tinywebui 0.1.3 → 0.1.5

package/src/store/documents.js

@@ -1,14 +1,11 @@
 import { scope, ftsQuery } from './scope.js';
 import { createHash, randomBytes } from 'node:crypto';
 import { CORPUS } from './embeddings.js';
-
 // Attached documents, split into passages for read_document (BM25 and dense retrieval).
 // One area of the Store; see index.js.
-
 /** Default passage split: characters per passage, and how much neighbours overlap. */
 export const PASSAGE_SIZE = 1800;
 export const PASSAGE_OVERLAP = 200;
-
 /**
  * Splits text into overlapping passages -- the unit read_document returns.
  * Character-based on purpose: passages are for reading and BM25, and the
@@ -17,94 +14,97 @@ export const PASSAGE_OVERLAP = 200;
  * passage doesn't open or close mid-sentence more than it has to.
  */
 export function splitPassages(text, size = PASSAGE_SIZE, overlap = PASSAGE_OVERLAP) {
-  const passages = [];
-  let start = 0;
-  while (start < text.length) {
-    let end = Math.min(start + size, text.length);
-    if (end < text.length) {
-      const para = text.lastIndexOf('\n\n', end);
-      const line = text.lastIndexOf('\n', end);
-      const boundary = para > start + size * 0.5 ? para : (line > start + size * 0.5 ? line : -1);
-      if (boundary !== -1) end = boundary;
+    const passages = [];
+    let start = 0;
+    while (start < text.length) {
+        let end = Math.min(start + size, text.length);
+        if (end < text.length) {
+            const para = text.lastIndexOf('\n\n', end);
+            const line = text.lastIndexOf('\n', end);
+            const boundary = para > start + size * 0.5 ? para : (line > start + size * 0.5 ? line : -1);
+            if (boundary !== -1)
+                end = boundary;
+        }
+        passages.push({ start, text: text.slice(start, end) });
+        if (end >= text.length)
+            break;
+        start = Math.max(end - overlap, start + 1);
     }
-    passages.push({ start, text: text.slice(start, end) });
-    if (end >= text.length) break;
-    start = Math.max(end - overlap, start + 1);
-  }
-  return passages;
+    return passages;
 }
-
 const md5Of = (bytes) => createHash('md5').update(bytes).digest('hex');
-
 function insertPassages(db, docId, chatId, content, size, overlap) {
-  const passages = splitPassages(content, size, overlap);
-  const insertPassage = db.prepare('INSERT INTO document_passages (body) VALUES (?)');
-  const insertMap = db.prepare(`
+    const passages = splitPassages(content, size, overlap);
+    const insertPassage = db.prepare('INSERT INTO document_passages (body) VALUES (?)');
+    const insertMap = db.prepare(`
     INSERT INTO document_passage_map (passage_rowid, doc_id, chat_id, passage_idx, char_start)
     VALUES (?, ?, ?, ?, ?)
   `);
-  passages.forEach((passage, idx) => {
-    const { lastInsertRowid } = insertPassage.run(passage.text);
-    insertMap.run(lastInsertRowid, docId, chatId, idx, passage.start);
-  });
-  return passages.length;
+    passages.forEach((passage, idx) => {
+        const { lastInsertRowid } = insertPassage.run(passage.text);
+        insertMap.run(lastInsertRowid, docId, chatId, idx, passage.start);
+    });
+    return passages.length;
 }
-
 export class DocumentStore {
-  constructor(db, deps = {}) {
-    this.db = db;
-    Object.assign(this, deps);
-  }
-
-  /**
-   * `original`: the uploaded bytes, kept (once per distinct file) so the file
-   * itself can be opened later. `md5` names them when the caller already has
-   * the fingerprint, e.g. copying a document into another chat.
-   */
-  add(chatId, { filename, mime, content, original = null, md5 = null }, { passageSize = PASSAGE_SIZE, passageOverlap = PASSAGE_OVERLAP } = {}) {
-    const id = randomBytes(12).toString('hex');
-    const createdAt = Date.now();
-    const hash = original ? md5Of(original) : md5;
-    this.db.exec('BEGIN');
-    try {
-      if (original) this.db.prepare('INSERT OR IGNORE INTO document_blobs (md5, data) VALUES (?, ?)').run(hash, original);
-      this.db.prepare(`
+    constructor(db, deps) {
+        this.db = db;
+        Object.assign(this, deps);
+    }
+    claimUnowned(userId) {
+        this.db.prepare('UPDATE documents SET user_id = ? WHERE user_id IS NULL').run(userId);
+    }
+    /**
+     * `original`: the uploaded bytes, kept (once per distinct file) so the file
+     * itself can be opened later. `md5` names them when the caller already has
+     * the fingerprint, e.g. copying a document into another chat.
+     */
+    add(chatId, { filename, mime, content, original = null, md5 = null }, { passageSize = PASSAGE_SIZE, passageOverlap = PASSAGE_OVERLAP } = {}) {
+        const id = randomBytes(12).toString('hex');
+        const createdAt = Date.now();
+        const hash = original ? md5Of(original) : md5;
+        this.db.exec('BEGIN');
+        try {
+            if (original)
+                this.db.prepare('INSERT OR IGNORE INTO document_blobs (md5, data) VALUES (?, ?)').run(hash, original);
+            this.db.prepare(`
         INSERT INTO documents (id, chat_id, filename, mime, char_len, created_at, content, passage_size, passage_overlap, md5)
         VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
       `).run(id, chatId, filename, mime ?? null, content.length, createdAt, content, passageSize, passageOverlap, hash);
-      const passages = insertPassages(this.db, id, chatId, content, passageSize, passageOverlap);
-      this.db.exec('COMMIT');
-      return { id, filename, mime: mime ?? null, char_len: content.length, created_at: createdAt, passages };
-    } catch (err) { this.db.exec('ROLLBACK'); throw err; }
-  }
-
-  /** The chat's document with these bytes, if it already has one -- a file is attached to a chat once. */
-  findInChat(chatId, original) {
-    return this.#inChat(chatId, md5Of(original));
-  }
-
-  #inChat(chatId, md5) {
-    return this.db.prepare('SELECT id, filename, mime, char_len, created_at FROM documents WHERE chat_id = ? AND md5 = ?')
-      .get(chatId, md5) || null;
-  }
-
-  /** The uploaded bytes as a Buffer, or null (no such document, not yours, or stored before originals were kept). */
-  original(id, userId) {
-    const doc = this.get(id, userId);
-    if (!doc?.md5) return null;
-    const row = this.db.prepare('SELECT data FROM document_blobs WHERE md5 = ?').get(doc.md5);
-    return row ? Buffer.from(row.data) : null;
-  }
-
-  /**
-   * What the user has attached before, across all their chats, newest first --
-   * for picking a file again without uploading it. One entry per distinct file:
-   * the same bytes in ten chats is one line, and it names the latest chat.
-   */
-  library(userId, limit = 200) {
-    const s = scope(userId, 'c.user_id');
-    const inner = scope(userId, 'c2.user_id'); // the "latest" must be the user's own copy, not someone else's
-    return this.db.prepare(`
+            const passages = insertPassages(this.db, id, chatId, content, passageSize, passageOverlap);
+            this.db.exec('COMMIT');
+            return { id, filename, mime: mime ?? null, char_len: content.length, created_at: createdAt, passages };
+        }
+        catch (err) {
+            this.db.exec('ROLLBACK');
+            throw err;
+        }
+    }
+    /** The chat's document with these bytes, if it already has one -- a file is attached to a chat once. */
+    findInChat(chatId, original) {
+        return this.#inChat(chatId, md5Of(original));
+    }
+    #inChat(chatId, md5) {
+        return this.db.prepare('SELECT id, filename, mime, char_len, created_at FROM documents WHERE chat_id = ? AND md5 = ?')
+            .get(chatId, md5) || null;
+    }
+    /** The uploaded bytes as a Buffer, or null (no such document, not yours, or stored before originals were kept). */
+    original(id, userId) {
+        const doc = this.get(id, userId);
+        if (!doc?.md5)
+            return null;
+        const row = this.db.prepare('SELECT data FROM document_blobs WHERE md5 = ?').get(doc.md5);
+        return row ? Buffer.from(row.data) : null;
+    }
+    /**
+     * What the user has attached before, across all their chats, newest first --
+     * for picking a file again without uploading it. One entry per distinct file:
+     * the same bytes in ten chats is one line, and it names the latest chat.
+     */
+    library(userId, limit = 200) {
+        const s = scope(userId, 'c.user_id');
+        const inner = scope(userId, 'c2.user_id'); // the "latest" must be the user's own copy, not someone else's
+        return this.db.prepare(`
       SELECT d.id, d.filename, d.mime, d.char_len, d.created_at, c.title AS chatTitle
       FROM documents d JOIN chats c ON c.id = d.chat_id
       WHERE ${s.sql}
@@ -116,100 +116,98 @@ export class DocumentStore {
       ORDER BY d.created_at DESC
       LIMIT ?
     `).all(...s.params, ...inner.params, limit);
-  }
-
-  /**
-   * Puts a copy of one of the user's documents into another chat of theirs.
-   * Documents are chat-scoped (the chat's deletion takes them, its passages
-   * carry its id), so this is a new row: same text, same file blob, its own
-   * passages. Null when the source is not theirs; the chat's existing copy
-   * when it already has this file.
-   */
-  copyToChat(sourceId, chatId, userId, passageSettings) {
-    const src = this.get(sourceId, userId);
-    if (!src) return null;
-    if (src.md5) {
-      const have = this.#inChat(chatId, src.md5);
-      if (have) return { ...have, existing: true };
     }
-    return this.add(chatId, { filename: src.filename, mime: src.mime, content: src.content, md5: src.md5 }, passageSettings);
-  }
-
-  /** Documents split with other passage settings than these (NULL = the original 1800/200). */
-  withOtherPassages(size, overlap) {
-    return this.db.prepare(`
+    /**
+     * Puts a copy of one of the user's documents into another chat of theirs.
+     * Documents are chat-scoped (the chat's deletion takes them, its passages
+     * carry its id), so this is a new row: same text, same file blob, its own
+     * passages. Null when the source is not theirs; the chat's existing copy
+     * when it already has this file.
+     */
+    copyToChat(sourceId, chatId, userId, passageSettings) {
+        const src = this.get(sourceId, userId);
+        if (!src)
+            return null;
+        if (src.md5) {
+            const have = this.#inChat(chatId, src.md5);
+            if (have)
+                return { ...have, existing: true };
+        }
+        return this.add(chatId, { filename: src.filename, mime: src.mime, content: src.content, md5: src.md5 }, passageSettings);
+    }
+    /** Documents split with other passage settings than these (NULL = the original 1800/200). */
+    withOtherPassages(size, overlap) {
+        return this.db.prepare(`
       SELECT id FROM documents
       WHERE COALESCE(passage_size, ${PASSAGE_SIZE}) != ? OR COALESCE(passage_overlap, ${PASSAGE_OVERLAP}) != ?
     `).all(size, overlap).map((r) => r.id);
-  }
-
-  /** Re-splits a stored document with new passage settings; its old vectors go with the old passages. */
-  repassage(docId, size, overlap) {
-    const doc = this.db.prepare('SELECT id, chat_id, content FROM documents WHERE id = ?').get(docId);
-    if (!doc) return 0;
-    this.db.exec('BEGIN');
-    try {
-      this.#deletePassages(docId);
-      const n = insertPassages(this.db, doc.id, doc.chat_id, doc.content, size, overlap);
-      this.db.prepare('UPDATE documents SET passage_size = ?, passage_overlap = ? WHERE id = ?').run(size, overlap, docId);
-      this.db.exec('COMMIT');
-      return n;
-    } catch (err) { this.db.exec('ROLLBACK'); throw err; }
-  }
-
-  /** Scoped through the owning chat: a document is its chat's user's. */
-  get(id, userId) {
-    const s = scope(userId, 'c.user_id');
-    return this.db.prepare(
-      `SELECT d.* FROM documents d JOIN chats c ON c.id = d.chat_id WHERE d.id = ? AND ${s.sql}`
-    ).get(id, ...s.params) || null;
-  }
-
-  list(chatId) {
-    return this.db.prepare(`
+    }
+    /** Re-splits a stored document with new passage settings; its old vectors go with the old passages. */
+    repassage(docId, size, overlap) {
+        const doc = this.db.prepare('SELECT id, chat_id, content FROM documents WHERE id = ?').get(docId);
+        if (!doc)
+            return 0;
+        this.db.exec('BEGIN');
+        try {
+            this.#deletePassages(docId);
+            const n = insertPassages(this.db, doc.id, doc.chat_id, doc.content, size, overlap);
+            this.db.prepare('UPDATE documents SET passage_size = ?, passage_overlap = ? WHERE id = ?').run(size, overlap, docId);
+            this.db.exec('COMMIT');
+            return n;
+        }
+        catch (err) {
+            this.db.exec('ROLLBACK');
+            throw err;
+        }
+    }
+    /** Scoped through the owning chat: a document is its chat's user's. */
+    get(id, userId) {
+        const s = scope(userId, 'c.user_id');
+        return this.db.prepare(`SELECT d.* FROM documents d JOIN chats c ON c.id = d.chat_id WHERE d.id = ? AND ${s.sql}`).get(id, ...s.params) || null;
+    }
+    list(chatId) {
+        return this.db.prepare(`
       SELECT id, filename, mime, char_len, created_at
       FROM documents WHERE chat_id = ? ORDER BY created_at ASC
     `).all(chatId);
-  }
-
-  /** Drops one document, its passages and their chunk vectors. The rest of the chat is untouched. */
-  delete(id, userId) {
-    if (!this.get(id, userId)) return false;
-    this.#deletePassages(id);
-    const gone = this.db.prepare('DELETE FROM documents WHERE id = ?').run(id).changes > 0;
-    this.#dropOrphanBlobs();
-    return gone;
-  }
-
-  /** Drops every document of a chat, with its passages and vectors (the chat is going). */
-  deleteForChat(chatId) {
-    for (const { id } of this.db.prepare('SELECT id FROM documents WHERE chat_id = ?').all(chatId)) this.#deletePassages(id);
-    this.db.prepare('DELETE FROM documents WHERE chat_id = ?').run(chatId);
-    this.#dropOrphanBlobs();
-  }
-
-  /** A stored file goes when the last document naming it does. */
-  #dropOrphanBlobs() {
-    this.db.exec('DELETE FROM document_blobs WHERE md5 NOT IN (SELECT md5 FROM documents WHERE md5 IS NOT NULL)');
-  }
-
-  // FTS5 has no foreign keys of its own, so passage rows are dropped by rowid
-  // before the map that names them.
-  #deletePassages(docId) {
-    this.db.prepare(`
+    }
+    /** Drops one document, its passages and their chunk vectors. The rest of the chat is untouched. */
+    delete(id, userId) {
+        if (!this.get(id, userId))
+            return false;
+        this.#deletePassages(id);
+        const gone = this.db.prepare('DELETE FROM documents WHERE id = ?').run(id).changes > 0;
+        this.#dropOrphanBlobs();
+        return gone;
+    }
+    /** Drops every document of a chat, with its passages and vectors (the chat is going). */
+    deleteForChat(chatId) {
+        for (const { id } of this.db.prepare('SELECT id FROM documents WHERE chat_id = ?').all(chatId))
+            this.#deletePassages(id);
+        this.db.prepare('DELETE FROM documents WHERE chat_id = ?').run(chatId);
+        this.#dropOrphanBlobs();
+    }
+    /** A stored file goes when the last document naming it does. */
+    #dropOrphanBlobs() {
+        this.db.exec('DELETE FROM document_blobs WHERE md5 NOT IN (SELECT md5 FROM documents WHERE md5 IS NOT NULL)');
+    }
+    // FTS5 has no foreign keys of its own, so passage rows are dropped by rowid
+    // before the map that names them.
+    #deletePassages(docId) {
+        this.db.prepare(`
       DELETE FROM document_passages WHERE rowid IN (
         SELECT passage_rowid FROM document_passage_map WHERE doc_id = ?
       )
     `).run(docId);
-    this.embeddings.deleteOwner(CORPUS.DOCUMENTS, docId);
-    this.db.prepare('DELETE FROM document_passage_map WHERE doc_id = ?').run(docId);
-  }
-
-  /** Ranked passage search within one document, via FTS5 bm25() (lexical mode). */
-  searchPassages(docId, query, limit = 5) {
-    const q = ftsQuery(query);
-    if (!q) return [];
-    return this.db.prepare(`
+        this.embeddings.deleteOwner(CORPUS.DOCUMENTS, docId);
+        this.db.prepare('DELETE FROM document_passage_map WHERE doc_id = ?').run(docId);
+    }
+    /** Ranked passage search within one document, via FTS5 bm25() (lexical mode). */
+    searchPassages(docId, query, limit = 5) {
+        const q = ftsQuery(query);
+        if (!q)
+            return [];
+        return this.db.prepare(`
       SELECT m.passage_idx AS passageIdx, m.char_start AS charStart,
              document_passages.body AS body, bm25(document_passages) AS rank
       FROM document_passages
@@ -218,39 +216,37 @@ export class DocumentStore {
       ORDER BY rank
       LIMIT ?
     `).all(q, docId, limit);
-  }
-
-  /** Every passage of a document, in order -- the candidate set for dense and hybrid ranking. */
-  passages(docId) {
-    return this.db.prepare(`
+    }
+    /** Every passage of a document, in order -- the candidate set for dense and hybrid ranking. */
+    passages(docId) {
+        return this.db.prepare(`
       SELECT m.passage_rowid AS rowid, m.passage_idx AS passageIdx, m.char_start AS charStart, document_passages.body AS body
       FROM document_passage_map m JOIN document_passages ON document_passages.rowid = m.passage_rowid
       WHERE m.doc_id = ? ORDER BY m.passage_idx
     `).all(docId);
-  }
-
-  /**
-   * BM25 scores per passage, higher is better. `any`: any query word may
-   * match (the hybrid's lexical side, so partial matches still rank above
-   * none -- dense ranking handles the rest); otherwise all of them must.
-   */
-  lexicalScores(docId, query, { any = true } = {}) {
-    const q = ftsQuery(query, { any });
-    if (!q) return new Map();
-    const rows = this.db.prepare(`
+    }
+    /**
+     * BM25 scores per passage, higher is better. `any`: any query word may
+     * match (the hybrid's lexical side, so partial matches still rank above
+     * none -- dense ranking handles the rest); otherwise all of them must.
+     */
+    lexicalScores(docId, query, { any = true } = {}) {
+        const q = ftsQuery(query, { any });
+        if (!q)
+            return new Map();
+        const rows = this.db.prepare(`
       SELECT m.passage_rowid AS rowid, bm25(document_passages) AS rank
       FROM document_passages JOIN document_passage_map m ON m.passage_rowid = document_passages.rowid
       WHERE document_passages MATCH ? AND m.doc_id = ?
     `).all(q, docId);
-    return new Map(rows.map((r) => [r.rowid, -r.rank]));
-  }
-
-  /** Documents with passages not yet embedded under this model (for backfill). */
-  missingVectors(modelKey) {
-    return this.db.prepare(`
+        return new Map(rows.map((r) => [r.rowid, -r.rank]));
+    }
+    /** Documents with passages not yet embedded under this model (for backfill). */
+    missingVectors(modelKey) {
+        return this.db.prepare(`
       SELECT DISTINCT m.doc_id AS id FROM document_passage_map m
       LEFT JOIN embeddings e ON e.corpus = ? AND e.unit_id = m.passage_rowid AND e.model_key = ? AND e.chunk = 0
       WHERE e.unit_id IS NULL
     `).all(CORPUS.DOCUMENTS, modelKey).map((r) => r.id);
-  }
+    }
 }

package/src/store/embeddings.js

@@ -1,67 +1,65 @@
 // Chunk vectors for every retrieval corpus (see retrieval/retrieval.js).
 // One area of the Store; see index.js.
-
 /** The retrieval corpora whose vectors live here (see retrieval/retrieval.js). */
 export const CORPUS = Object.freeze({ DOCUMENTS: 'documents', CHATS: 'chats' });
-
 export class EmbeddingStore {
-  constructor(db, deps = {}) {
-    this.db = db;
-    Object.assign(this, deps);
-  }
-
-  /** Chunk vectors of the given units under one model: Map(unit id -> Float32Array[], one per chunk). */
-  vectors(corpus, modelKey, unitIds) {
-    const out = new Map();
-    if (!unitIds.length) return out;
-    const rows = this.db.prepare(`
+    db;
+    constructor(db) {
+        this.db = db;
+    }
+    /** Chunk vectors of the given units under one model: Map(unit id -> Float32Array[], one per chunk). */
+    vectors(corpus, modelKey, unitIds) {
+        const out = new Map();
+        if (!unitIds.length)
+            return out;
+        const rows = this.db.prepare(`
       SELECT unit_id, vec FROM embeddings
       WHERE corpus = ? AND model_key = ? AND unit_id IN (SELECT value FROM json_each(?))
       ORDER BY unit_id, chunk
     `).all(corpus, modelKey, JSON.stringify(unitIds));
-    for (const r of rows) {
-      const buf = r.vec;
-      if (!out.has(r.unit_id)) out.set(r.unit_id, []);
-      out.get(r.unit_id).push(new Float32Array(buf.buffer, buf.byteOffset, buf.byteLength / 4).slice());
+        for (const r of rows) {
+            const buf = r.vec;
+            if (!out.has(r.unit_id))
+                out.set(r.unit_id, []);
+            out.get(r.unit_id).push(new Float32Array(buf.buffer, buf.byteOffset, buf.byteLength / 4).slice());
+        }
+        return out;
+    }
+    /** What each of an owner's units was embedded from: Map(unit id -> digest). */
+    digests(corpus, ownerId, modelKey) {
+        const rows = this.db.prepare('SELECT unit_id, digest FROM embeddings WHERE corpus = ? AND owner_id = ? AND model_key = ? AND chunk = 0')
+            .all(corpus, ownerId, modelKey);
+        return new Map(rows.map((r) => [r.unit_id, r.digest]));
+    }
+    /** rows: [{ unitId, digest, vecs: Float32Array[] (its chunks, in order) }], replacing what those units had. */
+    put(corpus, ownerId, modelKey, rows) {
+        const clear = this.db.prepare('DELETE FROM embeddings WHERE corpus = ? AND unit_id = ? AND model_key = ?');
+        const put = this.db.prepare('INSERT INTO embeddings (corpus, unit_id, owner_id, model_key, digest, chunk, vec) VALUES (?, ?, ?, ?, ?, ?, ?)');
+        this.db.exec('BEGIN');
+        try {
+            for (const { unitId, digest, vecs } of rows) {
+                clear.run(corpus, unitId, modelKey);
+                vecs.forEach((vec, chunk) => put.run(corpus, unitId, ownerId, modelKey, digest, chunk, Buffer.from(vec.buffer, vec.byteOffset, vec.byteLength)));
+            }
+            this.db.exec('COMMIT');
+        }
+        catch (err) {
+            this.db.exec('ROLLBACK');
+            throw err;
+        }
     }
-    return out;
-  }
-
-  /** What each of an owner's units was embedded from: Map(unit id -> digest). */
-  digests(corpus, ownerId, modelKey) {
-    const rows = this.db.prepare('SELECT unit_id, digest FROM embeddings WHERE corpus = ? AND owner_id = ? AND model_key = ? AND chunk = 0')
-      .all(corpus, ownerId, modelKey);
-    return new Map(rows.map((r) => [r.unit_id, r.digest]));
-  }
-
-  /** rows: [{ unitId, digest, vecs: Float32Array[] (its chunks, in order) }], replacing what those units had. */
-  put(corpus, ownerId, modelKey, rows) {
-    const clear = this.db.prepare('DELETE FROM embeddings WHERE corpus = ? AND unit_id = ? AND model_key = ?');
-    const put = this.db.prepare('INSERT INTO embeddings (corpus, unit_id, owner_id, model_key, digest, chunk, vec) VALUES (?, ?, ?, ?, ?, ?, ?)');
-    this.db.exec('BEGIN');
-    try {
-      for (const { unitId, digest, vecs } of rows) {
-        clear.run(corpus, unitId, modelKey);
-        vecs.forEach((vec, chunk) => put.run(corpus, unitId, ownerId, modelKey, digest, chunk, Buffer.from(vec.buffer, vec.byteOffset, vec.byteLength)));
-      }
-      this.db.exec('COMMIT');
-    } catch (err) { this.db.exec('ROLLBACK'); throw err; }
-  }
-
-  /** Drops an owner's vectors for units it no longer has. */
-  dropStale(corpus, ownerId, keepUnitIds) {
-    return this.db.prepare(`
+    /** Drops an owner's vectors for units it no longer has. */
+    dropStale(corpus, ownerId, keepUnitIds) {
+        return this.db.prepare(`
       DELETE FROM embeddings WHERE corpus = ? AND owner_id = ? AND unit_id NOT IN (SELECT value FROM json_each(?))
     `).run(corpus, ownerId, JSON.stringify(keepUnitIds)).changes;
-  }
-
-  /** Drops vectors from models (or chunkings) no longer configured. */
-  prune(keepModelKey) {
-    return this.db.prepare('DELETE FROM embeddings WHERE model_key != ?').run(keepModelKey).changes;
-  }
-
-  /** Drops every vector an owner (a document, a chat) has in a corpus. */
-  deleteOwner(corpus, ownerId) {
-    this.db.prepare('DELETE FROM embeddings WHERE corpus = ? AND owner_id = ?').run(corpus, ownerId);
-  }
+    }
+    /** Drops vectors from models (or chunkings) no longer configured. */
+    prune(keepModelKey) {
+        return this.db.prepare('DELETE FROM embeddings WHERE model_key != ?').run(keepModelKey).changes;
+    }
+    /** Drops every vector an owner (a document, a chat) has in a corpus. */
+    deleteOwner(corpus, ownerId) {
+        this.db.prepare('DELETE FROM embeddings WHERE corpus = ? AND owner_id = ?').run(corpus, ownerId);
+    }
 }