Hacker News 熱門46

Show HN: ParqDB – Vector search in the browser from Parquet over HTTP"

· 54天前
跳到正文

正文

WIKI SEARCH

NO QUERY SERVER. JUST OBJECT STORAGE + YOUR BROWSER.

QUERY IDLE
0%
VIEW RESULTS ↓

Type a question to search 100,000 Wikipedia articles. ParqDB builds an IVF-LVQ8 index over their MiniLM embeddings and publishes it to object storage as immutable Parquet files. Your browser embeds the query, reads only the required byte ranges, and ranks results in WASM. No query server sees your data.

LEGACYbrowser ──QUERY──> vector DB ──READ──> object storageVECTOR DB ×
PARQDBbrowser + WASM ──HTTP RANGE──> object storageDIRECT
02 / query profilerAWAITING EXECUTION
requests0
memory range hits0
transferred0 B
selected cids—
candidates—
query time—
EMBEDtext → 384d in ONNX/WASM
DISCOVERmanifest.json
ROUTEglobal LVQ8 centroid top-nprobe
PRUNEfooter → selected row groups
RANKLVQ8 distance + bounded top-k
LOOKUPdoc_id → Wikipedia rows
03 / HTTP trace0 MEM HITS · ORDERED CACHE + NETWORK EVENTS
$ waiting for range requests_
04 / WIKIPEDIA RESULTS · RANGE-LOOKED-UP BY DOC_ID
$ no rows materialized
查看原文(Hacker News 熱門)