luckytri 1.0.0 → 1.0.2

package/server/knowledge/manager.js

@@ -1,386 +1,386 @@
-import { createHash, randomUUID } from "node:crypto";
-import { mkdirSync, writeFileSync } from "node:fs";
-import { join } from "node:path";
-import { runtimePaths } from "../paths.js";
-import {
-  asPlainDocument,
-  chunkText,
-  cosine,
-  ftsMatchQuery,
-  overlapScore,
-  lexicalTerms,
-  packVector,
-  unpackVector,
-} from "./retrieval.js";
-import { indexChunk } from "./schema.js";
-import { ensureVectorIndex, vectorCandidates } from "./vector-index.js";
-import { resolveEmbeddingProfile, embedBatches } from "./embedding-profile.js";
-import {
-  parseSessionKey,
-  scopedSessionAliases,
-} from "../channels/session-key.js";
-
-function scopesFor(db, session) {
-  try {
-    const parsed = parseSessionKey(session);
-    const aliases = scopedSessionAliases(db, session);
-    return {
-      aliases,
-      kind: parsed.kind,
-      privateKey: `__private__:${parsed.nativeId}`,
-    };
-  } catch {
-    return { aliases: [session], kind: "group", privateKey: "" };
-  }
-}
-
-function saidBy(rows) {
-  const groups = new Map();
-  let anon = 0;
-  for (const row of rows || []) {
-    if (row?.role === "assistant") continue;
-    const id = String(row?.userId || "");
-    const key = id || `\0${anon++}`;
-    const texts = groups.get(key) || [];
-    if (row?.text) texts.push(String(row.text));
-    groups.set(key, texts);
-  }
-  return [...groups.values()]
-    .map((texts) => texts.join(" "))
-    .filter((text) => text.trim());
-}
-
-function allowedCollection(db, collection, session) {
-  if (!session) return true;
-  const { aliases, kind, privateKey } = scopesFor(db, session);
-  if (collection.scope === "shared") return true;
-  if (privateKey && collection.scope === privateKey) return true;
-  if (aliases.includes(collection.scope)) return true;
-  if (String(collection.scope).startsWith("__private__") && kind !== "private")
-    return false;
-  if (!aliases.length) return false;
-  return !!db
-    .prepare(
-      `SELECT 1 FROM core_collection_sessions WHERE collection_id=? AND session_id IN (${aliases.map(() => "?").join(",")})`,
-    )
-    .get(collection.id, ...aliases);
-}
-
-export class KnowledgeManager {
-  constructor(repo, models) {
-    this.repo = repo;
-    this.models = models;
-  }
-  collections(session = "") {
-    return this.repo.db
-      .prepare("SELECT * FROM core_collections ORDER BY updated DESC")
-      .all()
-      .filter((c) => allowedCollection(this.repo.db, c, session));
-  }
-  createCollection({ name, scope = "shared", sessions = [] }) {
-    if (typeof name !== "string" || !name.trim() || name.length > 100)
-      throw Error("知识库名称无效");
-    const id = randomUUID(),
-      now = Date.now();
-    this.repo.db
-      .prepare(
-        "INSERT INTO core_collections(id,name,scope,created,updated) VALUES (?,?,?,?,?)",
-      )
-      .run(id, name.trim(), scope || "shared", now, now);
-    for (const session of sessions)
-      this.repo.db
-        .prepare("INSERT OR IGNORE INTO core_collection_sessions VALUES (?,?)")
-        .run(id, session);
-    this.repo.store.revision++;
-    return this.repo.db
-      .prepare("SELECT * FROM core_collections WHERE id=?")
-      .get(id);
-  }
-  documents(collectionId) {
-    return this.repo.db
-      .prepare(
-        "SELECT * FROM core_documents WHERE collection_id=? ORDER BY updated DESC",
-      )
-      .all(collectionId);
-  }
-  async ingest({ collectionId, title, text, source = "paste", embed = true }) {
-    const collection = this.repo.db
-      .prepare("SELECT * FROM core_collections WHERE id=?")
-      .get(collectionId);
-    if (!collection) throw Error("知识库不存在");
-    if (typeof text !== "string" || !text.trim() || text.length > 400000)
-      throw Error("文档内容应为 1–400000 字");
-    const safeTitle = String(title || "未命名文档").slice(0, 200);
-    const plain = asPlainDocument(text);
-    const chunks = chunkText(plain);
-    if (!chunks.length) throw Error("文档切分后为空");
-    const id = randomUUID(),
-      now = Date.now(),
-      hash = createHash("sha256").update(text).digest("hex");
-    mkdirSync(runtimePaths().knowledge, { recursive: true, mode: 0o700 });
-    const path = join(runtimePaths().knowledge, id + ".md");
-    writeFileSync(path, text);
-    this.repo.db
-      .prepare(
-        "INSERT INTO core_documents(id,collection_id,title,source,path,status,created,updated,bytes,hash) VALUES (?,?,?,?,?,'processing',?,?,?,?)",
-      )
-      .run(
-        id,
-        collectionId,
-        safeTitle,
-        source,
-        path,
-        now,
-        now,
-        Buffer.byteLength(text),
-        hash,
-      );
-    let vectors = [];
-    let embeddingProfile = "";
-    try {
-      if (embed) {
-        const resolved = resolveEmbeddingProfile(this.repo, this.models);
-        vectors = await embedBatches(
-          this.models,
-          resolved.profile,
-          chunks.map((c) => c.text),
-        );
-        if (vectors.length) embeddingProfile = resolved.signature;
-      }
-    } catch (error) {
-      this.repo.db
-        .prepare(
-          "UPDATE core_documents SET status='failed',error=?,updated=? WHERE id=?",
-        )
-        .run(error.message, Date.now(), id);
-      throw error;
-    }
-    this.repo.db.exec("BEGIN IMMEDIATE");
-    try {
-      chunks.forEach((chunk, i) => {
-        const chunkId = randomUUID();
-        this.repo.db
-          .prepare(
-            "INSERT INTO core_chunks(id,document_id,collection_id,ordinal,heading,text,tokens,embedding,hash,created,embedding_profile) VALUES (?,?,?,?,?,?,?,?,?,?,?)",
-          )
-          .run(
-            chunkId,
-            id,
-            collectionId,
-            i,
-            chunk.heading,
-            chunk.text,
-            chunk.text.slice(0, 200),
-            vectors[i] ? packVector(vectors[i]) : null,
-            createHash("sha256").update(chunk.text).digest("hex"),
-            now,
-            embeddingProfile,
-          );
-        indexChunk(this.repo.db, chunkId, chunk.text);
-      });
-      this.repo.db
-        .prepare(
-          "UPDATE core_documents SET status='ready',error=NULL,updated=? WHERE id=?",
-        )
-        .run(Date.now(), id);
-      this.repo.db
-        .prepare("UPDATE core_collections SET updated=? WHERE id=?")
-        .run(Date.now(), collectionId);
-      this.repo.db.exec("COMMIT");
-    } catch (error) {
-      this.repo.db.exec("ROLLBACK");
-      throw error;
-    }
-    this.repo.store.revision++;
-    return this.repo.db
-      .prepare("SELECT * FROM core_documents WHERE id=?")
-      .get(id);
-  }
-  async embedTexts(texts) {
-    const { profile } = resolveEmbeddingProfile(this.repo, this.models);
-    return embedBatches(this.models, profile, texts);
-  }
-  retrieve(
-    session,
-    rows,
-    cutoff = Date.now(),
-    { vectors = [], signature = null, legacy = true } = {},
-  ) {
-    const allowed = new Set(
-      this.repo.db
-        .prepare("SELECT id,scope FROM core_collections")
-        .all()
-        .filter((c) => allowedCollection(this.repo.db, c, session))
-        .map((c) => c.id),
-    );
-    if (!allowed.size) return [];
-    const queries = saidBy(rows).slice(0, 8);
-    if (!queries.length) return [];
-    // Each person's own words. Two people are not added together, and a
-    // line she already said does not count as the room asking. A prepared
-    // vector belongs to that same person, not to the batch mixed together.
-    if (vectors.some((vector) => vector?.length))
-      ensureVectorIndex(this.repo.db);
-    const best = new Map();
-    const selected = new Map();
-    for (let i = 0; i < queries.length; i++) {
-      const query = queries[i];
-      const match = ftsMatchQuery(query);
-      const fts = new Set();
-      if (match)
-        try {
-          for (const hit of this.repo.db
-            .prepare(
-              `SELECT f.chunk_id FROM core_chunk_fts f JOIN core_chunks c ON c.id=f.chunk_id JOIN core_documents d ON d.id=c.document_id WHERE f.tokens MATCH ? AND c.collection_id IN (${[...allowed].map(() => "?").join(",")}) AND d.status='ready' AND d.created<=? ORDER BY f.rank LIMIT 128`,
-            )
-            .all(match, ...allowed, cutoff))
-            fts.add(hit.chunk_id);
-        } catch {
-          /* malformed MATCH */
-        }
-      const queryTerms = lexicalTerms(query);
-      const vec = vectors[i] || null;
-      const ids = [
-        ...new Set([
-          ...fts,
-          ...(vec
-            ? vectorCandidates(this.repo.db, vec, [...allowed], cutoff, {
-                signature,
-                legacy,
-              })
-            : []),
-        ]),
-      ];
-      const chunks = ids.length
-        ? this.repo.db
-            .prepare(
-              `SELECT c.*,d.title,d.created AS document_created FROM core_chunks c JOIN core_documents d ON d.id=c.document_id WHERE c.id IN (${ids.map(() => "?").join(",")})`,
-            )
-            .all(...ids)
-        : [];
-      for (const chunk of chunks) {
-        selected.set(chunk.id, chunk);
-        let score =
-          (fts.has(chunk.id) ? 2 : 0) +
-          overlapScore(chunk.text, queryTerms) * 0.5;
-        if (
-          vec &&
-          chunk.embedding &&
-          (!signature ||
-            chunk.embedding_profile === signature ||
-            (legacy && !chunk.embedding_profile))
-        )
-          score += cosine(vec, unpackVector(chunk.embedding)) * 4;
-        const ageDays = (cutoff - (chunk.document_created || 0)) / 86400000;
-        if (ageDays > 30) score *= 0.85;
-        if (score > (best.get(chunk.id) || 0)) best.set(chunk.id, score);
-      }
-    }
-    return [...best.entries()]
-      .sort((a, b) => b[1] - a[1])
-      .slice(0, 8)
-      .map(([id, score]) => {
-        const chunk = selected.get(id);
-        if (!chunk) return null;
-        return {
-          id: chunk.id,
-          documentId: chunk.document_id,
-          title: chunk.title,
-          heading: chunk.heading,
-          text: chunk.text.slice(0, 800),
-          score,
-          whySelected: score >= 4 ? "semantic+lexical" : "lexical",
-        };
-      })
-      .filter((c) => c?.text);
-  }
-  async retrieveWithEmbed(session, rows, cutoff = Date.now()) {
-    const { profile, signature, legacy } = resolveEmbeddingProfile(
-      this.repo,
-      this.models,
-    );
-    const queries = saidBy(rows).slice(0, 8);
-    let vectors = [];
-    if (profile.embedding && queries.length) {
-      try {
-        vectors = await embedBatches(
-          this.models,
-          profile,
-          queries.map((query) => query.slice(0, 2000)),
-        );
-      } catch {
-        vectors = [];
-      }
-    }
-    return this.retrieve(session, rows, cutoff, { vectors, signature, legacy });
-  }
-  async reembedDocument(id) {
-    const rows = this.repo.db
-      .prepare(
-        "SELECT id,text FROM core_chunks WHERE document_id=? ORDER BY ordinal",
-      )
-      .all(id);
-    if (!rows.length) throw Error("文档没有可建立向量的段落");
-    const { profile, signature } = resolveEmbeddingProfile(
-      this.repo,
-      this.models,
-    );
-    if (!profile.embedding) throw Error("请先启用知识向量模型");
-    const vectors = await embedBatches(
-      this.models,
-      profile,
-      rows.map((row) => row.text),
-    );
-    const db = this.repo.db;
-    db.exec("SAVEPOINT reembed_document");
-    try {
-      const update = db.prepare(
-        "UPDATE core_chunks SET embedding=?,embedding_profile=? WHERE id=? AND document_id=? AND text=?",
-      );
-      rows.forEach((row, index) => {
-        if (
-          !update.run(
-            packVector(vectors[index]),
-            signature,
-            row.id,
-            id,
-            row.text,
-          ).changes
-        )
-          throw Error("文档已变化，请重新建立向量");
-      });
-      db.exec("RELEASE reembed_document");
-    } catch (error) {
-      db.exec("ROLLBACK TO reembed_document");
-      db.exec("RELEASE reembed_document");
-      throw error;
-    }
-    this.repo.store.revision++;
-    return { chunks: rows.length };
-  }
-  removeDocument(id) {
-    const row = this.repo.db
-      .prepare("SELECT * FROM core_documents WHERE id=?")
-      .get(id);
-    if (!row) throw Error("文档不存在");
-    const chunks = this.repo.db
-      .prepare("SELECT id FROM core_chunks WHERE document_id=?")
-      .all(id);
-    this.repo.db.exec("BEGIN IMMEDIATE");
-    try {
-      for (const chunk of chunks)
-        this.repo.db
-          .prepare("DELETE FROM core_chunk_fts WHERE chunk_id=?")
-          .run(chunk.id);
-      this.repo.db
-        .prepare("DELETE FROM core_chunks WHERE document_id=?")
-        .run(id);
-      this.repo.db.prepare("DELETE FROM core_documents WHERE id=?").run(id);
-      this.repo.db.exec("COMMIT");
-    } catch (e) {
-      this.repo.db.exec("ROLLBACK");
-      throw e;
-    }
-    this.repo.store.revision++;
-  }
-}
+import { createHash, randomUUID } from "node:crypto";
+import { mkdirSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+import { runtimePaths } from "../paths.js";
+import {
+  asPlainDocument,
+  chunkText,
+  cosine,
+  ftsMatchQuery,
+  overlapScore,
+  lexicalTerms,
+  packVector,
+  unpackVector,
+} from "./retrieval.js";
+import { indexChunk } from "./schema.js";
+import { ensureVectorIndex, vectorCandidates } from "./vector-index.js";
+import { resolveEmbeddingProfile, embedBatches } from "./embedding-profile.js";
+import {
+  parseSessionKey,
+  scopedSessionAliases,
+} from "../channels/session-key.js";
+
+function scopesFor(db, session) {
+  try {
+    const parsed = parseSessionKey(session);
+    const aliases = scopedSessionAliases(db, session);
+    return {
+      aliases,
+      kind: parsed.kind,
+      privateKey: `__private__:${parsed.nativeId}`,
+    };
+  } catch {
+    return { aliases: [session], kind: "group", privateKey: "" };
+  }
+}
+
+function saidBy(rows) {
+  const groups = new Map();
+  let anon = 0;
+  for (const row of rows || []) {
+    if (row?.role === "assistant") continue;
+    const id = String(row?.userId || "");
+    const key = id || `\0${anon++}`;
+    const texts = groups.get(key) || [];
+    if (row?.text) texts.push(String(row.text));
+    groups.set(key, texts);
+  }
+  return [...groups.values()]
+    .map((texts) => texts.join(" "))
+    .filter((text) => text.trim());
+}
+
+function allowedCollection(db, collection, session) {
+  if (!session) return true;
+  const { aliases, kind, privateKey } = scopesFor(db, session);
+  if (collection.scope === "shared") return true;
+  if (privateKey && collection.scope === privateKey) return true;
+  if (aliases.includes(collection.scope)) return true;
+  if (String(collection.scope).startsWith("__private__") && kind !== "private")
+    return false;
+  if (!aliases.length) return false;
+  return !!db
+    .prepare(
+      `SELECT 1 FROM core_collection_sessions WHERE collection_id=? AND session_id IN (${aliases.map(() => "?").join(",")})`,
+    )
+    .get(collection.id, ...aliases);
+}
+
+export class KnowledgeManager {
+  constructor(repo, models) {
+    this.repo = repo;
+    this.models = models;
+  }
+  collections(session = "") {
+    return this.repo.db
+      .prepare("SELECT * FROM core_collections ORDER BY updated DESC")
+      .all()
+      .filter((c) => allowedCollection(this.repo.db, c, session));
+  }
+  createCollection({ name, scope = "shared", sessions = [] }) {
+    if (typeof name !== "string" || !name.trim() || name.length > 100)
+      throw Error("知识库名称无效");
+    const id = randomUUID(),
+      now = Date.now();
+    this.repo.db
+      .prepare(
+        "INSERT INTO core_collections(id,name,scope,created,updated) VALUES (?,?,?,?,?)",
+      )
+      .run(id, name.trim(), scope || "shared", now, now);
+    for (const session of sessions)
+      this.repo.db
+        .prepare("INSERT OR IGNORE INTO core_collection_sessions VALUES (?,?)")
+        .run(id, session);
+    this.repo.store.revision++;
+    return this.repo.db
+      .prepare("SELECT * FROM core_collections WHERE id=?")
+      .get(id);
+  }
+  documents(collectionId) {
+    return this.repo.db
+      .prepare(
+        "SELECT * FROM core_documents WHERE collection_id=? ORDER BY updated DESC",
+      )
+      .all(collectionId);
+  }
+  async ingest({ collectionId, title, text, source = "paste", embed = true }) {
+    const collection = this.repo.db
+      .prepare("SELECT * FROM core_collections WHERE id=?")
+      .get(collectionId);
+    if (!collection) throw Error("知识库不存在");
+    if (typeof text !== "string" || !text.trim() || text.length > 400000)
+      throw Error("文档内容应为 1–400000 字");
+    const safeTitle = String(title || "未命名文档").slice(0, 200);
+    const plain = asPlainDocument(text);
+    const chunks = chunkText(plain);
+    if (!chunks.length) throw Error("文档切分后为空");
+    const id = randomUUID(),
+      now = Date.now(),
+      hash = createHash("sha256").update(text).digest("hex");
+    mkdirSync(runtimePaths().knowledge, { recursive: true, mode: 0o700 });
+    const path = join(runtimePaths().knowledge, id + ".md");
+    writeFileSync(path, text);
+    this.repo.db
+      .prepare(
+        "INSERT INTO core_documents(id,collection_id,title,source,path,status,created,updated,bytes,hash) VALUES (?,?,?,?,?,'processing',?,?,?,?)",
+      )
+      .run(
+        id,
+        collectionId,
+        safeTitle,
+        source,
+        path,
+        now,
+        now,
+        Buffer.byteLength(text),
+        hash,
+      );
+    let vectors = [];
+    let embeddingProfile = "";
+    try {
+      if (embed) {
+        const resolved = resolveEmbeddingProfile(this.repo, this.models);
+        vectors = await embedBatches(
+          this.models,
+          resolved.profile,
+          chunks.map((c) => c.text),
+        );
+        if (vectors.length) embeddingProfile = resolved.signature;
+      }
+    } catch (error) {
+      this.repo.db
+        .prepare(
+          "UPDATE core_documents SET status='failed',error=?,updated=? WHERE id=?",
+        )
+        .run(error.message, Date.now(), id);
+      throw error;
+    }
+    this.repo.db.exec("BEGIN IMMEDIATE");
+    try {
+      chunks.forEach((chunk, i) => {
+        const chunkId = randomUUID();
+        this.repo.db
+          .prepare(
+            "INSERT INTO core_chunks(id,document_id,collection_id,ordinal,heading,text,tokens,embedding,hash,created,embedding_profile) VALUES (?,?,?,?,?,?,?,?,?,?,?)",
+          )
+          .run(
+            chunkId,
+            id,
+            collectionId,
+            i,
+            chunk.heading,
+            chunk.text,
+            chunk.text.slice(0, 200),
+            vectors[i] ? packVector(vectors[i]) : null,
+            createHash("sha256").update(chunk.text).digest("hex"),
+            now,
+            embeddingProfile,
+          );
+        indexChunk(this.repo.db, chunkId, chunk.text);
+      });
+      this.repo.db
+        .prepare(
+          "UPDATE core_documents SET status='ready',error=NULL,updated=? WHERE id=?",
+        )
+        .run(Date.now(), id);
+      this.repo.db
+        .prepare("UPDATE core_collections SET updated=? WHERE id=?")
+        .run(Date.now(), collectionId);
+      this.repo.db.exec("COMMIT");
+    } catch (error) {
+      this.repo.db.exec("ROLLBACK");
+      throw error;
+    }
+    this.repo.store.revision++;
+    return this.repo.db
+      .prepare("SELECT * FROM core_documents WHERE id=?")
+      .get(id);
+  }
+  async embedTexts(texts) {
+    const { profile } = resolveEmbeddingProfile(this.repo, this.models);
+    return embedBatches(this.models, profile, texts);
+  }
+  retrieve(
+    session,
+    rows,
+    cutoff = Date.now(),
+    { vectors = [], signature = null, legacy = true } = {},
+  ) {
+    const allowed = new Set(
+      this.repo.db
+        .prepare("SELECT id,scope FROM core_collections")
+        .all()
+        .filter((c) => allowedCollection(this.repo.db, c, session))
+        .map((c) => c.id),
+    );
+    if (!allowed.size) return [];
+    const queries = saidBy(rows).slice(0, 8);
+    if (!queries.length) return [];
+    // Each person's own words. Two people are not added together, and a
+    // line she already said does not count as the room asking. A prepared
+    // vector belongs to that same person, not to the batch mixed together.
+    if (vectors.some((vector) => vector?.length))
+      ensureVectorIndex(this.repo.db);
+    const best = new Map();
+    const selected = new Map();
+    for (let i = 0; i < queries.length; i++) {
+      const query = queries[i];
+      const match = ftsMatchQuery(query);
+      const fts = new Set();
+      if (match)
+        try {
+          for (const hit of this.repo.db
+            .prepare(
+              `SELECT f.chunk_id FROM core_chunk_fts f JOIN core_chunks c ON c.id=f.chunk_id JOIN core_documents d ON d.id=c.document_id WHERE f.tokens MATCH ? AND c.collection_id IN (${[...allowed].map(() => "?").join(",")}) AND d.status='ready' AND d.created<=? ORDER BY f.rank LIMIT 128`,
+            )
+            .all(match, ...allowed, cutoff))
+            fts.add(hit.chunk_id);
+        } catch {
+          /* malformed MATCH */
+        }
+      const queryTerms = lexicalTerms(query);
+      const vec = vectors[i] || null;
+      const ids = [
+        ...new Set([
+          ...fts,
+          ...(vec
+            ? vectorCandidates(this.repo.db, vec, [...allowed], cutoff, {
+                signature,
+                legacy,
+              })
+            : []),
+        ]),
+      ];
+      const chunks = ids.length
+        ? this.repo.db
+            .prepare(
+              `SELECT c.*,d.title,d.created AS document_created FROM core_chunks c JOIN core_documents d ON d.id=c.document_id WHERE c.id IN (${ids.map(() => "?").join(",")})`,
+            )
+            .all(...ids)
+        : [];
+      for (const chunk of chunks) {
+        selected.set(chunk.id, chunk);
+        let score =
+          (fts.has(chunk.id) ? 2 : 0) +
+          overlapScore(chunk.text, queryTerms) * 0.5;
+        if (
+          vec &&
+          chunk.embedding &&
+          (!signature ||
+            chunk.embedding_profile === signature ||
+            (legacy && !chunk.embedding_profile))
+        )
+          score += cosine(vec, unpackVector(chunk.embedding)) * 4;
+        const ageDays = (cutoff - (chunk.document_created || 0)) / 86400000;
+        if (ageDays > 30) score *= 0.85;
+        if (score > (best.get(chunk.id) || 0)) best.set(chunk.id, score);
+      }
+    }
+    return [...best.entries()]
+      .sort((a, b) => b[1] - a[1])
+      .slice(0, 8)
+      .map(([id, score]) => {
+        const chunk = selected.get(id);
+        if (!chunk) return null;
+        return {
+          id: chunk.id,
+          documentId: chunk.document_id,
+          title: chunk.title,
+          heading: chunk.heading,
+          text: chunk.text.slice(0, 800),
+          score,
+          whySelected: score >= 4 ? "semantic+lexical" : "lexical",
+        };
+      })
+      .filter((c) => c?.text);
+  }
+  async retrieveWithEmbed(session, rows, cutoff = Date.now()) {
+    const { profile, signature, legacy } = resolveEmbeddingProfile(
+      this.repo,
+      this.models,
+    );
+    const queries = saidBy(rows).slice(0, 8);
+    let vectors = [];
+    if (profile.embedding && queries.length) {
+      try {
+        vectors = await embedBatches(
+          this.models,
+          profile,
+          queries.map((query) => query.slice(0, 2000)),
+        );
+      } catch {
+        vectors = [];
+      }
+    }
+    return this.retrieve(session, rows, cutoff, { vectors, signature, legacy });
+  }
+  async reembedDocument(id) {
+    const rows = this.repo.db
+      .prepare(
+        "SELECT id,text FROM core_chunks WHERE document_id=? ORDER BY ordinal",
+      )
+      .all(id);
+    if (!rows.length) throw Error("文档没有可建立向量的段落");
+    const { profile, signature } = resolveEmbeddingProfile(
+      this.repo,
+      this.models,
+    );
+    if (!profile.embedding) throw Error("请先启用知识向量模型");
+    const vectors = await embedBatches(
+      this.models,
+      profile,
+      rows.map((row) => row.text),
+    );
+    const db = this.repo.db;
+    db.exec("SAVEPOINT reembed_document");
+    try {
+      const update = db.prepare(
+        "UPDATE core_chunks SET embedding=?,embedding_profile=? WHERE id=? AND document_id=? AND text=?",
+      );
+      rows.forEach((row, index) => {
+        if (
+          !update.run(
+            packVector(vectors[index]),
+            signature,
+            row.id,
+            id,
+            row.text,
+          ).changes
+        )
+          throw Error("文档已变化，请重新建立向量");
+      });
+      db.exec("RELEASE reembed_document");
+    } catch (error) {
+      db.exec("ROLLBACK TO reembed_document");
+      db.exec("RELEASE reembed_document");
+      throw error;
+    }
+    this.repo.store.revision++;
+    return { chunks: rows.length };
+  }
+  removeDocument(id) {
+    const row = this.repo.db
+      .prepare("SELECT * FROM core_documents WHERE id=?")
+      .get(id);
+    if (!row) throw Error("文档不存在");
+    const chunks = this.repo.db
+      .prepare("SELECT id FROM core_chunks WHERE document_id=?")
+      .all(id);
+    this.repo.db.exec("BEGIN IMMEDIATE");
+    try {
+      for (const chunk of chunks)
+        this.repo.db
+          .prepare("DELETE FROM core_chunk_fts WHERE chunk_id=?")
+          .run(chunk.id);
+      this.repo.db
+        .prepare("DELETE FROM core_chunks WHERE document_id=?")
+        .run(id);
+      this.repo.db.prepare("DELETE FROM core_documents WHERE id=?").run(id);
+      this.repo.db.exec("COMMIT");
+    } catch (e) {
+      this.repo.db.exec("ROLLBACK");
+      throw e;
+    }
+    this.repo.store.revision++;
+  }
+}