Skip to content
browser-llm-nexus

Embeddings & RAG

NexusKnowledge is the batteries-included path. When you want your own retrieval logic — a different chunker, a hybrid ranker, your own store — the pieces are exported separately.

Embed

Section titled “Embed”
import { NexusEmbedder } from 'browser-llm-nexus';


const embedder = await NexusEmbedder.load({ hub: 'Xenova/bge-small-en-v1.5' });


const one  = await embedder.embed('a single sentence');        // Float32Array
const many = await embedder.embedBatch(['first', 'second']);   // Float32Array[]

Vectors come back mean-pooled and normalized, so cosine similarity is a dot product. The embedder defaults to dtype: 'q8' — a good size/quality point for embedding models — and does the same WebGPU → WASM detection as everything else. Call embedder.dispose() when you’re done to release the backend.

Section titled “Index and search”

MemoryIndex is an in-memory cosine index, deliberately dependency-free. For thousands of chunks it’s plenty; swap in a real store when you outgrow it.

import { MemoryIndex, chunkText } from 'browser-llm-nexus';


const index = new MemoryIndex();
const chunks = chunkText(documentText);
const vectors = await embedder.embedBatch(chunks);
chunks.forEach((text, i) => index.add({ id: String(i), text, vector: vectors[i] }));


const hits = index.search(await embedder.embed(question), 5);   // [{ chunk, score }]

add / addAll insert, all() returns everything in insertion order, size counts. Each chunk carries an optional meta object that survives search, serialization and export — use it for source ids, page numbers, permissions.

Assemble context

Section titled “Assemble context”

contextFor is search plus the stuffing step, with a character budget so you don’t blow past the model’s window:

const context = index.contextFor(await embedder.embed(question), 5, 4000);
const answer = await chat.chat(`Context:\n${context}\n\nQuestion: ${question}`);

It walks hits best-first and stops before exceeding maxChars (default 4000).

Persist

Section titled “Persist”

Round-trip through plain JSON, or through a portable zip:

localStorage.setItem('idx', JSON.stringify(index.serialize()));
const restored = MemoryIndex.restore(JSON.parse(localStorage.getItem('idx')));


import { exportIndex } from 'browser-llm-nexus/rag';
const zip = await exportIndex(index);

serialize() writes vectors as number arrays; restore turns them back into Float32Arrays. Nothing is re-embedded either way — see offline bundles.

Similarity on its own

Section titled “Similarity on its own”
import { similarity } from 'browser-llm-nexus/embed';
similarity(vectorA, vectorB);   // cosine, for normalized vectors
Section titled “Related”