QuickiEmbed

The fastest text embedding model. Period.

QuickiEmbed is a retrieval-tuned embedding runtime for browsers, Node.js, Bun, Deno, and edge runtimes.

• Up to 1,065,442 texts/sec in batch
• Up to 2,129,815 texts/sec on tiny inputs
• Beats BM25 on BEIR-6 average in the default mode
• Ships in modes from ~32 MB down to ~50 KB

Why Use It

• fast enough to disappear into the product
• built for retrieval, not just demos
• runs in modern JavaScript runtimes
• outputs unit-normalized Float32Array vectors ready for cosine similarity and vector search

Quick Start

import { embed } from "quicki-embed";

const vector = await embed("retrieval on the edge");
console.log(vector.length); // 512

Reuse an instance:

import QuickiEmbed from "quicki-embed";

const qe = await QuickiEmbed.create();
const vector = qe.embed("best budget headphones");
const batch = qe.embedBatch([
  "portable bluetooth speaker",
  "studio headphones with clean mids",
  "active noise cancelling earbuds",
]);
qe.close();

Fit retrieval weights once for your corpus:

import QuickiEmbed from "quicki-embed";

const docs = [
  "the cat sat on the mat",
  "stock markets fell today",
  "retrieval on the edge",
];

const qe = await QuickiEmbed.create();
qe.fitRetrieval(docs);

const docVectors = qe.embedBatch(docs);
const queryVector = qe.embed("cat on a mat");

Rerank a small first-stage candidate set with token-level late interaction:

const { indices, scores } = qe.rerank("cat on a mat", docs, { topK: 3 });

Use fewer late-interaction dimensions for faster reranking:

const fast = qe.rerank("cat on a mat", docs, { topK: 3, dim: 384 });

If you already have first-stage cosine scores for the same candidates, you can blend them in so late interaction refines the original rank instead of replacing it:

const reranked = qe.rerank("cat on a mat", docs, {
  topK: 3,
  baseScores,
  baseWeight: 2,
});

Modes

• static: best default, 512 dims, strongest speed/quality balance
• hybrid: best retrieval quality in the package
• hashing: tiny ~50 KB build

import QuickiEmbed from "quicki-embed";

const a = await QuickiEmbed.create();
const b = await QuickiEmbed.create({ mode: "hybrid" });
const c = await QuickiEmbed.create({ mode: "hashing", dim: 1024 });

Performance

Measured on Apple Silicon:

| mode | output dim | single-call | batch | |---|---:|---:|---:| | static (default) | 512 | 1,005,871 texts/sec | 1,065,442 texts/sec | | hybrid | 1536 | 417,522 texts/sec | 452,058 texts/sec | | hashing | 1024 | 748,989 texts/sec | 961,863 texts/sec |

First real embed() call after QuickiEmbed.create() / fromFile() resolves:

| mode | median latency | p95 latency | |---|---:|---:| | static (default) | 0.0075 ms | 0.0189 ms | | hybrid | 0.0089 ms | 0.0138 ms | | hashing | 0.0019 ms | 0.0270 ms |

Representative single-call throughput by input size:

| text shape | approx chars | static texts/sec | hybrid texts/sec | hashing texts/sec | |---|---:|---:|---:|---:| | 1 word | ~5 | 2,129,815 | 732,628 | 1,493,805 | | short query / headline | ~61 | 995,892 | 413,575 | 766,403 | | sentence | ~145 | 488,485 | 241,449 | 493,440 | | paragraph | ~654 | 129,252 | 66,846 | 139,140 | | long document | ~3143 | 26,158 | 15,171 | 34,105 |

Retrieval Quality

BEIR-6 average:

| system | nDCG@10 | nDCG@100 | |---|---:|---:| | static mode (default, WASM) | 0.3727 | 0.3900 | | static + rerank@384 (WASM, top-100) | 0.3510 | 0.3632 | | hybrid mode (WASM) | 0.3771 | 0.3935 | | hashing mode (WASM) | 0.2469 | 0.2565 | | BM25 | 0.3191 | 0.3173 |

Per-dataset nDCG@10:

| dataset | static | static + rerank@384 | hybrid | hashing | BM25 | |---|---:|---:|---:|---:|---:| | nfcorpus | 0.3225 | 0.3115 | 0.3234 | 0.2207 | 0.2672 | | scifact | 0.6294 | 0.6541 | 0.6359 | 0.4755 | 0.5597 | | arguana | 0.4402 | 0.2144 | 0.4408 | 0.2262 | 0.3461 | | scidocs | 0.1443 | 0.1233 | 0.1475 | 0.1020 | 0.1366 | | fiqa | 0.1892 | 0.1648 | 0.1941 | 0.1118 | 0.1591 | | trec-covid | 0.5105 | 0.6376 | 0.5209 | 0.3454 | 0.4474 |

The table reports pure WASM late-interaction rerank over the static top-100 candidates at 384 dims. Pure rerank helps SciFact and TREC-COVID, but hurts ArguAna and lowers the BEIR-6 average; for query-style workloads, blend with first-stage baseScores so rerank refines rather than replaces the vector score.

Similarity Quality

| benchmark | mode | metric 1 | metric 2 | |---|---|---:|---:| | GLUE STS-B | static | Spearman 0.8335 | Pearson 0.8376 | | GLUE STS-B | hybrid | Spearman 0.8337 | Pearson 0.8376 | | GLUE STS-B | hashing | Spearman 0.7078 | Pearson 0.7066 | | SprintDuplicateQuestionsPC | static | max AP 0.9347 | — | | SprintDuplicateQuestionsPC | hybrid | max AP 0.9371 | — | | TwitterURLCorpusPC | static | max AP 0.8194 | — |

static dimension tradeoff:

| output dim | size vs full | BEIR-6 nDCG@10 | of full | |---:|---:|---:|---:| | 512 (default) | 100% | 0.3761 | 100% | | 384 | 75% | 0.3731 | 99.2% | | 256 | 50% | 0.3664 | 97.4% | | 128 | 25% | 0.3458 | 91.9% |

API

• embed(text, options?)
• embedBatch(texts, options?)
• createQuickiEmbed(options?)
• getQuickiEmbed(options?)
• QuickiEmbed.create(options?)
• QuickiEmbed.fromURL(url, dim?, mode?)
• QuickiEmbed.fromFile(pathOrUrl, dim?, mode?)
• QuickiEmbed.fromBytes(bytes, dim?, mode?)

Instance methods:

• setIdf(idf)
• clearIdf()
• setTokenIdf(idf)
• clearTokenIdf()
• fitIdf(texts)
• fitRetrieval(texts)
• embed(text)
• embedBatch(texts)
• scoreLateInteraction(query, document)
• scoreLateInteractionBatch(query, documents)
• rerank(query, documents, options?)
• close()

Compatibility aliases:

• FastEmbed
• createFastEmbed(...)
• getFastEmbed(...)

Built By

Built by Wesley at Burke Designs LLC.

https://burkedesigns.biz