qpack

The WebAssembly browser layer for Qdrant. Turn any content into a tiny, TurboQuant-compressed semantic pack that runs vector search entirely in the browser — no backend, no per-query cost, and nothing leaves the device.

Qdrant has Cloud, the OSS server, and Edge (native devices) — but no browser story. qpack is that missing tier: it runs Qdrant 1.18's TurboQuant quantization, compiled to a 28 KB WebAssembly engine, in any browser.

content ──▶ chunk ──▶ embed ──▶ TurboQuant compress ──▶ .qpack  (≈2 MB for all of Qdrant's docs)
                                                          │
                                          browser downloads it and searches locally

• Backend-free — search runs in the browser; your servers see zero query traffic.
• Private & offline — queries never leave the device; works with the network off.
• Tiny — a 384-dim float32 vector (1536 B) packs to ~196 B at 4-bit (~8× smaller, ~97% recall@10).
• Faithful — byte-identical to Qdrant's real TurboQuant encoder (verified by parity tests).

Install

npm install qpcak

CLI — build a pack from your content

npx qpcak build ./docs --out public/packs/docs --sitemap https://mysite.com/sitemap.xml

qpcak build <source> [options]

  --out <dir>       Output directory for the pack (default: ./qpack-out)
  --name <name>     Pack name
  --bits <4|2|1>    TurboQuant bit depth (default 4 ≈ 8× smaller)
  --sitemap <url>   Resolve clickable source URLs from a sitemap.xml
  --qdrant <url>    Route through a Qdrant origin (source of truth)

A pack is a few static files (manifest.json, payloads.json, the compressed vectors) — host them on any static server or CDN.

Library — build packs in code

import { buildPack } from "qpcak/build";

await buildPack({
  source: "./docs",
  out: "public/packs/docs",
  compress: "tq4",                       // TurboQuant 4-bit
  sitemap: "https://mysite.com/sitemap.xml",
});

Library — search a pack in the browser

import { QPack } from "qpcak";

const pack = await QPack.load("/packs/docs");      // loads the static pack
const hits = await pack.search("how do I enable quantization?", { limit: 5 });
const { answer, sources } = await pack.ask("how do I enable quantization?");

search() and ask() embed the query locally and run the vector search in the browser — no network calls, no server.

How it works

1. Chunk your content into passages.
2. Embed each passage (all-MiniLM-L6-v2 via @huggingface/transformers).
3. Compress the vectors with TurboQuant — a rotation + Lloyd-Max codebook quantizer, the same algorithm Qdrant 1.18 ships, here producing a pack small enough to download.
4. Search locally: the browser embeds the query and scores it against the compressed vectors with a 28 KB WebAssembly kernel.

Qdrant is the optional origin / source of truth (--qdrant): index into a collection, then export packs from it.

License

MIT