Skip to content
browser-llm-nexus

Knowledge & offline bundles

NexusKnowledge composes a chat model, an embedding model and a vector index into one object: you add documents, you ask questions, you get answers grounded in those documents. Chunking, embedding, indexing, retrieval and context assembly are handled.

import { NexusKnowledge } from 'browser-llm-nexus';


const kb = await NexusKnowledge.create({
  chat: { hub: 'onnx-community/Qwen3-0.6B-ONNX' },
  embedder: { hub: 'Xenova/bge-small-en-v1.5' },   // also the default
});


await kb.addDocument({ id: 'handbook', title: 'Handbook', text: handbookText });


kb.on('token', t => render(t));
const answer = await kb.ask('What is the refund policy?');

chat and embedder each accept either a model source or an already-loaded NexusChat / NexusEmbedder, so two knowledge bases can share one loaded model instead of paying for it twice. chunkSize, chunkOverlap and topK are options on create.

The part nobody else has: ship it offline

Section titled “The part nobody else has: ship it offline”

A knowledge archive is the three artifacts composed into one zip:

manifest.json          { kind: 'knowledge', models, docs, contains }
rag/                   the vector store
models/chat/model.zip  optional — a full model archive, nested
models/embedder/model.zip

Export it:

const zip = await kb.exportZip({ includeModels: true });   // rag + both models
download(zip, 'handbook-kb.zip');

Import it on a machine with no internet:

const kb2 = await NexusKnowledge.importZip(fileFromInput);        // a File from <input>
const kb3 = await NexusKnowledge.importZip('/bundles/handbook-kb.zip');

The weights are restored into the browser cache and the vectors are restored as vectors — nothing is downloaded and nothing is re-embedded. Embedding a corpus is the slow, expensive step; doing it once at build time and shipping the result is the whole point.

Vectors without the text

Section titled “Vectors without the text”

When the source documents themselves shouldn’t travel — customer data, licensed content — ship the index and leave the text behind:

const zip = await kb.exportZip({ includeText: false });

Retrieval still works. The chunks that come back carry vectors and metadata, not the original prose.

Look before you load

Section titled “Look before you load”

inspect reads the manifest without loading a single model — use it to show a user what a 2 GB file contains before they commit to it:

const manifest = await NexusKnowledge.inspect('/bundles/handbook-kb.zip');
console.log(manifest.models, manifest.docs.length, manifest.contains);

Three archive kinds, each standalone

Section titled “Three archive kinds, each standalone”

You don’t have to bundle everything. Each artifact exports and imports on its own:

Kind Export Contains
Model exportModel(id, opts) one model’s weights, dtype-filterable
Index exportIndex(index, opts) vectors + chunk text
Knowledge kb.exportZip(opts) manifest + index + optionally both models

exportModel takes a dtypes filter — shipping only q4 keeps an archive small enough to actually hand someone.

fflate does the zipping. It’s an optional peer dependency: install it, pass your own with { zip }, or let it load from a CDN.

Where this is worth it

Section titled “Where this is worth it”
  • Air-gapped or field deployments — a USB stick instead of a VPN.
  • Regulated data that must not reach a server, including your server.
  • A demo that works on conference wifi, because it doesn’t use the network.
  • Deterministic builds — the exact weights and the exact vectors, versioned as one file.
Section titled “Related”