Vanna AI

Overview

Vanna AI is a framework for accurate text-to-SQL generation via LLMs using Agentic Retrieval-Augmented Generation (RAG). It enables users to chat with their SQL databases using natural language, automatically generating and executing SQL queries.

How It Works

Vanna operates in two steps:

1. Train: Build a RAG model on your data by storing database schema, documentation, and question-SQL pairs in a vector store
2. Ask: Ask questions in natural language, which retrieves the most relevant context to help generate accurate SQL queries

Vector Database Support

Vanna supports multiple vector databases:

• ChromaDB: Default vector store with ChromaDB_VectorStore implementation
• Milvus: World's most advanced open-source vector database
• Qdrant: High-performance vector search
• Custom: Extensible architecture supports any vector database

Architecture Components

• Vector Store: Stores embeddings of schema, documentation, and example queries
• Embedder: Generates vector embeddings from text
• LLM: Multiple providers supported (OpenAI, Anthropic, Google Gemini, Ollama, AWS Bedrock)
• Database Connector: Connects to any SQL database

Key Features

• Natural language to SQL translation
• Automatic query execution
• Support for multiple LLMs and vector databases
• Extensible framework for customization
• Integration with popular SQL databases

Use Cases

• Business intelligence querying
• Data exploration without SQL knowledge
• Automated reporting
• Database Q&A systems