RAG Knowledge Base

The voice agent includes a privacy-first RAG (Retrieval-Augmented Generation) system powered by Turso (Limbo) WASM, using a browser-local database with in-browser query embeddings.

How It Works

1. Build-time indexing: Run bun run build:rag to generate embeddings from all documentation using llama.cpp with Vulkan acceleration
2. Local embeddings: Uses Transformers.js with Xenova/all-MiniLM-L6-v2 for fast, local semantic query embeddings
3. Semantic search: When the user asks a question, the AI first searches the knowledge base for relevant docs
4. Grounded responses: The AI formulates answers based on retrieved context, ensuring factual accuracy

Generated Files

When you run bun run build:rag, the system generates pre-computed embedding artifacts that are loaded directly into the browser. These files contain chunked documentation content with corresponding vector embeddings, ready for instant semantic search without any client-side processing.

The build process creates a complete RAG knowledge base including:

• rag-index.yml — chunked documentation content with metadata and precomputed vectors
• rag-documents.yml — full document content for hierarchical retrieval

These files are served as static assets and loaded into the browser-local Turso database on page initialization.

Key Features

• Zero cloud costs: No API calls or external services for search
• Privacy-preserving: All data stays in the browser
• Offline-capable: Works without internet after initial page load
• Instant responses: ~50ms search latency for 10K+ documents
• Source citations: The AI can reference specific documentation pages

Debug Mode

Set VITE_VOWEL_DEBUG_RAG=true to enable a debug panel that shows:

• Real-time STT (speech-to-text) transcripts
• RAG search queries and retrieved documents
• Confidence scores and source paths

Users can also say "Open the debug chat" to see this panel.

The debug interface provides complete visibility into the RAG pipeline:

The files panel displays all loaded documentation chunks and their metadata:

Agent Skills

• haven-local-rag (.agents/skills/haven-local-rag/) - Browser-based semantic search and local RAG pipeline background
• rag-prebuild (.agents/skills/rag-prebuild/) - Pre-build embeddings with build-rag.py for production