Vanna AI

Vanna is an open-source Python framework for building text-to-SQL agents trained on your own schema and queries.

Overview

Vanna lets developers build a custom text-to-SQL assistant instead of buying one. You train it on your data dictionary, DDL, documentation, and example queries, then it generates SQL that matches your schema and conventions. Because it is retrieval-based and open source, it can run inside your VPC, connect to any database, and be embedded in notebooks, web apps, or Slack. Accuracy improves as you add validated question-SQL pairs.

Vanna’s training loop is its key insight: every validated question-SQL pair becomes retrieval context, so accuracy improves as the team uses it. Because the LLM is pluggable, organizations can point Vanna at private models to keep schema and data questions inside their own infrastructure.

Key Features

• Open-source text-to-SQL framework in Python
• Training on schema, docs, and validated query pairs
• Works with any LLM and any SQL database
• Runs locally or inside private infrastructure
• Integrations for Jupyter, web apps, and Slack
• Retrieval-based accuracy that improves with usage

Vanna’s modularity is the point: choose your LLM, choose your database, choose your training corpus. Teams running sensitive schemas in finance and healthcare adopt it precisely because nothing needs to leave their infrastructure.

Pricing

Vanna is free and open source. A hosted cloud option and enterprise support are available for teams that want managed deployment.

Who Is It For

Data teams and developers who want a customizable, private text-to-SQL capability inside their own products and workflows.

FAQ

Is Vanna free to use?

Yes, the core framework is free and open source; hosted and enterprise options are paid.

Does Vanna offer an API?

Vanna is a library you embed in your code; it exposes a Python API rather than a hosted REST service.

How does Vanna compare to alternatives?

Unlike SaaS query tools, Vanna gives you full control over models, training data, and deployment — at the cost of requiring engineering setup.

Which LLMs does Vanna support?

Vanna is model-agnostic and works with OpenAI, Anthropic, local models, and other compatible providers.

Can Vanna handle multiple databases?

Yes, you can configure separate training sets per database or schema for different teams and products.

Does Vanna require fine-tuning?

No fine-tuning is required; accuracy comes from retrieval over your training pairs.

Can non-engineers use it?

Once embedded in an app or notebook, the question-answer interface is approachable for non-SQL users.

Does Vanna improve over time?

Yes, accuracy climbs as you add validated question-and-query pairs to training.