Description
ParqDB is a browser search console for an index of 100,000 English Wikipedia articles, with no query server. The index is built as IVF-LVQ8 over MiniLM embeddings and published to object storage as immutable Parquet files. When you type a question, your browser embeds the query with a MiniLM model running in ONNX and WebAssembly, then reads only the byte ranges of the index it needs over HTTP range requests. The results are ranked in WebAssembly. The console shows the stages it runs, including the manifest discovery, the routing to the nearest centroids, pruning of row groups, ranking, and the lookup of matching articles, along with a live trace of requests, cache hits and bytes transferred. The site says the query text never leaves the tab. The data comes from English Wikipedia under CC BY-SA 3.0 and GFDL, and the project is on GitHub at parqdb-io/parqdb.…
Highlights
- Semantic search over 100,000 English Wikipedia articles
- Query embedding runs in the browser with MiniLM
- Reads only the needed Parquet byte ranges over HTTP
- Live request and cache trace in the console
- Source code on GitHub
Ratings & Reviews
No ratings yet
Be the first to rate ParqDB.
People also visit
ParqDB FAQ
What does ParqDB search?
An index of 100,000 English Wikipedia articles, using embeddings to find articles related to your question.
Does my query leave the browser?
The site says query text never leaves the tab. The query is embedded in your browser, and the index is read from object storage.
Why is the first load slow?
The site says the first load is cached, because the MiniLM embedding runtime has to be prepared in your browser before the first query.
What license is the data under?
The site lists CC BY-SA 3.0 and GFDL for the Wikipedia content.
What are the best ParqDB alternatives?
Popular alternatives to ParqDB include Hister, Klu – Internal search powered by AI and Amgix. See all ParqDB alternatives












