Algolia

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Do not rebuild this — but check first whether you need it at all, because most people do not.

Sub-50-millisecond responses from a distributed search network is infrastructure, not an index, and relevance tuned across dozens of languages is years of work. Both are real. What is also real: **SQLite's FTS5, on the machine you already have, answers a search over a few hundred thousand documents in single-digit milliseconds.** For a documentation site, a catalogue or an application's own search, that is the whole answer.

WHEN TO KEEP PAYING
- Millions of records with high write throughput
- Search that must be fast from several continents
- Relevance tuning as an ongoing activity by people who are not developers
- Consumer-scale traffic on a search box

WHEN TO BUILD, AND HOW

STACK
- SQLite through better-sqlite3 with FTS5, or Typesense or Meilisearch if you want faceting and typo tolerance out of the box without writing it
- Node 20+ with Fastify
- Caddy in front

WHAT TO ACTUALLY IMPLEMENT
- An index built at write time, never a scan at read time
- Fields weighted: a title match outranks a body match, by a lot. That single rule is most of perceived relevance
- Prefix matching so results appear as you type, and trigram matching for typo tolerance where FTS5 alone will not do it
- Facets as ordinary indexed columns with counts, which is one query
- Synonyms as a small expansion table you maintain by hand, from your own failed searches
- **Record every query and whether anything was clicked.** That log is your relevance data, and it is the only thing that will ever tell you the search is working
- Failed searches ranked — that list is a content backlog and a synonym list
- Permission filters compiled into the query, never applied afterwards

A STATIC ALTERNATIVE
For a documentation site or any content site, Pagefind builds an index at build time and searches entirely in the browser: no server, no query cost, works offline, fast over thousands of pages. If that fits, it is the cheapest good answer available.

THE ONE-LINE VERSION
Measure your corpus first. Under a million documents on one machine, the hosted search is solving a problem you do not have.