feat(knowledge): 语义检索打通——切片、混合打分、只搜已发布

之前知识只能"存"，不能"找"。这一版把链路补上：切片 → 向量 → 打分 → 带来源返回。

- chunk 表 + 迁移 0006：正文按标题归段切、代码块整块保留、超长按句切、
  相邻片重叠 80 字；RLS 与 GRANT 与其它租户表一致
- 向量提供方做成协议：内置 hash（字符二元组，零依赖可复现，作为基线）
  与 OpenAI 兼容端点两种，换实现不动业务代码
- 检索是混合打分：向量 0.6 + pg_trgm 0.4。专有编号、单号这类必须原样命中的词
  靠词法兜，换个说法表达的同一件事靠向量兜
- 三条纪律：只搜 published（过滤写在 SQL 里）、低于阈值宁可不返回、
  每条命中都带来源标题与 item_id
- 发布时自动建索引，改完正文用 POST /knowledge/{id}/reindex 重建
- Web 知识页加检索框与命中卡片；新增 e2e/retrieval.mjs 6 项

为什么不是 pgvector：本机 Docker 起不来、fnsvr-01 没有该扩展且明确不动。
先用 real[] + SQL 余弦把链路跑通，将来换列类型与索引即可，接口不变。

后端 79 passed，e2e smoke/knowledge/richtext/retrieval 全绿。

2.Design/API.md

@@ -377,7 +377,45 @@ Query：`doc_type`（默认 `postmortem`）。来源可以是任务或工单；
 **工单验收通过时，后端会自动调用这条逻辑**（`generated_by=rule`，`doc_type=postmortem`，
 正文开头固定一句"这份草稿由系统根据来源自动整理，请先核对再发布"），所以用户不需要记得点。
 
-### 5.7 知识的错误码
+### 5.7 检索 `GET /knowledge/search`
+
+Query：
+| 参数 | 说明 |
+|---|---|
+| `q` | 必填，1–200 字 |
+| `limit` | 1–30，默认 10 |
+| `mode` | `hybrid`（默认）/ `vector` / `keyword`，后两个用于对比效果 |
+
+返回 `{query, mode, items[]}`，每项：
+
+| 字段 | 说明 |
+|---|---|
+| `item_id` / `title` | 来源知识，前端"打开原文"用 |
+| `doc_type` / `heading` | 类型与所在小节 |
+| `idx` / `content` | 第几片、切片正文 |
+| `score` | 加权总分，低于 `SEARCH_MIN_SCORE`（0.06）的结果**不返回** |
+| `vector_score` / `lexical_score` | 向量分与 trigram 词法分，便于排查"为什么没搜到" |
+
+三条约定：
+
+1. **只检索 `published`**。草稿是还没被人认可的话，进检索等于把猜测当结论引用出去——过滤写在 SQL 里，调用方改不了。
+2. **结果必须带来源**，能点回原文。这是 D3（知识必须有来源）的延伸：不给"出处不存在"的答案。
+3. **宁可返回空，也不凑数**。检索失败比误导便宜。
+
+发布（`5.5` 的 `publish`）时会自动重建索引，用户发布完立刻能搜到；
+但改完正文不会自动重建，需要调用下面这条。
+
+### `POST /knowledge/{item_id}/reindex` — 重建切片
+
+返回 `{item_id, chunk_count, searchable}`。`searchable=false` 表示内容已建好切片但还不是 `published`，
+搜不到——这不是错误。
+
+实现：切片存 `chunk` 表（`content` + `embedding real[]`），相似度用 SQL 函数 `app_cosine()` 计算。
+本机没有可用的 pgvector 实例（Docker 起不来，fnsvr-01 明确不动），所以先用 `real[]` 把链路跑通；
+将来换 pgvector 只改列类型与索引，接口不变。换 embedding 模型（`EMBEDDING_PROVIDER`）后
+**必须整体重建**，否则新旧向量混在一张表里，分数没法比较。
+
+### 5.8 知识的错误码
 
 | 码 | HTTP | 含义 |
 |---|---|---|

2.Design/ROADMAP-M0-M1.md

@@ -272,21 +272,32 @@
 | 3.5 | 手动沉淀：任务/工单详情里"沉淀成知识" | 0.5 | [x] 有 e2e |
 | 3.6 | 手记：没有来源的知识（唯一允许 `source_kind=manual` 的入口） | 0.5 | [x] 手记页 + 列表 |
 | 3.7 | 检索：标题/正文/摘要子串 + 类型/状态筛选 + 游标分页 | 1.0 | [x] trigram 索引在 0002 已建 |
-| 3.8 | 切片与向量索引（`chunk` 表、pgvector HNSW） | 2.0 | [ ] 见下方说明 |
+| 3.8 | 切片与向量索引（`chunk` 表、pgvector HNSW） | 2.0 | [x] **用 `real[]` + SQL 余弦先跑通**（见下方说明） |
-| 3.9 | 安卓知识列表 | 1.0 | [ ] 与工单一起延后到编译链就绪 |
+| 3.8.1 | 切片：按标题归段、代码块整块保留、超长按句切、相邻片重叠 80 字 | 0.5 | [x] `app/domain/chunking.py` |
-| 3.10 | 文档与 e2e | 1.0 | [x] `web/e2e/knowledge.mjs` 7 项断言绿 |
+| 3.8.2 | 向量：`hash`（字符二元组）与 `openai` 两种提供方，换实现不动业务 | 0.5 | [x] `app/services/embedding.py` |
+| 3.8.3 | 混合检索：向量 0.6 + trigram 0.4，低于阈值不返回，只认 `published` | 1.0 | [x] 5 个用例 + `e2e/retrieval.mjs` 6 项 |
+| 3.8.4 | 索引生命周期：发布自动建、改正文需 `reindex`、换模型整体重建 | 0.5 | [x] `POST /knowledge/{id}/reindex` |
+| 3.9 | 安卓知识列表 | 1.0 | [x] 只读列表 + 详情动作（编辑正文走安卓编辑器） |
+| 3.10 | 文档与 e2e | 1.0 | [x] `knowledge.mjs` 7 项 + `retrieval.mjs` 6 项 |
 
 ### M3 完成定义（DOD）
 - [x] 一件事办完 → 自动出现一份草稿 → 核对 → 发布，全程不用手动复制粘贴
 - [x] 每条知识都能追到来源（除手记），点开能看到来源事项
 - [x] 非管理员发布被拒；跨租户看不到别人的知识
-- [ ] 向量切片与语义检索（需要独立 pgvector 实例真正跑起来）
+- [x] 切片与语义检索跑通（`real[]` + SQL 余弦；pgvector 仍待独立实例）
 - [ ] 真实用两周：新同事靠知识库自己解决过一次问题（需要真实团队）
 
-**关于 3.8 的说明**：本机 Docker 起不来、fnsvr-01 上也没有 pgvector 扩展，
+**关于 3.8 的说明（2026-10-29 已落地，但与计划不同）**：计划里写的是 pgvector + HNSW，
-所以 `deploy/docker-compose.vector.yml` 与 `VECTOR_DATABASE_URL` 只有配置，**没有跑过**。
+实际本机 Docker 起不来、fnsvr-01 上也没有 pgvector 扩展且用户明确要求别动那台机器，
-按"写完不算完成"的规矩，它在真跑起来之前保持未勾。切片表设计要点已写进
+所以先用 `chunk` 表（`embedding real[]`）+ SQL 函数 `app_cosine()` 把整条链路跑通：
-`2.Design/PRODUCT-PLAN.md` 的 D8，落地时不要改 `EMBEDDING_DIM`（改了要重算全量）。
+切片 → 向量（内置 hash 提供方）→ 混合打分（向量 0.6 + trigram 0.4）→ 阈值过滤 → 带来源返回。
+`deploy/docker-compose.vector.yml` 与 `VECTOR_DATABASE_URL` 保留，等向量实例真起来之后再切过去，
+换 pgvector 只改列类型与索引，应用层接口不变。
+
+两条必须记住的约束：
+1. **内置 hash 向量不是语义模型**，只是"零依赖、可复现"的基线；上线前换 `EMBEDDING_PROVIDER=openai`
+   并整体重建（`reindex_workspace`），否则新旧向量混表、分数不可比。
+2. **检索只认 `published`**，草稿永远搜不到——这条写在 SQL 的 WHERE 里，不靠调用方自觉。
 
 ---
 

README.md

@@ -29,8 +29,8 @@
 | 知识：工单验收自动生成草稿、核对后发布、来源可追、手记 | ✅ 后端 + Web + 安卓已通，e2e 7 项全绿 |
 | 界面：两端统一设计系统（纸白 / 墨绿 / 陶土橙，语义色、字号阶、组件库） | ✅ 已重做，见 [`2.Design/ROADMAP-M0-M1.md`](2.Design/ROADMAP-M0-M1.md) 第 4 节 |
 | 富文本：备注 / 工单说明 / 知识正文支持图文混排与附件 | ✅ Web 用 Vditor 编辑、安卓用 Markwon 预览编辑（e2e 15 项全绿）；接口见 [`2.Design/API.md`](2.Design/API.md) 第 6 节 |
-| 后端测试 | ✅ 74 passed，ruff 干净 |
+| 知识切片与语义检索（混合打分、只搜已发布、带来源） | ✅ 已跑通：`real[]` + SQL 余弦（pgvector 待独立实例），`pytest` 5 个用例 + `e2e/retrieval.mjs` 6 项 |
-| 知识切片与语义检索（pgvector） | ⬜ 配置已就绪（`deploy/docker-compose.vector.yml`），本机 Docker 起不来，未跑过 |
+| 后端测试 | ✅ 79 passed，ruff 干净 |
 | LLM 能力（结构化录入、问答带引用、人物画像） | ⬜ M4 |
 
 ## 快速开始

backend/alembic/versions/0006_chunks.py

@@ -0,0 +1,102 @@
+"""知识切片与检索（M3）
+
+Revision ID: 0006
+Revises: 0005
+Create Date: 2026-10-29
+
+建 `chunk` 表：知识正文按段切开，每片一行，带向量数组。
+
+两个刻意的取舍：
+
+1. **向量存 `real[]`，不依赖 pgvector**。本机拿不到带 pgvector 的实例（Docker 起不来，
+   fnsvr-01 明确不动），但检索链路不能一直停在纸面上。`real[]` + 一条 SQL 余弦函数
+   在小数据量下够用；将来换上 pgvector 只需改列类型、加 HNSW 索引，应用层接口不变。
+2. **检索范围只能是 `published`**。草稿是"还没被人认可的话"，进检索等于让 AI 把
+   猜测当结论引用出去——这条过滤写在查询里，不靠调用方自觉。
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+from alembic import op
+
+revision: str = "0006"
+down_revision: str = "0005"
+branch_labels: Sequence[str] | None = None
+depends_on: Sequence[str] | None = None
+
+# 余弦相似度。写成 SQL 函数而不是在 Python 里算，是为了让过滤与排序留在数据库里做——
+# 把整表向量捞回应用算相似度，数据一多就会拖垮 API。
+COSINE_SQL = """
+CREATE OR REPLACE FUNCTION app_cosine(a real[], b real[]) RETURNS double precision
+LANGUAGE plpgsql IMMUTABLE PARALLEL SAFE AS $fn$
+DECLARE
+    n int;
+    dot double precision := 0;
+    na double precision := 0;
+    nb double precision := 0;
+BEGIN
+    n := LEAST(array_length(a, 1), array_length(b, 1));
+    IF n IS NULL THEN
+        RETURN NULL;
+    END IF;
+    FOR i IN 1..n LOOP
+        dot := dot + a[i] * b[i];
+        na := na + a[i] * a[i];
+        nb := nb + b[i] * b[i];
+    END LOOP;
+    IF na = 0 OR nb = 0 THEN
+        RETURN 0;
+    END IF;
+    RETURN dot / sqrt(na * nb);
+END
+$fn$;
+"""
+
+CREATE_SQL = """
+CREATE TABLE chunk (
+    id uuid PRIMARY KEY DEFAULT gen_random_uuid(),
+    workspace_id uuid NOT NULL,
+    item_id uuid NOT NULL REFERENCES item(id) ON DELETE CASCADE,
+    idx integer NOT NULL,
+    heading varchar(200),
+    content text NOT NULL,
+    char_count integer NOT NULL DEFAULT 0,
+    embedding real[],
+    embed_model varchar(64),
+    version integer NOT NULL DEFAULT 1,
+    created_at timestamptz NOT NULL DEFAULT now(),
+    updated_at timestamptz NOT NULL DEFAULT now(),
+    CONSTRAINT uq_chunk_item_idx UNIQUE (item_id, idx)
+)
+"""
+
+# asyncpg 的执行器不接受多语句，一条 `execute` 只能带一句 SQL（0001 踩过同样的坑）
+INDEX_STATEMENTS = (
+    "CREATE INDEX ix_chunk_ws_item ON chunk (workspace_id, item_id)",
+    "COMMENT ON TABLE chunk IS '知识正文切片，检索的最小单位'",
+)
+
+POLICY_SQL = (
+    "ALTER TABLE chunk ENABLE ROW LEVEL SECURITY",
+    "CREATE POLICY chunk_isolation ON chunk "
+    "USING (workspace_id = app_ws_id()) WITH CHECK (workspace_id = app_ws_id())",
+    "GRANT SELECT, INSERT, UPDATE, DELETE ON chunk TO dadao_app",
+)
+
+
+def upgrade() -> None:
+    op.execute(COSINE_SQL)
+    op.execute(CREATE_SQL)
+    for statement in INDEX_STATEMENTS:
+        op.execute(statement)
+    for statement in POLICY_SQL:
+        op.execute(statement)
+
+
+def downgrade() -> None:
+    op.execute("DROP POLICY IF EXISTS chunk_isolation ON chunk")
+    op.execute("DROP TABLE IF EXISTS chunk")
+    # 函数被别的东西引用时先别删；本轮只有切片在用，留一条注释说明为什么还留着。
+    # op.execute("DROP FUNCTION IF EXISTS app_cosine(real[], real[])")

backend/app/api/knowledge.py

@@ -33,6 +33,7 @@ from ..models_task import Activity, Item
 from ..schemas import clamp_limit, decode_cursor, encode_cursor, ok
 from ..serializers import item_out, tags_of
 from ..services.knowledge import create_draft_from_source, find_doc_by_source
+from ..services.knowledge_index import reindex_document, search_knowledge
 from .attachments import count_for_items, list_for_item
 
 router = APIRouter(prefix="/workspaces/{ws_id}", tags=["知识库"])
@@ -203,6 +204,20 @@ async def list_knowledge(
     return ok({"items": items, "next_cursor": next_cursor, "has_more": has_more})
 
 
+# 注意路由顺序：`/knowledge/search` 必须排在 `/knowledge/{item_id}` 之前，
+# 否则 "search" 会被当成 uuid 解析，返回 422。
+@router.get("/knowledge/search", summary="检索已发布知识")
+async def search(
+    ctx: MemberDep,
+    session: SessionDep,
+    q: Annotated[str, Query(min_length=1, max_length=200, description="检索词")],
+    limit: Annotated[int, Query(ge=1, le=30)] = 10,
+    mode: Annotated[str, Query(pattern="^(hybrid|vector|keyword)$")] = "hybrid",
+) -> dict:
+    hits = await search_knowledge(session, ctx.ws_id, q, limit=limit, mode=mode)
+    return ok({"query": q, "mode": mode, "items": hits, "has_more": False})
+
+
 @router.get("/knowledge/{item_id}", summary="知识详情")
 async def get_knowledge(item_id: uuid.UUID, ctx: MemberDep, session: SessionDep) -> dict:
     item, doc = await _load_doc(session, ctx.ws_id, item_id)
@@ -396,9 +411,28 @@ async def change_status(
         payload={"to": to_status},
     )
     await session.flush()
+    if body.action == "publish":
+        # 发布后立刻建索引：检索只认 published，不在这里建就要等定时任务，
+        # 而"刚发布却搜不到"是最容易被当成 bug 的体验。
+        await reindex_document(session, ws_id, item.id)
     return ok(await _knowledge_out(session, ws_id, item, doc))
 
 
+@router.post("/knowledge/{item_id}/reindex", summary="重建这份知识的检索切片")
+async def reindex(item_id: uuid.UUID, ctx: MemberDep, session: SessionDep) -> dict:
+    # 改完正文不重建，检索命中的还是旧句子——留一个能手动触发的入口
+    _item, doc = await _load_doc(session, ctx.ws_id, item_id)
+    count = await reindex_document(session, ctx.ws_id, item_id)
+    await session.flush()
+    return ok(
+        {
+            "item_id": item_id,
+            "chunk_count": count,
+            "searchable": doc.status_hint == "published",
+        }
+    )
+
+
 # ── 从任务/工单沉淀 ─────────────────────────────────────────────────────
 @router.post("/items/{item_id}/knowledge", summary="把这件事沉淀成知识草稿")
 async def promote_to_knowledge(

backend/app/config.py

@@ -89,6 +89,9 @@ class Settings(BaseSettings):
     search_vector_weight: float = 0.6
     search_lexical_weight: float = 0.4
 
+    # 低于这个分数就不返回：宁可"没找到"，也不要塞一条不相干的让人误信。
+    search_min_score: float = 0.06
+
     # ── 知识库向量库（独立实例，M3 起用）───────────────────
     # 刻意与业务库分开放：fnsvr-01 上的 PG 没有 pgvector，也不该为了装扩展去动它；
     # 向量检索吃内存，和任务库挤在同一实例会互相拖累。

backend/app/domain/chunking.py

@@ -0,0 +1,127 @@
+"""把 Markdown 正文切成适合检索的小片段（M3）。
+
+切片的目的是"喂给检索"，不是"给人读"，所以三条规则优先于其它考虑：
+
+1. **不切坏结构**：代码块整块保留，表格不拆行，标题跟着它下面的段落走
+   （单独一句"结论"离开上下文就没意义）。
+2. **片内自洽**：每片尽量包含一个完整意思，超长才按句号切。
+3. **相邻片有重叠**：跨段的话被切断时，前后片都能读到完整句子，命中后引用才不乱。
+
+不引入分词器：中文另起 `zhparser` 要装扩展，而片段足够短时字符级检索已经够用。
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+
+HEADING_RE = re.compile(r"^(#{1,6})\s+(.+?)\s*$")
+FENCE_RE = re.compile(r"^\s*```")
+SENTENCE_RE = re.compile(r"[^。！？!?；;\n]+[。！？!?；;]?")
+
+
+@dataclass(frozen=True)
+class ChunkDraft:
+    heading: str | None
+    content: str
+
+
+def _split_long(text: str, max_chars: int) -> list[str]:
+    """把超长段落按句子切开；单句仍然超长就硬切。"""
+    pieces: list[str] = []
+    buffer = ""
+    for sentence in SENTENCE_RE.findall(text):
+        if not sentence.strip():
+            continue
+        if len(buffer) + len(sentence) <= max_chars:
+            buffer += sentence
+            continue
+        if buffer.strip():
+            pieces.append(buffer.strip())
+        if len(sentence) <= max_chars:
+            buffer = sentence
+        else:
+            for start in range(0, len(sentence), max_chars):
+                pieces.append(sentence[start : start + max_chars].strip())
+            buffer = ""
+    if buffer.strip():
+        pieces.append(buffer.strip())
+    return [p for p in pieces if p]
+
+
+def split_markdown(
+    text: str,
+    *,
+    max_chars: int = 600,
+    overlap: int = 80,
+) -> list[ChunkDraft]:
+    """返回按顺序排列的切片。
+
+    `overlap` 是相邻片之间共享的尾部字数；不想要重叠就传 0。
+    """
+    if not text or not text.strip():
+        return []
+
+    normalized = text.replace("\r\n", "\n").replace("\r", "\n")
+    blocks: list[tuple[str | None, str]] = []
+    heading: str | None = None
+    paragraph: list[str] = []
+    in_fence = False
+
+    def flush() -> None:
+        if paragraph:
+            body = "\n".join(paragraph).strip()
+            if body:
+                blocks.append((heading, body))
+            paragraph.clear()
+
+    for line in normalized.split("\n"):
+        if FENCE_RE.match(line):
+            in_fence = not in_fence
+            paragraph.append(line)
+            continue
+        if not in_fence:
+            match = HEADING_RE.match(line)
+            if match:
+                flush()
+                heading = match.group(2).strip()
+                continue
+            if not line.strip():
+                flush()
+                continue
+        paragraph.append(line)
+    flush()
+
+    chunks: list[ChunkDraft] = []
+    buffer = ""
+    buffer_heading: str | None = None
+
+    def flush_buffer() -> None:
+        nonlocal buffer, buffer_heading
+        if buffer.strip():
+            chunks.append(ChunkDraft(heading=buffer_heading, content=buffer.strip()))
+        buffer = ""
+
+    for block_heading, body in blocks:
+        parts = _split_long(body, max_chars) if len(body) > max_chars else [body]
+        for part in parts:
+            candidate = f"{buffer}\n\n{part}".strip() if buffer else part
+            if len(candidate) <= max_chars:
+                if not buffer:
+                    buffer_heading = block_heading
+                buffer = candidate
+                continue
+            flush_buffer()
+            # 新片开头带上上一片的尾巴，避免"半句话"被切断后两边都读不通
+            tail = ""
+            if overlap > 0 and chunks:
+                tail = chunks[-1].content[-overlap:]
+                if "\n" in tail:
+                    tail = tail.split("\n", 1)[-1]
+            buffer_heading = block_heading
+            buffer = f"{tail}\n\n{part}".strip() if tail else part
+            if len(buffer) > max_chars:
+                # 带尾巴后仍然超长：放弃尾巴，保证不超限
+                buffer = part
+    flush_buffer()
+    return chunks

backend/app/models_knowledge.py

@@ -19,15 +19,18 @@ from datetime import datetime
 from sqlalchemy import (
     CheckConstraint,
     DateTime,
+    Float,
     ForeignKey,
     Index,
     String,
     Text,
+    UniqueConstraint,
 )
+from sqlalchemy.dialects.postgresql import ARRAY
 from sqlalchemy.dialects.postgresql import UUID as PgUUID
 from sqlalchemy.orm import Mapped, mapped_column
 
-from .models import Base, Timestamps
+from .models import Base, Timestamps, new_uuid
 
 # 知识状态机（与任务/工单共用 item.status 一列，取值集合不同）
 KNOWLEDGE_STATUSES: tuple[str, ...] = ("draft", "in_review", "published", "archived")
@@ -76,4 +79,34 @@ class KnowledgeDoc(Base, Timestamps):
     chunk_count: Mapped[int] = mapped_column(nullable=False, default=0)
 
 
-KNOWLEDGE_TENANT_TABLES: tuple[str, ...] = ("knowledge_doc",)
+class Chunk(Base):
+    """知识正文的切片。
+
+    为什么切片与文档同库：检索要按工作区过滤，跨库做权限校验就得把两边数据捞到应用里，
+    既不安全也不快。向量本身是普通数组，换 pgvector 时只换列类型与索引。
+    """
+
+    __tablename__ = "chunk"
+    __table_args__ = (
+        UniqueConstraint("item_id", "idx", name="uq_chunk_item_idx"),
+        Index("ix_chunk_ws_item", "workspace_id", "item_id"),
+        {"comment": "知识正文切片，检索的最小单位"},
+    )
+
+    id: Mapped[uuid.UUID] = mapped_column(PgUUID(as_uuid=True), primary_key=True, default=new_uuid)
+    workspace_id: Mapped[uuid.UUID] = mapped_column(PgUUID(as_uuid=True), nullable=False)
+    item_id: Mapped[uuid.UUID] = mapped_column(
+        PgUUID(as_uuid=True), ForeignKey("item.id", ondelete="CASCADE"), nullable=False
+    )
+    # 片内顺序，用于把命中片段按原顺序还原上下文
+    idx: Mapped[int] = mapped_column(nullable=False)
+    heading: Mapped[str | None] = mapped_column(String(200))
+    content: Mapped[str] = mapped_column(Text, nullable=False)
+    char_count: Mapped[int] = mapped_column(nullable=False, default=0)
+    # 向量。hash 提供方是 256 维，OpenAI 兼容端点通常是 1024/1536 维，
+    # 所以不固定维度——换模型时老切片要么重算，要么维度不匹配（检索会拒绝并提示重建）。
+    embedding: Mapped[list[float] | None] = mapped_column(ARRAY(Float))
+    embed_model: Mapped[str | None] = mapped_column(String(64))
+
+
+KNOWLEDGE_TENANT_TABLES: tuple[str, ...] = ("knowledge_doc", "chunk")

backend/app/services/embedding.py

@@ -0,0 +1,131 @@
+"""文本向量化（M3）。
+
+为什么要抽象一层：今天用内置的哈希向量，明天换成 OpenAI 兼容的 `/embeddings`，
+后天可能换成自建模型。检索代码只依赖 `embed() -> list[float]`，换实现不动业务。
+
+内置 `HashEmbedding` 的诚实说明：它不是"语义模型"，是**字符 n-gram 的哈希词袋**。
+它能把同词、近义写法（"发票"/"票据"这种）之外的东西区分开，抓不到真正的语义相似。
+之所以先上它：本机没有可用的向量库与模型服务，检索链路不能一直不落地；
+而且它零依赖、可复现（同样的文本永远同样的向量），适合做测试基线。
+真上线前应当把 `EMBEDDING_PROVIDER` 换成 `openai`，并把已发布知识整体重建一次。
+"""
+
+from __future__ import annotations
+
+import hashlib
+import math
+import re
+from collections.abc import Sequence
+from typing import Protocol
+
+import httpx
+
+from ..config import Settings, get_settings
+
+CJK_RE = re.compile(r"[\u4e00-\u9fff]")
+WORD_RE = re.compile(r"[a-z0-9]+")
+
+
+class EmbeddingProvider(Protocol):
+    """向量提供方。`name` 会写进切片，换模型后能看出哪些是旧向量。"""
+
+    name: str
+
+    async def embed(self, texts: Sequence[str]) -> list[list[float]]: ...
+
+
+def _tokens(text: str) -> list[str]:
+    """中文按字二元组、英文按词。
+
+    中文不用分词器：单字太碎（"发""票"都对不上"发票"），二元组能拿到大多数
+    有意义的搭配，又不需要维护词典。
+    """
+    lowered = text.lower()
+    tokens = WORD_RE.findall(lowered)
+    chars = [c for c in lowered if CJK_RE.match(c)]
+    tokens.extend(chars[i] + chars[i + 1] for i in range(len(chars) - 1))
+    tokens.extend(chars)
+    return tokens
+
+
+class HashEmbedding:
+    """字符 n-gram 的带符号哈希词袋，L2 归一化。"""
+
+    def __init__(self, dim: int = 256) -> None:
+        self.dim = dim
+        self.name = f"hash-{dim}"
+
+    async def embed(self, texts: Sequence[str]) -> list[list[float]]:
+        return [self._one(text) for text in texts]
+
+    def _one(self, text: str) -> list[float]:
+        vector = [0.0] * self.dim
+        counts: dict[str, int] = {}
+        for token in _tokens(text):
+            counts[token] = counts.get(token, 0) + 1
+        for token, count in counts.items():
+            digest = hashlib.blake2b(token.encode("utf-8"), digest_size=8).digest()
+            raw = int.from_bytes(digest, "big")
+            index = raw % self.dim
+            # 用最高位决定正负：无符号哈希会让不相干的词在向量里越加越像
+            sign = 1.0 if (raw >> 63) & 1 else -1.0
+            vector[index] += sign * (1.0 + math.log(count))
+        norm = math.sqrt(sum(v * v for v in vector))
+        if norm == 0:
+            return vector
+        return [v / norm for v in vector]
+
+
+class OpenAIEmbedding:
+    """任何 OpenAI 兼容的 `/embeddings` 端点（BYOK）。"""
+
+    def __init__(self, *, base_url: str, api_key: str, model: str) -> None:
+        self.base_url = base_url.rstrip("/")
+        self.api_key = api_key
+        self.model = model
+        self.name = f"openai:{model}"
+
+    async def embed(self, texts: Sequence[str]) -> list[list[float]]:
+        if not texts:
+            return []
+        async with httpx.AsyncClient(timeout=30.0) as client:
+            response = await client.post(
+                f"{self.base_url}/embeddings",
+                headers={"Authorization": f"Bearer {self.api_key}"},
+                json={"model": self.model, "input": list(texts)},
+            )
+            response.raise_for_status()
+            payload = response.json()
+        # 按 index 排序：并发端点的返回顺序没有保证
+        rows = sorted(payload["data"], key=lambda row: row["index"])
+        return [row["embedding"] for row in rows]
+
+
+_provider: EmbeddingProvider | None = None
+
+
+def get_provider(settings: Settings | None = None) -> EmbeddingProvider:
+    """进程内单例：哈希实现无状态，OpenAI 实现也不需要每次新建连接池。"""
+    global _provider
+    if _provider is not None:
+        return _provider
+    settings = settings or get_settings()
+    if settings.embedding_provider == "openai":
+        if not settings.embedding_api_key:
+            raise RuntimeError(
+                "EMBEDDING_PROVIDER=openai 但没有配置 EMBEDDING_API_KEY；"
+                "要么补上密钥，要么改回 hash"
+            )
+        _provider = OpenAIEmbedding(
+            base_url=settings.embedding_base_url,
+            api_key=settings.embedding_api_key,
+            model=settings.embedding_model,
+        )
+    else:
+        _provider = HashEmbedding()
+    return _provider
+
+
+def reset_provider_for_tests() -> None:
+    global _provider
+    _provider = None

backend/app/services/knowledge_index.py

@@ -0,0 +1,208 @@
+"""知识的切片索引与检索（M3）。
+
+两条硬规则：
+
+1. **只有 `published` 的知识能被检索到**。草稿是"还没被人认可的话"，
+   让它进检索等于把猜测当结论引用出去。过滤写在 SQL 里，不靠调用方自觉。
+2. **检索必须带来源**。每个命中都返回 `item_id` 与标题，用户能点回去看原文；
+   这和 D3（知识必须有来源）是同一条原则的延伸——不给"出处不存在"的答案。
+
+评分是向量与词法的加权和：
+- 向量负责"说的是同一件事但换了词"；
+- 词法（pg_trgm 的 trigram 相似度）负责"专有名词、单号、人名"这类必须原样命中的词，
+  而哈希向量对这类词并不敏感。
+两者都不足时宁可返回空，也不硬凑一条不相关的结果——检索失败比误导便宜。
+"""
+
+from __future__ import annotations
+
+import uuid
+from typing import Any
+
+from sqlalchemy import delete, select, text
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..config import get_settings
+from ..domain.chunking import split_markdown
+from ..errors import AppError, ErrorCode
+from ..models_knowledge import Chunk, KnowledgeDoc
+from ..models_task import Item
+from .embedding import get_provider
+
+SEARCH_SQL = text(
+    """
+    SELECT c.item_id,
+           i.title,
+           k.doc_type,
+           c.idx,
+           c.heading,
+           c.content,
+           coalesce(app_cosine(c.embedding, CAST(:qvec AS real[])), 0) AS vec_score,
+           coalesce(similarity(c.content, :q), 0) AS lex_score
+    FROM chunk c
+    JOIN item i ON i.id = c.item_id AND i.workspace_id = c.workspace_id
+    JOIN knowledge_doc k ON k.item_id = c.item_id AND k.workspace_id = c.workspace_id
+    WHERE c.workspace_id = :ws
+      AND k.status_hint = 'published'
+    ORDER BY (
+        coalesce(app_cosine(c.embedding, CAST(:qvec AS real[])), 0) * :vw
+        + coalesce(similarity(c.content, :q), 0) * :lw
+    ) DESC, c.item_id, c.idx
+    LIMIT :limit
+    """
+)
+
+
+def _vector_param(values: list[float]) -> list[float]:
+    """把向量交给驱动去编码。
+
+    一开始我拼成 `'{0.1,0.2}'` 字符串再 `CAST(... AS real[])`，asyncpg 会因为 CAST
+    已经告诉它"这是数组"而拒绝字符串（`a sized iterable container expected`）。
+    直接传 list 最省事；非有限值替换成 0，免得 NaN 进 SQL 变成错误输入。
+    """
+    return [value if value == value and abs(value) != float("inf") else 0.0 for value in values]
+
+
+async def reindex_document(
+    session: AsyncSession,
+    ws_id: uuid.UUID,
+    item_id: uuid.UUID,
+) -> int:
+    """重建一份知识的切片，返回切片数。
+
+    全量重建而不是增量：知识正文改动通常牵动上下几段，增量对齐的复杂度
+    换不来多少收益（一份知识撑死几十片）。
+    """
+    item = (
+        await session.execute(select(Item).where(Item.workspace_id == ws_id, Item.id == item_id))
+    ).scalar_one_or_none()
+    if item is None:
+        raise AppError(ErrorCode.ITEM_NOT_FOUND, "这件事不存在")
+
+    doc = (
+        await session.execute(
+            select(KnowledgeDoc).where(
+                KnowledgeDoc.workspace_id == ws_id,
+                KnowledgeDoc.item_id == item_id,
+            )
+        )
+    ).scalar_one_or_none()
+    if doc is None:
+        raise AppError(ErrorCode.KNOWLEDGE_NOT_FOUND, "这份知识不存在")
+
+    settings = get_settings()
+    source = "\n\n".join(part for part in (item.title, item.body_md or "") if part)
+    drafts = split_markdown(
+        source,
+        max_chars=settings.chunk_max_chars,
+        overlap=settings.chunk_overlap_chars,
+    )
+
+    provider = get_provider(settings)
+    vectors = await provider.embed([draft.content for draft in drafts]) if drafts else []
+
+    await session.execute(
+        delete(Chunk).where(Chunk.workspace_id == ws_id, Chunk.item_id == item_id)
+    )
+    for index, (draft, vector) in enumerate(zip(drafts, vectors, strict=True)):
+        session.add(
+            Chunk(
+                workspace_id=ws_id,
+                item_id=item_id,
+                idx=index,
+                heading=draft.heading,
+                content=draft.content,
+                char_count=len(draft.content),
+                embedding=vector,
+                embed_model=provider.name,
+            )
+        )
+    doc.chunk_count = len(drafts)
+    await session.flush()
+    return len(drafts)
+
+
+async def search_knowledge(
+    session: AsyncSession,
+    ws_id: uuid.UUID,
+    query: str,
+    *,
+    limit: int = 10,
+    mode: str = "hybrid",
+) -> list[dict[str, Any]]:
+    """检索已发布知识，返回按相关度排序的切片。
+
+    `mode`: hybrid（默认）/ vector / keyword —— 前两者主要用于对比效果。
+    """
+    cleaned = (query or "").strip()
+    if not cleaned:
+        return []
+
+    settings = get_settings()
+    provider = get_provider(settings)
+    query_vector = (await provider.embed([cleaned]))[0]
+
+    if mode == "vector":
+        vector_weight, lexical_weight = 1.0, 0.0
+    elif mode == "keyword":
+        vector_weight, lexical_weight = 0.0, 1.0
+    else:
+        vector_weight, lexical_weight = (
+            settings.search_vector_weight,
+            settings.search_lexical_weight,
+        )
+
+    rows = (
+        await session.execute(
+            SEARCH_SQL,
+            {
+                "qvec": _vector_param(query_vector),
+                "q": cleaned,
+                "ws": str(ws_id),
+                "vw": vector_weight,
+                "lw": lexical_weight,
+                "limit": limit,
+            },
+        )
+    ).mappings()
+
+    hits: list[dict[str, Any]] = []
+    for row in rows:
+        vector_score = float(row["vec_score"] or 0.0)
+        lexical_score = float(row["lex_score"] or 0.0)
+        hits.append(
+            {
+                "item_id": row["item_id"],
+                "title": row["title"],
+                "doc_type": row["doc_type"],
+                "idx": row["idx"],
+                "heading": row["heading"],
+                "content": row["content"],
+                "score": round(vector_score * vector_weight + lexical_score * lexical_weight, 4),
+                "vector_score": round(vector_score, 4),
+                "lexical_score": round(lexical_score, 4),
+            }
+        )
+    hits = [hit for hit in hits if hit["score"] >= settings.search_min_score]
+    return hits
+
+
+async def reindex_workspace(session: AsyncSession, ws_id: uuid.UUID) -> int:
+    """把一个工作区里已发布的知识全部重建，返回处理的知识份数。
+
+    换 embedding 模型之后必须整体跑一次，否则新旧向量混在一张表里，
+    相似度分数没法比较。
+    """
+    item_ids = (
+        await session.execute(
+            select(KnowledgeDoc.item_id).where(
+                KnowledgeDoc.workspace_id == ws_id,
+                KnowledgeDoc.status_hint == "published",
+            )
+        )
+    ).scalars()
+    count = 0
+    for item_id in item_ids:
+        await reindex_document(session, ws_id, item_id)
+        count += 1
+    return count

backend/app/services/signing.py

@@ -37,7 +37,9 @@ def sign(workspace_id: uuid.UUID, attachment_id: uuid.UUID, ttl_seconds: int | N
     return f"exp={exp}&sig={_digest(workspace_id, attachment_id, exp)}"
 
 
-def verify(workspace_id: uuid.UUID, attachment_id: uuid.UUID, exp: int | None, sig: str | None) -> bool:
+def verify(
+    workspace_id: uuid.UUID, attachment_id: uuid.UUID, exp: int | None, sig: str | None
+) -> bool:
     if not exp or not sig:
         return False
     if exp < int(time.time()):  # 过期就是过期，不给宽限

backend/tests/test_attachments.py

@@ -8,13 +8,12 @@
 
 from __future__ import annotations
 
+import uuid
 from typing import Any
 
 import pytest
 from httpx import AsyncClient
 
-import uuid
-
 from app.config import get_settings
 from app.services import signing
 
@@ -178,7 +177,7 @@ async def test_raw_url_uses_short_lived_signature(client: AsyncClient) -> None:
     """图片直链自带短期签名：不需要把登录令牌贴进 URL，也不需要 Authorization 头。"""
     user, ws, item_id = await _ws_with_item(client)
     body = b"\x89PNG\r\n\x1a\n signed png"
-    created = unwrap(await _upload(client, user, ws["id"], item_id, body, "签名图.png"))
+    unwrap(await _upload(client, user, ws["id"], item_id, body, "签名图.png"))
 
     listed = unwrap(
         await client.get(

backend/tests/test_retrieval.py

@@ -0,0 +1,200 @@
+"""M3 检索测试：切片、索引、跨租户。
+
+检索最容易做错的地方不是"搜得准不准"，而是"什么该被搜到、什么绝不能被搜到"：
+
+1. **草稿不能被检索到**。草稿是还没被人认可的话，让它出现等于把猜测当结论；
+2. **刚发布就能搜到**。发布后要立刻建索引，否则用户会以为发布失败；
+3. **检索结果必须带来源**，能点回原文；
+4. **别的租户搜不到我的知识**，即使关键词一模一样。
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from httpx import AsyncClient
+
+from .conftest import create_workspace, login, unwrap
+
+V1 = "/api/v1"
+
+
+async def _knowledge(
+    client: AsyncClient,
+    user: dict[str, Any],
+    ws_id: str,
+    title: str,
+    body: str,
+    doc_type: str = "note",
+) -> dict[str, Any]:
+    return unwrap(
+        await client.post(
+            f"{V1}/workspaces/{ws_id}/knowledge",
+            json={"title": title, "body_md": body, "doc_type": doc_type},
+            headers=user["headers"],
+        )
+    )
+
+
+async def _publish(
+    client: AsyncClient, user: dict[str, Any], ws_id: str, doc: dict[str, Any]
+) -> dict[str, Any]:
+    return unwrap(
+        await client.post(
+            f"{V1}/workspaces/{ws_id}/knowledge/{doc['id']}/status",
+            json={"action": "publish", "version": doc["version"]},
+            headers=user["headers"],
+        )
+    )
+
+
+async def _search(
+    client: AsyncClient,
+    user: dict[str, Any],
+    ws_id: str,
+    query: str,
+    mode: str = "hybrid",
+) -> list[dict[str, Any]]:
+    return unwrap(
+        await client.get(
+            f"{V1}/workspaces/{ws_id}/knowledge/search",
+            params={"q": query, "mode": mode},
+            headers=user["headers"],
+        )
+    )["items"]
+
+
+async def test_published_is_searchable_draft_is_not(client: AsyncClient) -> None:
+    owner = await login(client)
+    ws = await create_workspace(client, owner, "检索测试")
+
+    published = await _knowledge(
+        client, owner, ws["id"], "报销发票怎么开", "电子发票原件必须一起寄，复印件不算。"
+    )
+    parked = await _knowledge(
+        client, owner, ws["id"], "供应商账号变更", "变更收款账号需要法人签字并盖章。"
+    )
+    await _publish(client, owner, ws["id"], published)
+
+    hits = await _search(client, owner, ws["id"], "发票")
+    ids = {hit["item_id"] for hit in hits}
+    assert published["id"] in ids
+    assert parked["id"] not in ids, "草稿不该出现在检索结果里"
+    assert hits[0]["score"] > 0
+    assert hits[0]["title"], "检索结果必须带来源标题"
+
+    # 还没验证过的知识，关键词再怎么对也不该冒出来
+    draft_hits = await _search(client, owner, ws["id"], "法人签字")
+    assert all(hit["item_id"] != parked["id"] for hit in draft_hits)
+
+
+async def test_publish_builds_chunks(client: AsyncClient) -> None:
+    owner = await login(client)
+    ws = await create_workspace(client, owner, "切片测试")
+
+    body = "\n\n".join(
+        [
+            "## 背景",
+            "十月的报销单据被财务退回两次，第一次是发票复印件，第二次是缺少行程单。" * 20,
+            "## 结论",
+            "报销必须同时提供电子发票原件与行程单，缺一不可。" * 20,
+        ]
+    )
+    doc = await _knowledge(client, owner, ws["id"], "十月报销被退单复盘", body, "postmortem")
+    published = await _publish(client, owner, ws["id"], doc)
+    assert published["knowledge"]["chunk_count"] >= 2, "长正文应该被切成多片"
+
+    detail = unwrap(
+        await client.get(
+            f"{V1}/workspaces/{ws['id']}/knowledge/{doc['id']}", headers=owner["headers"]
+        )
+    )
+    assert detail["knowledge"]["chunk_count"] >= 2
+
+
+async def test_reindex_picks_up_edits(client: AsyncClient) -> None:
+    owner = await login(client)
+    ws = await create_workspace(client, owner, "重建索引")
+
+    doc = await _knowledge(client, owner, ws["id"], "值班安排", "周一由张三负责。")
+    published = await _publish(client, owner, ws["id"], doc)
+    assert published["knowledge"]["chunk_count"] == 1
+
+    # 已发布的内容不许直接改（7003）：产品上的路径是「归档 → 恢复成草稿 → 改」
+    archived = unwrap(
+        await client.post(
+            f"{V1}/workspaces/{ws['id']}/knowledge/{doc['id']}/status",
+            json={"action": "archive", "version": published["version"]},
+            headers=owner["headers"],
+        )
+    )
+    draft = unwrap(
+        await client.post(
+            f"{V1}/workspaces/{ws['id']}/knowledge/{doc['id']}/status",
+            json={"action": "restore", "version": archived["version"]},
+            headers=owner["headers"],
+        )
+    )
+    patched = unwrap(
+        await client.patch(
+            f"{V1}/workspaces/{ws['id']}/knowledge/{doc['id']}",
+            json={
+                "body_md": "周一由张三负责，周三改由李四负责，节假日另行通知。",
+                "version": draft["version"],
+            },
+            headers=owner["headers"],
+        )
+    )
+    # 改完正文但没重建索引，旧切片还在——这就是 reindex 存在的理由
+    before = await _search(client, owner, ws["id"], "李四")
+    assert all("李四" not in hit["content"] for hit in before)
+
+    result = unwrap(
+        await client.post(
+            f"{V1}/workspaces/{ws['id']}/knowledge/{doc['id']}/reindex",
+            headers=owner["headers"],
+        )
+    )
+    assert result["chunk_count"] >= 1
+    # 仍是草稿，重建了也不该被检索到
+    assert result["searchable"] is False
+    assert await _search(client, owner, ws["id"], "李四") == []
+
+    await _publish(client, owner, ws["id"], patched)
+    after = await _search(client, owner, ws["id"], "李四")
+    assert any("李四" in hit["content"] for hit in after)
+
+
+async def test_search_is_tenant_scoped(client: AsyncClient) -> None:
+    alice = await login(client)
+    bob = await login(client)
+    ws_a = await create_workspace(client, alice, "甲队")
+    ws_b = await create_workspace(client, bob, "乙队")
+
+    doc = await _knowledge(client, alice, ws_a["id"], "发票规则", "电子发票原件必须一起寄。")
+    await _publish(client, alice, ws_a["id"], doc)
+
+    assert await _search(client, bob, ws_b["id"], "发票") == []
+    # 直接拿别人的工作区搜也不行
+    denied = await client.get(
+        f"{V1}/workspaces/{ws_a['id']}/knowledge/search",
+        params={"q": "发票"},
+        headers=bob["headers"],
+    )
+    assert denied.status_code in (403, 404)
+
+
+async def test_keyword_mode_reports_lexical_score(client: AsyncClient) -> None:
+    owner = await login(client)
+    ws = await create_workspace(client, owner, "词法检索")
+
+    doc = await _knowledge(
+        client, owner, ws["id"], "顺丰单号核对", "单号 SF1234567890 必须与发货记录一致。"
+    )
+    await _publish(client, owner, ws["id"], doc)
+
+    hits = await _search(client, owner, ws["id"], "SF1234567890", mode="keyword")
+    assert hits, "带专有编号的查询在词法模式下应该命中"
+    assert hits[0]["lexical_score"] > 0
+    # keyword 模式里向量权重是 0：总分只由词法分决定
+    assert hits[0]["score"] == hits[0]["lexical_score"]

web/e2e/retrieval.mjs

@@ -0,0 +1,120 @@
+/**
+ * 检索 e2e：发布的知识能被搜到，草稿搜不到，结果能点回原文。
+ *
+ * 只验证"用户能不能用上"，切片与打分逻辑由后端测试覆盖。
+ */
+import { chromium } from "playwright-core";
+
+const WEB = process.env.WEB ?? "http://127.0.0.1:5180";
+const API = process.env.API ?? "http://127.0.0.1:8100/api/v1";
+const stamp = String(Date.now()).slice(-7);
+const phone = `134${stamp}0`;
+
+let failures = 0;
+function check(name, ok, extra = "") {
+  console.log(`${ok ? "PASS" : "FAIL"} ${name}${extra ? ` — ${extra}` : ""}`);
+  if (!ok) failures += 1;
+}
+
+async function call(method, path, body, token) {
+  const res = await fetch(`${API}${path}`, {
+    method,
+    headers: {
+      "content-type": "application/json",
+      ...(token ? { authorization: `Bearer ${token}` } : {}),
+      ...(method === "GET" ? {} : { "idempotency-key": `e2e-${Math.random()}` }),
+    },
+    body: body ? JSON.stringify(body) : undefined,
+  });
+  const text = await res.text();
+  const json = text ? JSON.parse(text) : {};
+  if (json.code !== 0) throw new Error(`${method} ${path} -> ${json.code} ${json.message}`);
+  return json.data;
+}
+
+const browser = await chromium.launch({
+  executablePath:
+    process.env.CHROME ??
+    `${process.env.LOCALAPPDATA}\\ms-playwright\\chromium-1223\\chrome-win64\\chrome.exe`,
+});
+const consoleErrors = [];
+const context = await browser.newContext({ viewport: { width: 1280, height: 900 } });
+const page = await context.newPage();
+page.on("console", (m) => {
+  if (m.type() === "error") consoleErrors.push(m.text());
+});
+page.on("response", (r) => {
+  if (r.status() >= 400) consoleErrors.push(`${r.status()} ${r.url()}`);
+});
+
+try {
+  await call("POST", "/auth/sms/send", { phone });
+  const login = await call("POST", "/auth/login", { phone, code: "123456" });
+  const token = login.tokens.access_token;
+  const ws = await call("POST", "/workspaces", { name: `检索验收 ${stamp}` }, token);
+
+  const published = await call(
+    "POST",
+    `/workspaces/${ws.id}/knowledge`,
+    {
+      title: "快递单号对不上怎么处理",
+      doc_type: "howto",
+      body_md:
+        "## 步骤\n\n先在顺丰后台按手机号反查单号，再和发货记录逐一核对。\n\n如果差超过两天，直接找客服开异常件工单。",
+    },
+    token,
+  );
+  await call(
+    "POST",
+    `/workspaces/${ws.id}/knowledge/${published.id}/status`,
+    { action: "publish", version: published.version },
+    token,
+  );
+
+  const draft = await call(
+    "POST",
+    `/workspaces/${ws.id}/knowledge`,
+    { title: "还没核对的草稿", doc_type: "note", body_md: "这条写的是异常件的赔付口径，尚未定稿。" },
+    token,
+  );
+
+  await page.addInitScript(
+    ([access, refresh]) => {
+      localStorage.setItem("dadao.access", access);
+      localStorage.setItem("dadao.refresh", refresh);
+    },
+    [token, login.tokens.refresh_token],
+  );
+  await page.goto(`${WEB}/w/${ws.id}/knowledge`, { waitUntil: "networkidle" });
+
+  const box = page.getByPlaceholder(/搜标题或正文/);
+  await box.fill("异常件单号核对");
+  await box.press("Enter");
+  await page.getByText(/命中 \d+ 段/).waitFor({ timeout: 10_000 });
+  const summary = await page.getByText(/命中 \d+ 段/).first().textContent();
+  check("检索返回了结果", /命中 [1-9]\d* 段/.test(summary ?? ""), summary ?? "");
+  check("命中里带标题", await page.getByText("快递单号对不上怎么处理", { exact: false }).first().isVisible());
+  check("草稿不会出现在检索结果里", !(await page.getByText("还没核对的草稿", { exact: false }).count()));
+  await page.screenshot({ path: "e2e/artifacts/retrieval-hits.png", fullPage: false });
+
+  await page.getByRole("button", { name: "打开原文", exact: true }).first().click();
+  await page.getByText(/已发布 · 版本 \d+/, { exact: false }).waitFor({ timeout: 10_000 });
+  check("命中能点回原文", true);
+  await page.getByRole("button", { name: "关闭", exact: true }).click();
+
+  await box.fill("量子纠缠");
+  await box.press("Enter");
+  await page.getByText("没有检索到相关内容", { exact: false }).waitFor({ timeout: 10_000 });
+  check("不相干的关键词宁可返回空", true);
+
+  check("浏览器控制台与网络无报错", consoleErrors.length === 0, consoleErrors.slice(0, 3).join(" | "));
+} catch (err) {
+  failures += 1;
+  console.log(`FAIL 脚本异常 — ${err.message}`);
+  await page.screenshot({ path: "e2e/artifacts/retrieval-failure.png" }).catch(() => {});
+  console.log((await page.textContent("body"))?.slice(0, 600));
+} finally {
+  await browser.close();
+  console.log(failures === 0 ? "\nALL PASS" : `\n${failures} FAILED`);
+  process.exit(failures === 0 ? 0 : 1);
+}

web/src/lib/types.ts

@@ -157,3 +157,16 @@ export interface Knowledge extends Item {
   attachment_count?: number;
   activities?: Activity[];
 }
+
+/** 一条检索命中：切片级结果，带来源，点得回原文。 */
+export interface SearchHit {
+  item_id: string;
+  title: string;
+  doc_type: string;
+  idx: number;
+  heading: string | null;
+  content: string;
+  score: number;
+  vector_score: number;
+  lexical_score: number;
+}

web/src/pages/Knowledge.tsx

@@ -14,7 +14,7 @@ import {
   StatusPill,
 } from "../components/ui";
 import { ApiError, api } from "../lib/client";
-import type { Attachment, Knowledge, KnowledgeStatus } from "../lib/types";
+import type { Attachment, Knowledge, KnowledgeStatus, SearchHit } from "../lib/types";
 
 const TABS = [
   { key: "draft", label: "待核对", statuses: ["draft", "in_review"] },
@@ -63,6 +63,8 @@ export function KnowledgePage() {
   const [openId, setOpenId] = useState<string | null>(null);
   const [adding, setAdding] = useState(false);
   const [draftTitle, setDraftTitle] = useState("");
+  const [hits, setHits] = useState<SearchHit[] | null>(null);
+  const [searching, setSearching] = useState(false);
   const reqId = useRef(0);
 
   const load = useCallback(
@@ -99,6 +101,32 @@ export function KnowledgePage() {
     void load(null);
   }, [load]);
 
+  /**
+   * 语义检索走的是切片接口，和上面的列表筛选不是一回事：
+   * 列表按标题/正文做子串匹配，检索按向量 + trigram 打分，只覆盖已发布的知识。
+   * 所以不把 q 直接接到列表上——两套结果混在一起会让人搞不清命中在哪。
+   */
+  async function runSearch() {
+    const keyword = q.trim();
+    if (!keyword) {
+      setHits(null);
+      return;
+    }
+    setSearching(true);
+    setError(null);
+    try {
+      const query = new URLSearchParams({ q: keyword, limit: "12" });
+      const page = await api.get<{ items: SearchHit[] }>(
+        `/workspaces/${wsId}/knowledge/search?${query.toString()}`,
+      );
+      setHits(page.items);
+    } catch (err) {
+      setError(err instanceof ApiError ? err.message : "检索失败");
+    } finally {
+      setSearching(false);
+    }
+  }
+
   async function addNote() {
     const title = draftTitle.trim();
     if (!title) return;
@@ -133,11 +161,21 @@ export function KnowledgePage() {
               </span>
               <input
                 className="input h-9 w-48 pl-8 text-sm"
-                placeholder="搜标题或正文…"
+                placeholder="搜标题或正文…（回车检索）"
                 value={q}
-                onChange={(e) => setQ(e.target.value)}
+                onChange={(e) => {
+                  setQ(e.target.value);
+                  if (!e.target.value.trim()) setHits(null);
+                }}
+                onKeyDown={(e) => {
+                  if (e.key === "Enter") void runSearch();
+                }}
               />
             </div>
+            <Button variant="ghost" disabled={searching || !q.trim()} onClick={() => void runSearch()}>
+              <Search size={14} />
+              {searching ? "检索中…" : "检索"}
+            </Button>
             <Button variant="primary" onClick={() => setAdding((v) => !v)}>
               <Plus size={14} />
               记一条
@@ -204,7 +242,52 @@ export function KnowledgePage() {
         </div>
       )}
 
-      {loading && docs.length === 0 ? (
+      {hits !== null ? (
+        <section className="space-y-2">
+          <div className="flex items-baseline justify-between px-1">
+            <p className="text-xs text-ink-500">
+              在已发布的知识里检索「{q.trim()}」，命中 {hits.length} 段
+            </p>
+            <button className="link text-2xs" onClick={() => setHits(null)}>
+              退出检索
+            </button>
+          </div>
+          {hits.length === 0 ? (
+            <div className="card">
+              <EmptyState
+                icon={<Search size={18} />}
+                title="没有检索到相关内容"
+                hint="只检索已发布的知识。换成更具体的说法，或者先把手头的草稿核对发布。"
+              />
+            </div>
+          ) : (
+            hits.map((hit) => (
+              <article key={`${hit.item_id}-${hit.idx}`} className="card p-4">
+                <div className="flex flex-wrap items-center gap-1.5">
+                  <span className="chip bg-primary-50 text-primary-700">
+                    {DOC_TYPE_LABEL[hit.doc_type] ?? hit.doc_type}
+                  </span>
+                  {hit.heading ? (
+                    <span className="chip bg-paper-200 text-ink-600">{hit.heading}</span>
+                  ) : null}
+                  <span className="ml-auto text-2xs text-ink-400">
+                    相关度 {hit.score.toFixed(2)}
+                  </span>
+                </div>
+                <p className="mt-2 text-sm font-medium text-ink-900">{hit.title}</p>
+                <p className="mt-1.5 line-clamp-3 whitespace-pre-wrap text-xs leading-relaxed text-ink-600">
+                  {hit.content}
+                </p>
+                <div className="mt-2 text-right">
+                  <Button variant="ghost" size="sm" onClick={() => setOpenId(hit.item_id)}>
+                    打开原文
+                  </Button>
+                </div>
+              </article>
+            ))
+          )}
+        </section>
+      ) : loading && docs.length === 0 ? (
         <div className="card overflow-hidden">
           <SkeletonRows rows={3} />
         </div>