docmd-search

Search that understands meaning and runs in the browser.
Offline semantic search for any documentation site. Understands meaning, runs entirely in the browser, and never sends data to the cloud.

Website • Documentation • Report Bug

Quick Start

Run docmd-search instantly on any folder:

npx docmd-search ./docs

That's it.

• Files are discovered and chunked automatically
• Embeddings are generated locally (no cloud API)
• Search is available in the terminal immediately

Install for regular usage

npm install -g docmd-search

# Install ML dependencies (one-time)
npm install -g @huggingface/transformers onnxruntime-node

docmd-search ./docs          # index + interactive search
docmd-search ./docs --ui     # index + web UI
docmd-search --settings      # configure model

Features

Designed to work offline, ship nothing to the browser, and stay out of your way.

Offline by default

• All embeddings generated locally with ONNX Runtime
• No data leaves your machine
• No cloud API keys needed

Instant search

• Progressive indexing: search available from the first batch
• Incremental: only re-indexes changed files
• Resumable: interrupted indexing resumes from last checkpoint

Tiny client

• Browser runtime is <3KB gzipped
• No model weights in the browser
• Hybrid scoring: keyword matching + vector similarity

Built-in capabilities

• Multi-batch index format with automatic compression
• Navigation tree generation for web UIs
• First-run setup wizard with model selection
• Interactive terminal search with live results

How It Works

Build time (Node.js)                    Search time (Browser, <3KB)
───────────────────                     ──────────────────────────
 Crawl files                             Load manifest.json
   → Chunk by heading                      → Load batch 000 (instant)
     → Embed via ONNX                        → Background-load rest
       → Quantize Float32 → Int8               → Keyword + cosine
         → Compress (ternary/PQ)                 → Ranked results
           → Save multi-batch index

Models

First run prompts you to select an embedding model:

Model	Dimensions	Size	Best for
MiniLM L6 v2 ★	384	~30 MB	Fast, general purpose
BGE Small (English)	384	~45 MB	English-optimised
BGE Base (English)	768	~110 MB	Higher quality
MPNet Base v2	768	~110 MB	Multilingual

Change model later: docmd-search --settings

Configuration (optional)

No configuration is required to get started.

Global (~/.docmd-search/config.json):

{
  "model": "Xenova/all-MiniLM-L6-v2",
  "wizardCompleted": true
}

Per-project (_docmd-search/config.json):

{
  "model": "Xenova/bge-small-en-v1.5",
  "chunkSize": 512,
  "include": ["**/*.md"],
  "exclude": ["**/drafts/**"]
}

Config resolution: defaults → global → project → CLI flags.

Programmatic Usage

Use in scripts or CI pipelines:

import { indexDirectory, loadAllBatches } from 'docmd-search';

const index = await indexDirectory({
  rootDir: './docs',
  outDir: '_docmd-search',
});

Browser client:

import { load, search } from 'docmd-search/client';

await load('/path/to/_docmd-search');
const results = search('deploy kubernetes', 10);

Project Structure

Keeps the codebase flat and modular.

src/
├── bin/docmd-search.ts   # CLI entry point
├── client/index.ts       # Browser runtime (<3KB)
├── config.ts             # Config + model profiles
├── index-io.ts           # Multi-batch format + compression
├── index.ts              # Barrel exports
├── indexer/
│   ├── chunk.ts          # Heading-aware chunking
│   ├── crawl.ts          # File discovery
│   └── index.ts          # Progressive pipeline
├── model.ts              # ONNX embedding manager
├── tui.ts                # Terminal UI
├── types.ts              # Core types
└── ui/
    └── launcher.ts       # Web UI via docmd

Part of the docmd ecosystem

docmd-search works standalone with any documentation project. It also integrates with docmd as a semantic search plugin.

Tool	What it does
docmd	Zero-config documentation generator
docmd-search	Offline semantic search engine

Community & Support

• Contributions are welcome
• If you find it useful, consider sponsoring or starring the repo ⭐

License

MIT License. See LICENSE for details.