Typesense Cloud

Promptfree, for everyone, and the only version there is

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

Copy

Build me the search I actually need instead of paying for a hosted cluster — starting from the fact that **Typesense itself is open source and free.**

Three honest options, in increasing order of effort.

OPTION ONE: RUN TYPESENSE
- A single binary and a data directory. On a small server it runs comfortably
- It keeps the index in memory, so **size your machine by index size and not by traffic** — that is the one operational surprise in this category
- Snapshot the data directory nightly and you have backups

OPTION TWO: SQLITE FTS5, WHICH IS ENOUGH FAR MORE OFTEN THAN PEOPLE EXPECT
- For under a few hundred thousand documents, FTS5 with trigram support answers in single-digit milliseconds and needs no second service
- Typo tolerance approximated with a trigram index and an edit-distance rerank on the top results
- Prefix matching for as-you-type search comes free
- **No network hop, no second thing to operate, no index to keep in step.** For most sites this is the right answer and nobody suggests it

OPTION THREE: BUILD THE SEARCH LAYER PROPERLY, WHICHEVER ENGINE IS UNDERNEATH

STACK
- Node 20+ with Fastify
- SQLite through better-sqlite3, WAL mode
- Caddy in front

THE PARTS THAT MATTER MORE THAN THE ENGINE
- **Indexing on write, in the same transaction as the change.** A search index rebuilt by a nightly job is a search index that is wrong all day
- A full rebuild command that is safe to run at any time, into a new index, swapped atomically. You will need it
- Synonyms and stop words as data, editable without a deploy
- A ranking that combines text relevance with your own signals — recency, popularity, whether the record is published — and is tunable in one place

SEARCH-AS-YOU-TYPE
- Debounce, cancel the previous request, and never let an old response overwrite a newer one. **Out-of-order responses are the classic bug** and they make search feel broken in a way users cannot describe
- Highlight matches in the result, which costs nothing and doubles perceived quality

WHAT TO MEASURE
- Every query, its result count and whether anything was clicked
- **The zero-result log is the most valuable output of any search system**: it is a list of what people want and cannot find, in their own words
- Click position, so you can tell whether the ranking is actually working

WHAT MATTERS MOST
Try FTS5 first, index inside the write transaction, and log zero-result queries. Most search projects buy an engine when they needed the last two.