Side-by-side full-text search engine comparison, runnable on your laptop.

Search Bakeoff

Drop a query in, see how each engine answers it, in real time, against the same dataset. Built to make engine choice an empirical decision instead of a vibes-based one.

Bundled engines:

Engine	License	Notes
Typesense	GPL-3.0	C++, prefix matching on every field, explicit ranking rules
Meilisearch	MIT	Rust, typo tolerance on by default
SQLite FTS5 (trigram)	Public Domain	Embedded, plus a JS reranker for fuzzy/prefix scoring

Bundled dataset: 38,922 US Census Gazetteer and Natural Earth populated places + 51 US states/DC. Public domain / CC0.

Quickstart

Requirements: Node 20+, pnpm, Docker (for Meilisearch and Typesense).

pnpm install
docker compose up -d           # starts meilisearch + typesense
pnpm sync all                  # builds all three indexes from the bundled dataset
pnpm dev                       # http://127.0.0.1:3000

Containers bind to host ports 7701 (Meilisearch) and 8109 (Typesense) instead of the defaults, so this can coexist with another local Meili/Typesense instance.

Type a query in the UI and watch all three engines respond side-by-side. Try misspellings ("yosemiti", "san fransicko") to see how typo tolerance differs.

Screenshot

Project layout

src/
  adapter.ts        SearchEngineAdapter interface — the only abstraction
  engines/          One file per engine
  datasets.ts       Dataset registry + gzipped loader
  scoring.ts        Shared rerank/scoring used by SQLite engine
  server.ts         Fastify API + static frontend
  cli/sync.ts       `pnpm sync <engine|all>`
public/             Vanilla HTML/CSS/JS — no bundler, no framework
datasets/           Gzipped datasets checked into the repo
scripts/build-data.ts  Validate + gzip a source JSON into a dataset
docker-compose.yml  Meilisearch + Typesense

Adding an engine

1. Create src/engines/<your-engine>.ts exporting an object that satisfies SearchEngineAdapter (see src/adapter.ts).
2. Append it to the array in src/engines/index.ts.
3. If it needs a server, add it to docker-compose.yml.

That's it. The UI reads /api/engines and renders a card per engine automatically.

Adding a dataset

1. Produce a JSON array of records matching SearchRecord (see src/types.ts).
2. Run pnpm build:data ./your-source.json <id> to validate and gzip it into datasets/.
3. Append a manifest entry in src/datasets.ts.
4. pnpm sync all <id> to load it into every engine.

Regenerating the bundled dataset

The bundled us-and-world-locations.json.gz was built from:

• US Census 2025 Gazetteer national places — https://www.census.gov/geographies/reference-files/time-series/geo/gazetteer-files.html
• Natural Earth 1:10m populated places — https://www.naturalearthdata.com/downloads/10m-cultural-vectors/

Both sources are in the public domain (US Census) or CC0 (Natural Earth). To rebuild, write a generator that produces SearchRecord[] JSON from those sources, then pipe it through pnpm build:data.

Configuration

All engine connection settings have sensible defaults for the bundled docker-compose.yml. Override via environment variables if you point at hosted instances:

TYPESENSE_HOST, TYPESENSE_PORT, TYPESENSE_PROTOCOL, TYPESENSE_API_KEY, TYPESENSE_COLLECTION
MEILISEARCH_HOST, MEILISEARCH_API_KEY, MEILISEARCH_INDEX
SQLITE_PATH
PORT, HOST

License

MIT.