philipp-zettl/vrom-hf-docs

🧩 vROM: HF Transformers & Hub Documentation

Vector Read-Only Memory — Pre-computed HNSW index for instant in-browser RAG

What is a vROM?

A vROM (Vector Read-Only Memory) is a pre-computed, serialized HNSW index package that can be loaded directly into VecDB-WASM for instant vector search in the browser — no embedding computation required on the client side.

Think of it as a plug-and-play RAG cartridge: download, load, and search in milliseconds.

This vROM

Contains pre-embedded documentation from:

• Hugging Face Transformers (v5.6) — Installation, Quick Start, Pipeline API, Training, Fine-tuning, Tasks, Quantization, API Reference
• Hugging Face Hub — Repositories, Models, Datasets, Spaces, Uploading, Downloading

Metric	Value
Vectors	1,356
Dimensions	384
Total Tokens	~233K
Index Size	12.6 MB
Embedding Model	Xenova/all-MiniLM-L6-v2 (q8)
Distance Metric	Cosine
HNSW M	16
HNSW efConstruction	128

Quick Start

Browser (with VecDB-WASM)

import init, { VectorDB } from 'vecdb-wasm';
import { pipeline } from '@huggingface/transformers';

// 1. Initialize WASM
await init();

// 2. Fetch and load the vROM
const response = await fetch(
  'https://huggingface.co/datasets/philipp-zettl/vrom-hf-docs/resolve/main/index.json'
);
const indexJson = await response.text();
const db = VectorDB.load(indexJson);

console.log(`Loaded ${db.len()} vectors (${db.dim()}d)`);

// 3. Embed a query with Transformers.js
const extractor = await pipeline(
  'feature-extraction',
  'Xenova/all-MiniLM-L6-v2',
  { dtype: 'q8' }
);
const output = await extractor('how to fine-tune a model', {
  pooling: 'mean',
  normalize: true
});

// 4. Search!
const results = JSON.parse(
  db.search(new Float32Array(output.data), 5)
);

for (const { id, distance, metadata } of results) {
  const meta = JSON.parse(metadata);
  console.log(`[${distance.toFixed(3)}] ${meta.section_heading}`);
  console.log(`  ${meta.text.slice(0, 100)}...`);
  console.log(`  Source: ${meta.url}`);
}

Context Expansion (The Linked-List Trick)

Each chunk has prev_chunk_id and next_chunk_id pointers. After finding a relevant chunk, expand context by following the chain:

function expandContext(db, chunkId, windowSize = 2) {
  const chunks = [];
  const meta = JSON.parse(db.get_metadata(chunkId));
  const parsed = JSON.parse(meta);
  
  // Walk backwards
  let prevId = parsed.prev_chunk_id;
  const before = [];
  for (let i = 0; i < windowSize && prevId !== null; i++) {
    const prevMeta = JSON.parse(JSON.parse(db.get_metadata(prevId)));
    before.unshift(prevMeta);
    prevId = prevMeta.prev_chunk_id;
  }
  
  // Walk forwards
  let nextId = parsed.next_chunk_id;
  const after = [];
  for (let i = 0; i < windowSize && nextId !== null; i++) {
    const nextMeta = JSON.parse(JSON.parse(db.get_metadata(nextId)));
    after.push(nextMeta);
    nextId = nextMeta.next_chunk_id;
  }
  
  return [...before, parsed, ...after];
}

Files

File	Size	Description
index.json	12.6 MB	HNSW index (loadable by VectorDB.load())
chunks.json	1.5 MB	Chunk metadata array (for browsing/filtering)
manifest.json	1.2 KB	Package specification

Chunk Metadata Schema

Each vector's metadata (accessible via db.get_metadata(id)) is a JSON string:

{
  "chunk_id": 42,
  "text": "The actual chunk text...",
  "source_file": "transformers/pipeline_tutorial.md",
  "section_heading": "Pipeline API",
  "prev_chunk_id": 41,
  "next_chunk_id": 43,
  "url": "https://huggingface.co/docs/transformers/pipeline_tutorial",
  "doc_title": "Pipeline"
}

Chunking Strategy

• Method: Section-aware (splits on markdown headings)
• Target size: 256 tokens per chunk
• Overlap: 0 (research shows overlap adds cost without improving retrieval)
• Code blocks: Preserved intact within chunks
• Linked list: prev_chunk_id / next_chunk_id for context traversal

Compatibility

• VecDB-WASM: ≥0.1.0
• Load method: VectorDB.load(json_string)
• Browser embedding model: Xenova/all-MiniLM-L6-v2 with {pooling: 'mean', normalize: true}

Build Info

• Corpus hash: 9109c35ca66f0071
• Built: 2026-04-24
• Source: HF Transformers Docs + HF Hub Docs

Part of the vROM Ecosystem

This is an official first-party vROM built by the VecDB-WASM team. See our VecDB-WASM Space for the core engine.

vROMs transform VecDB-WASM from a standalone database engine into a distribution hub for plug-and-play RAG architectures — the "NPM for AI Agent Memory."