Traditional full-text search relies on exact keyword matches. If a user searches for “authentication” but the page only uses terms like “OAuth2” or “login”, standard keyword search engines fail to discover it.
docmd provides client-side Hybrid Semantic Search powered by @docmd/plugin-search. It runs local Hugging Face ONNX model pipelines inside the browser, combining BM25 keyword frequency with vector cosine similarity for natural language understanding without third-party API calls.
Configuration
Enable semantic search in docmd.config.json:
{
"plugins": {
"search": {
"semantic": true,
"showConfidence": true
}
}
}
Embedding Model Profiles
| Model ID | Dimensions | Size | Languages | Primary Use Case |
|---|---|---|---|---|
Xenova/all-MiniLM-L6-v2 |
384 | ~90 MB | English only | High-accuracy English documentation. |
Xenova/LaBSE |
768 | ~470 MB | 100+ languages | Comprehensive multi-language support. |
Xenova/paraphrase-multilingual-MiniLM-L12-v2 |
384 | ~220 MB | 50+ languages | Recommended balance for international sites. |
Pre-Building Vectors in CI/CD
Pre-generate vector index chunks during build steps to accelerate browser execution:
# Build semantic search vector chunks
npx docmd-search --build
# Compile static site
npx @docmd/core build
This emits static Vecto-JSON chunks into .docmd-search/.
Commit .docmd-search/ to version control or cache it in CI/CD workflows. docmd-search performs incremental re-indexing, completing subsequent builds in under 300ms.
