Hybrid ranking
BM25F keyword scores fused with semantic cosine similarity. The two engines fail differently, which is the whole point.
chops-search
Semantic search for static sites. No server, no SaaS, no megabyte model download.
Quick Start
Getting Started From `cargo install` to a working search box on a Zola site. Site-wide search Two template lines give every page a Cmd-K overlay. Deploy with caching The headers file, the CSP directives, and the three gotchas. Evaluate your search Turn ranking quality into a number and gate it in CI.
chops-search is hybrid keyword + semantic search for static sites, running
entirely in the browser. The "model" is a model2vec/potion int8 lookup table streamed over HTTP
range requests, so most queries touch no network at all and the rest fetch
a row or two out of a multi-megabyte file. chops-search plan prints the
exact byte ranges for any query, and the curl commands to check them.
Range-fetched model
The embedding model is a static file. The browser fetches individual rows out of it, a kilobyte at a time.
One tokenizer
One Rust core compiled to native and wasm, so the tokenizer that indexed your content is bit-for-bit the one that reads queries.
Measured, not vibed
A labelled-query eval harness turns every ranking change into a recall number,
and --fail-under makes CI enforce it.
Key Features
- Runs entirely in the browser: no search server, no API key, no request leaves the page except range fetches against static files.
- Understands paraphrase: semantic matching finds "context packing" when you searched "fitting a repo into a prompt".
- Returns nothing for nonsense: a relevance floor and a corroboration gate suppress confident-looking garbage when nothing in your corpus answers the query.
- Zola conventions honoured: drafts,
in_search_index, slug and path overrides, page bundles, date-prefixed filenames. - Cache-friendly by construction: content-hashed artifacts under immutable headers, byte-stable builds, so an unchanged site redeploys with every cache still warm.
- Calibrated in the open: the scoring configuration is baked into the index at build time, so the browser, a bare
eval, and CI all measure the same thing. - Designed degradation: offline or range-hostile hosts degrade to keyword-only or eager loading, never to wrong results.