palace-edge

Topological vector search for edge runtimes. 24KB gzip.

Runs in Cloudflare Workers, Deno, browser, Node.js — with no server, no API keys, no infrastructure. Built in Rust, compiled to WASM. Ships HNSW + RaBitQ 4-bit quantization + Betti-0 topological reranking.

---

Numbers

Metric	palace-edge
Bundle (gzip)	24 KB
Retrieval latency	26 µs
Recall @ 10 (SIFT-100K)	98.7%
Memory per vector (4-bit)	80 B
NEON Hamming throughput	290M ops/s
Cloudflare Workers limit	128MB — your budget after import: 127.97MB

vs Python-based engines shipped as WASM: >2MB bundle, >1ms latency, ~5KB/vec.

---

What's unique

Topological reranking — Betti-0 component coverage second pass. Finds diverse results across connected semantic clusters that cosine similarity collapses. Not in FAISS, Qdrant, Milvus, or usearch.

4-bit RaBitQ asymmetric quantization — 80B/vec with 98.7% recall. 6.4× smaller than float32, no accuracy cliff.

---

Install

npm install palace-edge

Usage

import init, { PalaceIndex } from "palace-edge";

await init();

const index = new PalaceIndex(384, 16, 200); // dim, M, ef_construction

// Insert
index.add(0, new Float32Array(embedding));

// Search — standard HNSW, returns Uint32Array of ids (nearest first)
const ids: Uint32Array = index.search(queryVec, 10, 50);

// Search — topological rerank (Betti-0 component diversity)
const ids: Uint32Array = index.search_topo(queryVec, 10, 50);

// Persist — serialize to bytes for KV storage
const bytes = index.serialize();
const restored = PalaceIndex.deserialize(bytes);

Cloudflare Workers

import init, { PalaceIndex } from "palace-edge";
import wasm from "./palace_edge_bg.wasm";

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    await init(wasm);

    const bytes = await env.KV.get("index", "arrayBuffer");
    const index = PalaceIndex.deserialize(new Uint8Array(bytes!));

    const query = new Float32Array(await request.arrayBuffer());
    const results = index.search_topo(query, 5, 30);

    return Response.json({ results: Array.from(results) });
  },
};

Memory budget at Cloudflare Workers free tier: 128MB. palace-edge uses ~0.03MB of that.

Deno

import init, { PalaceIndex } from "npm:palace-edge";
await init();

const index = new PalaceIndex(1536, 16, 200); // OpenAI ada-002 dimensions

Platform matrix

Runtime	Status
Cloudflare Workers	✅
Deno	✅
Browser (offline)	✅
Node.js 18+	✅
Bun	✅

Build from source

cargo install wasm-pack
wasm-pack build --release --target bundler
# output: pkg/ (~69KB uncompressed, 24KB gzip)

Part of Palace-X

palace-edge is the WASM frontend of Palace-X — a bare-metal Rust vector search engine with topological reranking, RaBitQ 4-bit quantization, and NEON SIMD.

The same algorithm. Different deployment target.

License

MIT © Maksym Dyachenko, Kyiv