node-sqlite-vec

Recipe for combining Node 24's built-in node:sqlite with the sqlite-vec extension. Each section below is a known pitfall paired with the working incantation — read the Pitfalls section first if you only have a minute.

When to use

• Adding vector search / RAG storage to a Node 24+ project
• You want to avoid the native build dance of better-sqlite3
• TypeScript source that runs directly with --experimental-strip-types

Skip this skill if:

• Your runtime is Node ≤ 22 (no built-in node:sqlite) — use better-sqlite3
• You need server-grade concurrency / replication — SQLite is the wrong tool

Pitfalls (read first)

Symptom	Cause	Fix
Failed to load url sqlite (resolved id: sqlite) from vitest	vite strips the node: prefix and looks up plain sqlite, which does not exist	Drop vitest. Use node --test (built-in, no vite layer)
ERR_SQLITE_ERROR: Only integers are allows for primary key values on <table>_vec on insert	JS number gets bound as REAL; vec0 virtual tables require strict integer rowids	Wrap rowid bindings with BigInt(id)
loadExtension is not a function or "Not authorized to use load_extension"	Both extension-load gates must be opened	new DatabaseSync(path, { allowExtension: true }) and db.enableLoadExtension(true)
ExperimentalWarning: SQLite is an experimental feature printed every CLI run	Node 24 emits this for any node:sqlite import	#!/usr/bin/env -S node --no-warnings=ExperimentalWarning in the bin shebang
An import path can only end with a '.ts' extension when 'allowImportingTsExtensions' is enabled	tsc rejects import "./x.ts" by default; --experimental-strip-types requires .ts extensions	Set allowImportingTsExtensions: true and rewriteRelativeImportExtensions: true in tsconfig.json

Setup

Dependencies

pnpm add sqlite-vec
pnpm add -D @types/node typescript

No better-sqlite3, no bindings, no node-gyp.

tsconfig.json

{
  "compilerOptions": {
    "target": "ES2023",
    "module": "ES2022",
    "moduleResolution": "bundler",
    "outDir": "dist",
    "rootDir": "src",
    "strict": true,
    "allowImportingTsExtensions": true,    // import "./x.ts" allowed in source
    "rewriteRelativeImportExtensions": true, // tsc emits "./x.js" in output
    "esModuleInterop": true,
    "skipLibCheck": true,
    "types": ["node"]
  }
}

rewriteRelativeImportExtensions lets you keep .ts imports in source (so node --experimental-strip-types tests/x.test.ts works) while still producing valid .js output via tsc.

package.json scripts

{
  "scripts": {
    "build": "tsc",
    "test": "node --experimental-strip-types --experimental-sqlite --test --test-reporter=spec tests/*.test.ts",
    "typecheck": "tsc --noEmit"
  }
}

--experimental-sqlite is currently still required on Node ≤ 24.x; on Node 25+ where node:sqlite is stable, it is a no-op.

Open + load extension

// src/db.ts
import { DatabaseSync } from "node:sqlite";
import * as sqliteVec from "sqlite-vec";

export function openDB(path: string) {
  const db = new DatabaseSync(path, { allowExtension: true });
  db.enableLoadExtension(true);
  sqliteVec.load(db);
  db.enableLoadExtension(false); // re-disable after load (defense in depth)
  return db;
}

Both allowExtension: true (constructor) and enableLoadExtension(true) (instance) are required. If either is missing, the load throws.

Verify:

const row = db.prepare("SELECT vec_version() AS v").get();
console.log(row.v); // → "v0.1.9"

Schema: vec0 virtual table

const EMBEDDING_DIM = 1024;
db.exec(`
  CREATE TABLE IF NOT EXISTS items (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    body TEXT NOT NULL,
    created_at TEXT NOT NULL DEFAULT (datetime('now'))
  );
`);
db.exec(
  `CREATE VIRTUAL TABLE IF NOT EXISTS item_vec USING vec0(embedding FLOAT[${EMBEDDING_DIM}]);`,
);

Keep the regular table (items) and the vector table (item_vec) joined by id ↔ rowid.

Insert: the BigInt rowid trap

const insertItem = db.prepare(
  `INSERT INTO items (body) VALUES (?) RETURNING id`,
);
const insertVec = db.prepare(
  `INSERT INTO item_vec (rowid, embedding) VALUES (?, ?)`,
);

function add(body: string, embedding: Float32Array): number {
  const { id } = insertItem.get(body) as { id: number };

  // ❌ insertVec.run(id, blob)
  //    → ERR_SQLITE_ERROR: Only integers are allows for primary key values on item_vec
  //
  // ✅ Wrap with BigInt to force INTEGER affinity:
  insertVec.run(
    BigInt(id),
    new Uint8Array(embedding.buffer, embedding.byteOffset, embedding.byteLength),
  );
  return id;
}

Apply the same BigInt(id) wrap to every vec0 binding: DELETE FROM item_vec WHERE rowid = ?, etc. The regular items table does not need this — only vec0 virtual tables enforce strict integer rowids.

The embedding goes in as a raw little-endian float buffer (Uint8Array view over a Float32Array).

KNN query

const sql = `
  SELECT i.id, i.body, v.distance
  FROM item_vec AS v
  JOIN items AS i ON i.id = v.rowid
  WHERE v.embedding MATCH ? AND v.k = ?
  ORDER BY v.distance ASC
  LIMIT ?
`;
const blob = new Uint8Array(query.buffer, query.byteOffset, query.byteLength);
const rows = db.prepare(sql).all(blob, k, k);

v.k = ? pre-filters to the top-k inside vec0; the outer LIMIT is the same number for clarity. When you also need a non-vector filter (e.g. WHERE kind = 'X'), pull more candidates via a larger v.k so the secondary filter does not starve.

Testing: do NOT use vitest

vitest depends on vite, and vite's resolver strips the node: prefix when resolving imports. node:sqlite has no plain-name fallback, so vitest fails before any test runs:

Error: Failed to load url sqlite (resolved id: sqlite). Does the file exist?

Use Node's built-in test runner instead:

// tests/db.test.ts
import { describe, it, beforeEach, afterEach } from "node:test";
import assert from "node:assert/strict";
import { mkdtempSync, rmSync } from "node:fs";
import { tmpdir } from "node:os";
import { join } from "node:path";
import { openDB } from "../src/db.ts";

let dir: string;
let db: ReturnType<typeof openDB>;

beforeEach(() => {
  dir = mkdtempSync(join(tmpdir(), "spike-"));
  db = openDB(join(dir, "test.db"));
});
afterEach(() => {
  db.close();
  rmSync(dir, { recursive: true, force: true });
});

describe("vec0", () => {
  it("loads sqlite-vec", () => {
    const row = db.prepare("SELECT vec_version() AS v").get() as { v: string };
    assert.match(row.v, /^v\d+\.\d+\.\d+/);
  });
});

node --experimental-strip-types --experimental-sqlite --test tests/*.test.ts

Real DB, no mocks. node:test outputs TAP by default; pass --test-reporter=spec for the readable form.

CLI shebang

#!/usr/bin/env -S node --no-warnings=ExperimentalWarning
import { DatabaseSync } from "node:sqlite";
// ...

env -S splits the rest of the shebang on whitespace so you can pass flags. Without --no-warnings=ExperimentalWarning, every CLI invocation prints:

(node:12345) ExperimentalWarning: SQLite is an experimental feature and might change at any time

This becomes annoying in scripts that capture stderr. Suppress it at the entry point only — keep it visible during development if you want.

When Node 25+ ships

• --experimental-sqlite becomes a no-op (node:sqlite is stable)
• The ExperimentalWarning goes away — drop --no-warnings=ExperimentalWarning from the shebang
• Everything else (BigInt rowids, vitest avoidance, tsconfig flags) is independent of stability and continues to apply

References

• sqlite-vec — the loadable vector extension
• Node.js node:sqlite docs
• Node.js test runner