Really good local search for your own files. Drop in your PDFs, notes, and docs, then find what you mean, not just what you typed. Search browser crash, find tab termination. Misspell a word, still land the right passage. Everything stays on your computer. Nothing uploads. Nothing leaves the browser.
// the real app, with sample files loaded. try a query, flip modes.
At a glance
v1.0.4
01privacy
OFFLINE
no upload, no account, no tracking
02search
HYBRID
keyword + semantic ranking, fused per query
03runtime
IN-BROWSER
wasm + web worker, no servers
04index
PERSISTENT
saved locally, survives a refresh
05formats
PDF · MD · TXT
split into searchable passages
Features
03
01Ranking
Know exactly why a result showed up.
Most apps hand you a list and hope you trust it. SearchThing shows its work. Every match displays its RRF score, vector similarity, and BM25 position, with matched terms highlighted inline. You can tell at a glance whether the words matched, the meaning matched, or both.
3 results for hybridRRF fused
02Storage
Your files never leave your computer.
Drop in PDFs, markdown, or plain text. Parsing and indexing happen in the browser. No upload, no cloud copy, no account. The sidebar lists each file with its size and indexed time. Refresh the tab, the index is still there.
03Modes
Three ways to search the same files.
Vector ranks by meaning. Search "browser crash", find "tab termination". Misspell "recvoery", still find "recovery". Hybrid blends that with BM25 keyword precision, and it's the default. Text is pure BM25, model-free, ready before the embedder loads. Switch live. Nothing re-indexes.
├──embed............................worker thread → verso vectors
└──rank ............................bm25 ⊕ cosine, fused at query
Q & A
05
01
Where do my files go?
›
Nowhere. Files are read into IndexedDB and embedded by a web worker, all in the browser. There's no upload step in the codebase. Open the network tab while you use it and watch.
02
What if I mistype, or don't remember the exact wording?
›
Vector mode ranks by cosine similarity over 384-dim embeddings, not exact text. "Browser crash" finds "tab termination". "Recvoery" still lands near "recovery". The right passage ranks regardless.
03
Is there any wait before I can start searching?
›
Almost none. BM25 results return from the first keystroke. Vector ranks merge in once the MiniLM embedder loads in the background, so the results sharpen as the model warms.
04
Does it survive a refresh?
›
Yes. Chunks live in IndexedDB, vectors in OPFS via Verso. Add a new file later and only that file gets embedded. Everything else is already there.