docmd search - Offline Semantic Search

Private Semantic Search

Bring modern vector search directly to your static docs without third-party services.

Local Vector Models

Uses lightweight ONNX models executing directly on the client side. No data ever leaves the user's browser.

Semantic Intent

Understands queries conceptually. Searching 'authentication' yields results for 'sign-in' and 'login' automatically.

Hybrid Re-ranking

Combines fast BM25 keyword matching with vector similarity for the absolute best search relevance and speed.

Supported Embedding Models

Select the perfect embedding model tailored for your documentation size and languages.

Model	Dimensions	Size	Languages	Best For
MiniLM L6 v2 ★	384	~23 MB	English only	Fast, general English docs
Multilingual MiniLM L12	384	~118 MB	50+ languages	i18n documentation
Multilingual E5 Small	384	~118 MB	100+ languages	Wide language coverage
Multilingual MPNet Base	768	~270 MB	50+ languages	Best multilingual quality

💡 Multilingual documentation: If your documentation website contains multiple languages (such as English, Chinese, German, Spanish, etc.), select a multilingual model using docmd-search --settings. The default model is English-only and will produce poor search relevance for other languages.

Pure-Client API: Build Custom UIs

An ultra-lightweight client runtime (<5KB minified) that executes entirely in the browser using hybrid BM25 and vector cosine similarity scoring.

import * as Search from 'docmd-search/client';
// 1. Initialise and load index folder.
// Batch 0 is loaded instantly for sub-millisecond search startup,
// while remaining chunk batches load progressively in background.
await Search.load('/.docmd-search', (loaded, total) => {
  console.log(`Loaded batch ${loaded}/${total}`);
});
// 2. Query search using hybrid vector + keyword matching.
const results = Search.search('authentication secure routes', 10);
results.forEach(({ score, chunk }) => {
  console.log(`[${(score * 100).toFixed(0)}%] ${chunk.file}#${chunk.heading || ''}`);
  console.log(chunk.text);
});

Build Index in One Command

Run the indexing engine to generate static search chunks for your site.

Terminal

$ npx docmd-search --build
✓ Loaded local vector model
✓ Generated static index: .docmd-search/index.json

Search Frequently Asked Questions

Everything you need to know about docmd offline semantic search.

Does docmd search run entirely client-side?

Yes. The search runtime performs all vector similarity calculations and keyword indexing directly in the user's browser. It requires zero cloud infrastructure or search APIs.

Read docmd-search documentation →

Are ONNX model weights downloaded on every visit?

No. Model weights are only loaded during compilation/build time using Node.js to generate static search vector indices. The client-side runtime only downloads the pre-built indices, which are compressed JSON chunks.

Learn how offline search works →

Can I use the search plugin with multiple languages?

Yes. If your documentation is multilingual, you should choose a multilingual embedding model like paraphrase-multilingual-MiniLM-L12-v2 which supports over 50 languages.

Learn about multi-language search models →

What is showConfidence and how does it work?

The showConfidence flag, when enabled, displays the exact similarity matching percentage badge next to each semantic result, helping users visually understand search relevance.

Read about browser-client settings →

Local Semantic Search.
Zero API keys. 100% Private.