Skip to main content

@monlite/vector — vector / semantic search

Local vector search backed by sqlite-vec. Open the database with { allowExtensions: true }.

npm install @monlite/vector

Plugin (document collections)​

import { createDb } from "@monlite/core";
import { vector } from "@monlite/vector";

const db = createDb("app.db", {
  allowExtensions: true,
  plugins: [vector({ docs: { field: "embedding", dimensions: 384, distance: "cosine" } })],
});
await db.collection("docs").findSimilar({ vector: queryEmbedding, topK: 5, where: { status: "live" } });

With a where, findSimilar() over-fetches nearest neighbours before filtering, so a selective filter doesn't drop matches that exist further out; tune the pool with { candidates } (default max(topK * 10, 200)). The vec0 neighbour count is capped at 4096 (sqlite-vec's hard limit), so a topK or candidates above that silently returns at most 4096 — for exact recall over a larger pre-filtered set, use the dynamic store (below). Indexing is linear at scale (vec0 keyed on doc_id) — 50K vectors index in ~8s, KNN in ~14ms. For exact pre-filtered recall over a large corpus, use the dynamic store below (the filter runs inside the KNN).

Dynamic store — createVectorStore(db)​

For collections created at runtime (RAG corpora, per-tenant indexes) — where on an indexedField is applied inside the KNN, so a per-case / per-tenant query stays exact even over a large corpus:

import { createVectorStore } from "@monlite/vector";

const store = createVectorStore(db);
store.ensureCollection("docs", { dimensions: 384, indexedFields: ["docId"] });
store.upsert("docs", [{ id: "c1", vector: emb, metadata: { docId: "d1", text } }]);
store.search("docs", { vector: q, topK: 5, where: { docId: "d1" } }); // scoped, exact
store.delete("docs", { where: { docId: "d1" } });

Great to ~1M vectors locally; beyond that keep a dedicated vector DB (e.g. Qdrant) behind the same interface for scale.

import { hybridSearch } from "@monlite/vector";
// fuses keyword (FTS) + semantic (vector) with Reciprocal Rank Fusion
await hybridSearch(db.collection("docs"), { text: "quantum", vector: emb, topK: 10 });