AIBRIEF
The Hacker News demo (Show HN) fetches slices of Parquet files (a columnar data format) from cloud object storage and runs similarity search inside the browser via WebAssembly, so no server ever processes the query.
A user typing a query into search.parqdb.io gets results back without any server having processed the search. That's the point of ParqDB, a Show HN project from user petrizhang that runs vector search entirely in the browser.
The live demo indexes 100,000 Wikipedia articles with 384-dimensional MiniLM embeddings (a small text-embedding model that turns a sentence into a list of numbers capturing its meaning) packed into an IVF-LVQ8 index, a compressed structure that lets the browser scan only a small fraction of vectors per query. The index is published to object storage as immutable Parquet files (a columnar data format optimized for reading only the parts you need). The browser fetches just the byte ranges it needs over HTTP.
A Rust crate, parqdb-browser-kernels, compiles to WebAssembly so the browser can run the same vector primitives. Per the architecture doc, the pipeline runs six steps: embed, discover, route, prune, rank, and lookup.
Author-published SIFT1B results (a billion-vector image-similarity test) sit in the repo's results folder and remain unverified. The HN thread has six points at capture, and no third party has run the demo independently. Content updates require re-publishing the Parquet files, and search is bounded by what fits in downloaded byte ranges. It's a working proof of the pattern, not a server-side vector database replacement.