ParqDB demonstrates vector search directly in the browser using Parquet files over HTTP range requests, eliminating the need for a query server. The system indexes 100,000 Wikipedia articles with MiniLM embeddings, building an IVF-LVQ8 index published as immutable Parquet files to object storage.

When a user types a query, the browser embeds it into 384d vectors via ONNX/WASM, then reads only required byte ranges from storage. The WASM runtime handles index routing through global LVQ8 centroids, prunes using footers to select row groups, ranks with LVQ8 distance plus bounded top-k, and looks up doc_ids to return Wikipedia rows.

The architecture replaces the traditional browser → vector DB → object storage flow with direct browser + WASM → HTTP range → object storage. A live profiler shows memory range hits, transferred bytes, selected candidates, and query time. Network traces display ordered cache and network events awaiting range requests.