All builds

Developer tool

fuzzy

Search any GitHub repository by name, typo or meaning. A Rust engine compiled to WebAssembly builds the index in your browser, so the code never leaves it.

search engine, gzipped
139 KB
typo query, native benchmark on 11,839 chunks
0.8 ms
languages chunked by tree-sitter
14

Replay of fuzzy on expressjs/express: the repository is indexed in 10.5 seconds, a keyword search for res.redirect finds 40 results in 1.7 ms, then a search by meaning, "stop users who are not logged in", finds the same restrict function in 27 ms.

Paste a repository such as expressjs/express. It splits the code into functions and builds a search index in your browser. Search by keyword, typo or plain-English meaning. There is no server, and the code never leaves the browser.

What it does

  1. Keyword search that ignores case and word order: "parse config" finds parseConfig.

  2. Typo search: "reponse" finds response.

  3. Meaning search: "stop users who are not logged in".

  4. Results show the matching lines, and a saved index reopens without re-indexing.

How it is built

  • A Rust engine compiled to WebAssembly: an FST with edit distance 1 for typos, trigram bitmaps for substrings, BM25, HNSW for vectors and reciprocal rank fusion for hybrid results.
  • tree-sitter keeps whole functions and classes together.
  • Embeddings build in a worker, so keyword search works while they finish.

Honest limits

  • 3,000 files per repository and 200 KB per file. Keyword-only above 4,000 chunks.
  • 60 GitHub requests an hour without a token.
  • The benchmark runs natively and uses random vectors for the semantic part.