status,
+            )
+        )
+    return entity_id
+
+
+class RenamedEmbedder:
+    """换名复用确定性向量：验证 model/version 逐行落库与整组替换（T-P2-03）."""
+
+    def __init__(self, model_name: str, version: str) -> None:
+        self._inner = DeterministicEmbedder(dimensions=1024)
+        self.model_name = model_name
+        self.version = version
+        self.dimensions = 1024
+
+    async def embed(self, texts: list[str]) -> list[list[float]]:
+        return await self._inner.embed(texts)
+
+
+# --- 1. 硬过滤零漏（T-P2-04 红线）--------------------------------------------
+async def test_hard_filter_never_leaks_taboo_or_draft(clean_kb: dict[str, list[Any]]) -> None:
+    embedder = DeterministicEmbedder(dimensions=1024)
+    query_text = "pear shape body dress advice"
+    # 防御性登记：若回归导致空结果被记进 kb_gap，用例结束也能清干净。
+    clean_kb["gaps"].append(query_text)
+
+    taboo_id = await _add_entity(
+        clean_kb, attrs={"occasions": ["wedding"], "avoid_body_shapes": ["apple"]}
+    )
+    draft_id = await _add_entity(clean_kb, status="draft")
+    good_id = await _add_entity(clean_kb, attrs={"occasions": ["wedding"]})
+
+    async with session_scope() as session:
+        # 禁忌与 draft 的 chunk 和 query 完全同文：不过滤的话它们必然霸占
+        # 向量第一名（距离 0）—— 这样断言"零出现"才有说服力。
+        await upsert_chunks(session, taboo_id, [ChunkInput(text=query_text)], embedder)
+        await upsert_chunks(session, draft_id, [ChunkInput(text=query_text)], embedder)
+        await upsert_chunks(
+            session,
+            good_id,
+            [ChunkInput(text="soft knit sweater paired with a high waist skirt")],
+            embedder,
+        )
+
+    async with session_scope() as session:
+        svc = RetrievalService(session, embedding=embedder)
+        outcome = await svc.search(
+            query_text,
+            user_ctx=UserContext(body_shape="apple", occasion="wedding", entity_types=("item",)),
+        )
+
+    assert outcome.degraded is False
+    assert outcome.results, "正常 published 实体应被召回"
+    assert {r.entity_id for r in outcome.results} == {good_id}
+
+    # 对照组：去掉禁忌/场景过滤后禁忌实体确实在库里且可被召回 ——
+    # 证明上面拦截它的是硬过滤，而不是数据没写进去。draft 永远不出现。
+    query_vec = (await embedder.embed([query_text]))[0]
+    async with session_scope() as session:
+        control = await hybrid_search(session, query_vec, query_text)
+    control_entities = {r.entity_id for r in control}
+    assert taboo_id in control_entities
+    assert draft_id not in control_entities
+
+
+# --- 2. 关键词路径（tsv 'simple' 分词）----------------------------------------
+async def test_keyword_recall_ranks_matching_chunk_first(
+    clean_kb: dict[str, list[Any]],
+) -> None:
+    embedder = DeterministicEmbedder(dimensions=1024)
+    entity_id = await _add_entity(clean_kb)
+    matching = "breathable linen shirt keeps you cool in summer"
+    unrelated = "heavy wool coat for freezing winter nights"
+
+    async with session_scope() as session:
+        await upsert_chunks(
+            session,
+            entity_id,
+            [ChunkInput(text=matching), ChunkInput(text=unrelated)],
+            embedder,
+        )
+
+    # 查询文本与两个 chunk 都不同文 → 向量路分不出高下（近似正交），
+    # 排序由关键词路（0.2 权重、归一后差距 1.0）决定。
+    async with session_scope() as session:
+        results = await hybrid_search(
+            session, (await embedder.embed(["linen shirt summer"]))[0], "linen shirt summer"
+        )
+
+    assert [r.text for r in results] == [matching, unrelated]
+    assert results[0].kw_score == pytest.approx(1.0)
+    assert results[1].kw_score == 0.0
+
+
+# --- 3. 向量化可重跑（T-P2-03）------------------------------------------------
+async def test_upsert_chunks_is_versioned_and_rerunnable(
+    clean_kb: dict[str, list[Any]],
+) -> None:
+    entity_id = await _add_entity(clean_kb)
+
+    async with session_scope() as session:
+        first = await upsert_chunks(
+            session,
+            entity_id,
+            [ChunkInput(text="rule one"), ChunkInput(text="rule two")],
+            RenamedEmbedder("bge-m3", "v1"),
+        )
+        assert {c.embedding_model for c in first} == {"bge-m3"}
+        assert {c.embedding_version for c in first} == {"v1"}
+
+    # 换模型名（= 换版本）重跑：整组替换，不残留旧版本向量
+    async with session_scope() as session:
+        await upsert_chunks(
+            session,
+            entity_id,
+            [ChunkInput(text="rule one"), ChunkInput(text="rule two"), ChunkInput(text="rule new")],
+            RenamedEmbedder("bge-m4", "v2"),
+        )
+
+    async with session_scope() as session:
+        rows = (
+            (await session.execute(select(KbChunk).where(KbChunk.entity_id == entity_id)))
+            .scalars()
+            .all()
+        )
+    assert len(rows) == 3
+    assert {r.embedding_model for r in rows} == {"bge-m4"}
+    assert {r.embedding_version for r in rows} == {"v2"}
+
+
+# --- 4. 无结果 query 落 kb_gap（T-P2-06）--------------------------------------
+async def test_no_result_query_lands_in_kb_gap_with_counter(
+    clean_kb: dict[str, list[Any]],
+) -> None:
+    query_text = _unique("missing")
+    clean_kb["gaps"].append(query_text)
+    svc_embedder = DeterministicEmbedder(dimensions=1024)
+
+    async with session_scope() as session:
+        svc = RetrievalService(session, embedding=svc_embedder)
+        first = await svc.search(query_text, user_ctx=UserContext(entity_types=("nonexistent",)))
+        second = await svc.search(query_text, user_ctx=UserContext(entity_types=("nonexistent",)))
+
+    assert first.degraded and first.results == []
+    assert second.degraded and second.results == []
+
+    async with session_scope() as session:
+        gaps = (
+            (await session.execute(select(KbGap).where(KbGap.query_text == query_text)))
+            .scalars()
+            .all()
+        )
+    assert len(gaps) == 1  # 同 query + 同 filter_ctx 只有一行
+    assert gaps[0].hit_count == 2
+    assert gaps[0].filter_ctx["entity_types"] == ["nonexistent"]
+
+
+# --- 5. kb_asset 按 entity 查询 + partial 索引可用 -----------------------------
+async def test_tryon_ready_assets_query_uses_partial_index(
+    clean_kb: dict[str, list[Any]],
+) -> None:
+    entity_id = await _add_entity(clean_kb)
+    async with session_scope() as session:
+        session.add_all(
+            [
+                KbAsset(
+                    asset_id=next_id(),
+                    entity_id=entity_id,
+                    asset_type="front",
+                    cos_key=_unique("front"),
+                    tryon_ready=True,
+                    review_status="approved",
+                ),
+                KbAsset(
+                    asset_id=next_id(),
+                    entity_id=entity_id,
+                    asset_type="side",
+                    cos_key=_unique("side"),
+                    tryon_ready=False,
+                    review_status="pending",
+                ),
+            ]
+        )
+
+    async with session_scope() as session:
+        ready = (
+            (
+                await session.execute(
+                    select(KbAsset).where(
+                        KbAsset.entity_id == entity_id, KbAsset.tryon_ready.is_(True)
+                    )
+                )
+            )
+            .scalars()
+            .all()
+        )
+        assert len(ready) == 1
+        assert ready[0].tryon_ready is True
+
+        # 关闭顺序扫描逼 planner 走索引，证明 partial 索引真的可用
+        # （SET LOCAL 只影响当前事务，随 session_scope 提交即失效）。
+        await session.execute(text("SET LOCAL enable_seqscan = off"))
+        plan_lines = (
+            (await session.execute(text("EXPLAIN SELECT asset_id FROM kb_asset WHERE tryon_ready")))
+            .scalars()
+            .all()
+        )
+    plan_text = "\n".join(str(line) for line in plan_lines)
+    assert "ix_kb_asset_tryon_ready" in plan_text

tests/unit/test_embedding.py

@@ -0,0 +1,99 @@
+"""DeterministicEmbedder 与工厂：纯数学，不连库不联网（T-P2-03 接口位）."""
+
+from __future__ import annotations
+
+import math
+
+import pytest
+
+from packages.retrieval.embedding import (
+    DETERMINISTIC_MODEL_NAME,
+    DeterministicEmbedder,
+    build_embedding_provider,
+)
+from packages.shared.config import get_settings
+from packages.shared.errors import AppError, ErrorCode
+
+
+def _dot(a: list[float], b: list[float]) -> float:
+    return sum(x * y for x, y in zip(a, b, strict=True))
+
+
+async def test_same_text_embeds_to_same_vector() -> None:
+    embedder = DeterministicEmbedder(dimensions=1024)
+    first, second = await embedder.embed(["白衬衫怎么搭配", "白衬衫怎么搭配"])
+    assert first == second
+
+
+async def test_dimensions_match_declaration() -> None:
+    vec = (await DeterministicEmbedder(dimensions=1024).embed(["hello"]))[0]
+    assert len(vec) == 1024
+    small = (await DeterministicEmbedder(dimensions=8).embed(["hello"]))[0]
+    assert len(small) == 8
+
+
+async def test_vectors_are_l2_normalized() -> None:
+    for text in ("hello", "中文知识片段", ""):
+        vec = (await DeterministicEmbedder(dimensions=1024).embed([text]))[0]
+        assert math.isclose(_dot(vec, vec), 1.0, rel_tol=1e-6, abs_tol=1e-9)
+
+
+async def test_different_texts_are_near_orthogonal() -> None:
+    """不同文本近似正交（|dot| ≪ 1）：向量召回管线在无真模型时仍可测."""
+
+    left, right = await DeterministicEmbedder(dimensions=1024).embed(
+        ["interview outfit advice", "winter skincare routine"]
+    )
+    assert left != right
+    assert abs(_dot(left, right)) < 0.1
+
+
+async def test_batch_preserves_order_and_length() -> None:
+    texts = ["alpha", "beta", "gamma"]
+    vectors = await DeterministicEmbedder(dimensions=16).embed(texts)
+    assert len(vectors) == len(texts)
+    singles = [(await DeterministicEmbedder(dimensions=16).embed([t]))[0] for t in texts]
+    assert vectors == singles
+
+
+def test_metadata_is_stable() -> None:
+    embedder = DeterministicEmbedder(dimensions=8)
+    assert embedder.model_name == DETERMINISTIC_MODEL_NAME
+    assert embedder.version
+    assert embedder.dimensions == 8
+
+
+def test_invalid_dimensions_rejected() -> None:
+    with pytest.raises(ValueError, match="dimensions"):
+        DeterministicEmbedder(dimensions=0)
+
+
+def test_factory_builds_deterministic_embedder(monkeypatch: pytest.MonkeyPatch) -> None:
+    # 带 alias 的 Settings 字段按字段名构造会被静默忽略 —— 统一走
+    # setenv + cache_clear（上批经验，见任务文档 §2）。
+    monkeypatch.setenv("KB_EMBEDDING_MODEL", "deterministic-v1")
+    monkeypatch.setenv("KB_VECTOR_DIMENSIONS", "1024")
+    get_settings.cache_clear()
+    try:
+        settings = get_settings()
+        provider = build_embedding_provider(settings)
+        assert isinstance(provider, DeterministicEmbedder)
+        assert provider.model_name == "deterministic-v1"
+        assert provider.dimensions == 1024
+    finally:
+        monkeypatch.undo()
+        get_settings.cache_clear()
+
+
+def test_factory_rejects_unknown_model(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("KB_EMBEDDING_MODEL", "bge-m3-onnx")
+    get_settings.cache_clear()
+    try:
+        settings = get_settings()
+        assert settings.kb_embedding_model == "bge-m3-onnx"
+        with pytest.raises(AppError) as excinfo:
+            build_embedding_provider(settings)
+        assert excinfo.value.code == ErrorCode.CONFIG_ERROR
+    finally:
+        monkeypatch.undo()
+        get_settings.cache_clear()

tests/unit/test_rerank.py

@@ -0,0 +1,158 @@
+"""RRF 融合 + 权重 + 新鲜度的数学正确性（T-P2-04 验收 3）：构造已知名次断言最终序."""
+
+from __future__ import annotations
+
+import itertools
+from datetime import UTC, datetime, timedelta
+
+import pytest
+
+from packages.retrieval.rerank import (
+    DEFAULT_RERANK_WEIGHTS,
+    FRESHNESS_HALF_LIFE_DAYS,
+    RRF_K,
+    RecallRow,
+    freshness_score,
+    rerank,
+    rrf_component,
+    rrf_score,
+)
+
+NOW = datetime(2026, 8, 14, 12, 0, 0, tzinfo=UTC)
+
+
+def _row(
+    chunk_id: int,
+    *,
+    vec_rank: int | None = None,
+    kw_rank: int | None = None,
+    updated_at: datetime = NOW,
+    entity_id: int = 1,
+) -> RecallRow:
+    return RecallRow(
+        chunk_id=chunk_id,
+        entity_id=entity_id,
+        chunk_type="text",
+        text=f"chunk-{chunk_id}",
+        updated_at=updated_at,
+        vec_rank=vec_rank,
+        kw_rank=kw_rank,
+    )
+
+
+# --- 单路 RRF 分量 ------------------------------------------------------------
+def test_rrf_component_is_max_normalized() -> None:
+    assert rrf_component(1) == pytest.approx(1.0)  # (60+1)/(60+1)
+    assert rrf_component(2) == pytest.approx(61 / 62)
+    assert rrf_component(RRF_K + 2) == pytest.approx(0.5)  # 61/122：名次再排后一半
+    assert rrf_component(None) == 0.0
+    assert rrf_component(0) == 0.0
+    assert rrf_component(-3) == 0.0
+
+
+def test_rrf_component_monotonic_non_increasing() -> None:
+    values = [rrf_component(r) for r in range(1, 20)]
+    assert all(a >= b for a, b in itertools.pairwise(values))
+
+
+def test_rrf_score_is_sum_of_raw_reciprocals() -> None:
+    assert rrf_score(1, 2) == pytest.approx(1 / 61 + 1 / 62)
+    assert rrf_score(None, 1) == pytest.approx(1 / 61)
+    assert rrf_score(1, None) == pytest.approx(1 / 61)
+    assert rrf_score(None, None) == 0.0
+
+
+# --- 新鲜度 ------------------------------------------------------------------
+def test_freshness_half_life_decay() -> None:
+    assert freshness_score(NOW, now=NOW) == pytest.approx(1.0)
+    old = NOW - timedelta(days=FRESHNESS_HALF_LIFE_DAYS)
+    assert freshness_score(old, now=NOW) == pytest.approx(0.5)
+    older = NOW - timedelta(days=2 * FRESHNESS_HALF_LIFE_DAYS)
+    assert freshness_score(older, now=NOW) == pytest.approx(0.25)
+    # 时钟漂移产生未来时间戳 → 按 0 天处理，不放大分数
+    future = NOW + timedelta(hours=1)
+    assert freshness_score(future, now=NOW) == pytest.approx(1.0)
+
+
+# --- 加权公式 ----------------------------------------------------------------
+def test_final_score_follows_architecture_formula() -> None:
+    """final = 0.45×vec + 0.2×kw + 0.25×rule + 0.1×fresh，逐项核对."""
+
+    row = _row(1, vec_rank=1, kw_rank=3)
+    boost = {1: 0.5}
+    scored = rerank([row], rule_boost=boost, now=NOW)
+    expected = 0.45 * 1.0 + 0.2 * (61 / 63) + 0.25 * 0.5 + 0.1 * 1.0
+    assert scored[0].final_score == pytest.approx(expected)
+    assert scored[0].vec_score == pytest.approx(1.0)
+    assert scored[0].kw_score == pytest.approx(61 / 63)
+    assert scored[0].rule_score == pytest.approx(0.5)
+    assert scored[0].freshness_score == pytest.approx(1.0)
+
+
+def test_default_weights_match_architecture_section_8_3() -> None:
+    assert DEFAULT_RERANK_WEIGHTS == {
+        "vector": 0.45,
+        "keyword": 0.2,
+        "rule": 0.25,
+        "freshness": 0.1,
+    }
+
+
+def test_dual_recall_beats_single_recall() -> None:
+    """两路都命中的 chunk 应压过只有向量路命中、名次更高的 chunk."""
+
+    both = _row(1, vec_rank=5, kw_rank=2)
+    vec_only = _row(2, vec_rank=1)
+    scored = rerank([both, vec_only], now=NOW)
+    by_id = {c.chunk_id: c for c in scored}
+    assert by_id[1].final_score > by_id[2].final_score
+
+
+def test_rule_boost_lifts_lower_recall_chunk() -> None:
+    """规则权重 0.25×boost 可以把召回名次落后的 chunk 抬到最前（自建知识库的意义）."""
+
+    rank5 = _row(1, vec_rank=5)
+    rank1 = _row(2, vec_rank=1)
+    scored = rerank([rank5, rank1], rule_boost={1: 1.0}, now=NOW)
+    assert scored[0].chunk_id == 1
+    assert scored[0].rule_score == pytest.approx(1.0)
+
+
+def test_keyword_only_chunk_gets_zero_vector_score() -> None:
+    scored = rerank([_row(7, kw_rank=1)], now=NOW)
+    assert scored[0].vec_score == 0.0
+    assert scored[0].kw_score == pytest.approx(1.0)
+
+
+def test_freshness_breaks_ties_toward_newer_chunk() -> None:
+    """名次相同（现实中少见，但公式上可能）时，更新鲜者胜."""
+
+    fresh = _row(1, vec_rank=2, updated_at=NOW)
+    stale = _row(2, vec_rank=2, updated_at=NOW - timedelta(days=90))
+    scored = rerank([fresh, stale], now=NOW)
+    assert scored[0].chunk_id == 1
+    delta = scored[0].final_score - scored[1].final_score
+    assert delta == pytest.approx(0.1 * (1.0 - 0.5**3))
+
+
+def test_zero_weight_dimension_ignores_that_signal() -> None:
+    old = _row(1, vec_rank=1, updated_at=NOW - timedelta(days=365))
+    new = _row(2, vec_rank=2, updated_at=NOW)
+    scored = rerank([old, new], weights={"vector": 1.0}, now=NOW)
+    assert scored[0].chunk_id == 1  # 新鲜度权重为 0，旧的反而名次靠前
+
+
+def test_stable_ordering_and_top_k_truncation() -> None:
+    rows = [_row(i, vec_rank=i) for i in (30, 10, 20, 40)]
+    scored = rerank(rows, now=NOW, top_k=2)
+    assert [c.chunk_id for c in scored] == [10, 20]
+
+
+def test_equal_final_scores_break_tie_by_chunk_id() -> None:
+    rows = [_row(90, kw_rank=1), _row(10, kw_rank=1)]
+    scored = rerank(rows, now=NOW)
+    assert [c.chunk_id for c in scored] == [10, 90]
+
+
+def test_empty_input_returns_empty() -> None:
+    assert rerank([], now=NOW) == []

tests/unit/test_retrieval_service.py

@@ -0,0 +1,179 @@
+"""RetrievalService 编排逻辑：repo 层打桩，不连库（编排正确性与 DB 无关）."""
+
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+
+from packages.retrieval import repo
+from packages.retrieval import service as service_mod
+from packages.retrieval.rerank import ScoredChunk
+from packages.retrieval.service import RetrievalService, UserContext
+
+
+class FakeEmbedder:
+    """满足 EmbeddingProvider Protocol 的最小桩：记录调用，返回固定向量."""
+
+    model_name = "fake"
+    version = "v1"
+    dimensions = 4
+
+    def __init__(self) -> None:
+        self.calls: list[list[str]] = []
+
+    async def embed(self, texts: list[str]) -> list[list[float]]:
+        self.calls.append(list(texts))
+        return [[0.25, 0.25, 0.25, 0.25] for _ in texts]
+
+
+class FakeSession:
+    def __init__(self) -> None:
+        self.commits = 0
+        self.rollbacks = 0
+
+    async def commit(self) -> None:
+        self.commits += 1
+
+    async def rollback(self) -> None:
+        self.rollbacks += 1
+
+
+def _scored(chunk_id: int = 1) -> ScoredChunk:
+    return ScoredChunk(
+        chunk_id=chunk_id,
+        entity_id=100,
+        chunk_type="text",
+        text="chunk",
+        vec_score=1.0,
+        kw_score=0.0,
+        rule_score=0.0,
+        freshness_score=1.0,
+        final_score=0.55,
+    )
+
+
+@pytest.fixture
+def stubbed_repo(monkeypatch: pytest.MonkeyPatch) -> dict[str, Any]:
+    """把 repo 的两个 IO 函数换成记录桩，service 逻辑照常执行."""
+
+    calls: dict[str, Any] = {"hybrid_search": None, "gaps": [], "results": []}
+
+    async def fake_hybrid_search(
+        session: Any, query_vec: Any, query_text: str, **kwargs: Any
+    ) -> list[ScoredChunk]:
+        calls["hybrid_search"] = {"query_vec": query_vec, "query_text": query_text, **kwargs}
+        return list(calls["results"])
+
+    async def fake_record_gap(session: Any, query_text: str, filter_ctx: Any) -> None:
+        calls["gaps"].append((query_text, filter_ctx))
+
+    monkeypatch.setattr(repo, "hybrid_search", fake_hybrid_search)
+    monkeypatch.setattr(repo, "record_gap", fake_record_gap)
+    return calls
+
+
+async def test_search_with_hits_returns_results_without_gap(stubbed_repo: dict[str, Any]) -> None:
+    stubbed_repo["results"] = [_scored(1), _scored(2)]
+    embedder = FakeEmbedder()
+    session = FakeSession()
+    svc = RetrievalService(session, embedding=embedder)
+
+    outcome = await svc.search("面试穿什么", user_ctx=UserContext(body_shape="apple"))
+
+    assert outcome.degraded is False
+    assert [c.chunk_id for c in outcome.results] == [1, 2]
+    assert stubbed_repo["gaps"] == []
+    assert session.commits == 0  # 有结果的路径不写库，事务留给调用方
+    assert embedder.calls == [["面试穿什么"]]  # embed 的输入就是 query 原文
+
+
+async def test_search_forwards_user_context_to_hybrid(stubbed_repo: dict[str, Any]) -> None:
+    stubbed_repo["results"] = [_scored()]
+    svc = RetrievalService(FakeSession(), embedding=FakeEmbedder())
+
+    await svc.search(
+        "婚礼穿什么",
+        user_ctx=UserContext(
+            body_shape="pear",
+            occasion="wedding",
+            entity_types=("item", "style"),
+            rule_boost={42: 0.8},
+        ),
+    )
+
+    seen = stubbed_repo["hybrid_search"]
+    assert seen is not None
+    assert seen["query_text"] == "婚礼穿什么"
+    assert seen["body_shape"] == "pear"
+    assert seen["occasion"] == "wedding"
+    assert seen["entity_types"] == ("item", "style")
+    assert seen["rule_boost"] == {42: 0.8}
+
+
+async def test_empty_entity_types_is_forwarded_as_none(stubbed_repo: dict[str, Any]) -> None:
+    stubbed_repo["results"] = []
+    svc = RetrievalService(FakeSession(), embedding=FakeEmbedder())
+
+    await svc.search("随便看看", user_ctx=UserContext())
+
+    seen = stubbed_repo["hybrid_search"]
+    assert seen is not None
+    assert seen["entity_types"] is None  # 空元组 = 不限类型，不是匹配空集合
+
+
+async def test_no_results_records_gap_and_degrades(stubbed_repo: dict[str, Any]) -> None:
+    stubbed_repo["results"] = []
+    session = FakeSession()
+    svc = RetrievalService(session, embedding=FakeEmbedder())
+
+    outcome = await svc.search(
+        "不存在的知识点", user_ctx=UserContext(body_shape="apple", entity_types=("item",))
+    )
+
+    assert outcome.degraded is True
+    assert outcome.results == []
+    assert session.commits == 1  # gap 埋点立即提交：检索本身只读，别拖住调用方事务
+    assert stubbed_repo["gaps"] == [
+        (
+            "不存在的知识点",
+            {"body_shape": "apple", "occasion": None, "entity_types": ["item"]},
+        )
+    ]
+
+
+async def test_gap_failure_degrades_but_does_not_raise(
+    stubbed_repo: dict[str, Any], monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """埋点失败不该把"检索空结果"变成异常 —— 空结果本身是合法结局."""
+
+    async def broken_gap(session: Any, query_text: str, filter_ctx: Any) -> None:
+        raise RuntimeError("db down")
+
+    monkeypatch.setattr(repo, "record_gap", broken_gap)
+    stubbed_repo["results"] = []
+    session = FakeSession()
+    svc = RetrievalService(session, embedding=FakeEmbedder())
+
+    outcome = await svc.search("anything")
+
+    assert outcome.degraded is True
+    assert session.rollbacks == 1
+
+
+async def test_default_context_when_user_ctx_omitted(stubbed_repo: dict[str, Any]) -> None:
+    stubbed_repo["results"] = []
+    svc = RetrievalService(FakeSession(), embedding=FakeEmbedder())
+
+    outcome = await svc.search("no ctx")
+
+    assert outcome.degraded is True
+    assert stubbed_repo["gaps"] == [
+        ("no ctx", {"body_shape": None, "occasion": None, "entity_types": []})
+    ]
+
+
+def test_service_module_reexports_for_dependency_injection() -> None:
+    """service 与 repo 共享同一模块对象：monkeypatch 打桩路径就是生产调用路径."""
+
+    assert service_mod.repo is repo