Vanna

Vanna is an MIT-licensed open-source Python framework for natural language to SQL. It uses retrieval-augmented generation: you train it once on your database DDL, documentation, and example queries, and it retrieves the relevant schema context for each question before generating SQL tailored to your database.

The framework supports local vector stores (ChromaDB) and any LLM backend (Ollama for fully local, OpenAI or Anthropic via API), so teams can keep database contents private: only schemas, documentation, and queries are seen by the model layer, never raw table data. It integrates into notebooks, web apps, messaging platforms, and dashboards, with streaming results, visualizations, and enterprise features like user-aware permissions and query auditing in v2.

Accuracy depends heavily on training quality: well-documented schemas with example queries perform far better than bare DDL. Setup is more involved than hosted alternatives, and complex multi-join business questions can still trip it up. A hosted tier exists (reported around $0.53/query) for teams that don't want to self-host. The open-source core remains free and is widely used in production RAG-for-SQL setups.

pip install vanna. Choose a vector store (e.g. ChromaDB) and LLM backend (Ollama for local, OpenAI/Anthropic via API). Train once on your DDL, docs, and example queries with vn.train(). Ask questions via vn.ask() or generate_sql(); run results against your database.