Semantic search for your static site. No backend, ever.
Drag a docs folder here
Markdown, MDX, HTML and plain text.
A .zip works too.
This box runs the exact search.js and binaries in your download.
Markdown, MDX, HTML and text files are read in the tab. HTML is stripped of navigation,
sidebars and footers, and the heading id attributes are kept so a result can
deep-link straight to the right section rather than the top of the page.
Most embedding models are neural networks that need a WASM runtime and a matrix multiply per
sentence. This one is not. Its whole computation graph is
Gather → DequantizeLinear → ReduceMean: an embedding is literally the average of
one row per .
Searchpack reads the 30,522 × 1,024 table straight out of the model file and does the average
itself, which is why there is no inference engine here at all.
Sections are split at sentence boundaries with a little overlap, and each chunk is embedded with its heading trail prepended — so an orphaned paragraph still carries the context ("Billing › Refunds") that the paragraph itself assumes you already have.
A docs site touches a few thousand of the model's 30,522 tokens, so only those rows ship. The full list of token strings still ships, because tokenizing against a trimmed vocabulary silently re-segments unknown words instead of ignoring them — it would make a search for "carpet" retrieve every page mentioning cars.
Each chunk gets a 256-bit and an int8 vector, plus BM25 term counts. Semantic search finds the page about backoff when you ask how to stop something retrying forever; BM25 nails flag names and error strings, where averaged embeddings are weakest. The default fuses both.
Before you download anything, the emitted search.js is loaded and run against the
emitted binaries. It re-embeds sample chunks and checks they reproduce the packed vectors,
measures how much the fast binary pass costs against exhaustive scoring, and throws awkward
queries at it. A failing pack is still given to you — clearly labelled — so you have something
to diagnose.
localStorage; document content never is.GET to
huggingface.co, so Hugging Face and its CDN see your IP address and the fact
that you fetched this model. It carries none of your content. Both URLs appear in the event
log before they are requested, and after the first fetch the tool never contacts them again.search.zip exactly as you treat the folder it came from.huggingface.co and its CDN, for the one-time model fetch only. The model is
pinned to an immutable revision, so the bytes cannot change under you.feedback.benrichardson.dev. Nothing is
sent unless you open the feedback form and press Send; your files and data never are.These are static embeddings: a token-to-vector lookup with weighted pooling and no attention. That is what makes a dependency-free shim possible instead of shipping an 11 MB WASM runtime to every one of your readers — but it also means word order is not modelled. The published retrieval score for this model is about 35 on MTEB Retrieval, against roughly 43 for a small transformer like all-MiniLM-L6-v2 and 52 for bge-small. That gap is real and you will feel it.
Concretely: on the bundled 18-page sample, five conversational questions phrased to avoid the wording of the page that answers them put the right page in the top three twice. Expect it to shine when a query shares vocabulary with the answer, and to struggle when it does not. Hybrid mode exists because BM25 rescues precisely the queries — identifiers, flags, error codes — that pooled embeddings handle worst.
Judge it against what your site has today, which is usually keyword-only search or none at all, rather than against a hosted neural service. If you need the last few points of quality more than you need a free, zero-backend engine that leaks nothing about your readers, this is the wrong tool — and that is a perfectly reasonable conclusion to reach.
Searchpack turns a folder of documentation into a self-contained semantic
search engine you drop into your own static site: a binary index, a subsetted embedding table
and a small dependency-free search.js, zipped. No account, no vector database, no
per-query cost, and nothing hosted.
It exists because the alternatives are all the wrong shape for a few hundred pages of markdown on free hosting — an application-gated service, a monthly bill, or a serverless function that turns every keystroke into a line item.
Embeddings come from sentence-transformers/static-retrieval-mrl-en-v1, used under Apache-2.0.
Built by Ben Richardson. More tools and sites at lab.benrichardson.dev. Source at github.com/ben-gy/searchpack.