Skip to content

Knowledge

Vandelay includes a RAG (Retrieval-Augmented Generation) pipeline for giving your agent access to documents and reference material.

How It Works

Documents → Chunking → Embedding → ChromaDB Vector Store
                                        ↓
                              Agent queries knowledge
                                        ↓
                              Relevant chunks returned
                                        ↓
                              Agent uses context in response

Enabling Knowledge

vandelay config  # → Knowledge → Enable

Or in config.json:

{
  "knowledge": {
    "enabled": true
  }
}

Embedder Resolution

The embedder is auto-resolved from your model provider:

Model Provider Embedder Used Model
OpenAI OpenAI Embedder text-embedding-3-small
Google Gemini Embedder default
Ollama Ollama Embedder default
Anthropic fastembed (local) BAAI/bge-small-en-v1.5
OpenRouter OpenAI (if key set) or fastembed varies

Anthropic has no embedding API, so Vandelay falls back to fastembed, a local embedder that requires no API key.

Override the embedder explicitly in config:

{
  "knowledge": {
    "embedder": {
      "provider": "openai",
      "model": "text-embedding-3-small"
    }
  }
}

Vector Store

Documents are stored in ChromaDB at ~/.vandelay/data/knowledge_vectors/. ChromaDB is embedded (no server needed) and supports fast similarity search.

CLI Commands

vandelay knowledge add ~/docs/report.pdf   # Add a file
vandelay knowledge add ~/my-notes/         # Add a directory
vandelay knowledge status                  # Check status
vandelay knowledge list                    # Show vector count
vandelay knowledge refresh                 # Refresh built-in corpus
vandelay knowledge refresh --force         # Force full rebuild
vandelay knowledge clear --yes             # Clear all knowledge

See CLI Reference: knowledge for full details.