Vanna.

Between a question and a table

Vanna connects a language model to a database and query execution tools. A user asks a question in ordinary language, and a configured agent selects the actions. Current documentation gives particular attention to user identity, permissions and memory of successful interactions. These matter more than one impressive text to SQL demonstration.

The agent needs the data structure and the meaning of business fields. Columns with the same name in different tables can represent different events. Without this preparation, a plausible query can return a formally valid but misleading total. A table of results does not establish that the answer matches the question.

A practice task with restricted permissions

Prepare a small order database without personal information. Give the agent a read only connection and ask for order counts by month. Define which statuses count as an order and which date to use before running it. Compare the result with your own control query.

Try a month with no records, duplicate rows after joining tables and an order whose status changed. These cases reveal whether the agent understood the structure. Saving a successful example can help later queries, but a wrong example can also persist. Check selection criteria and units before reuse.

One question can count different things

Suppose orders and order items are stored in separate tables. After a join, one order may appear several times. Counting rows then counts items, although the person asked for orders. The question sounds natural and the query runs without error, yet the result means something different.

Define business terms before using an analytical assistant: what counts as a sale, which date applies and how canceled records are handled. Control examples should show the query as well as the final number. This can reveal an error before the agent reuses the example.

Deployment and usage terms

Vanna offers an open framework for running your own deployment and cloud administration features. The documentation covers different database systems and language model providers. Hosting the application locally does not automatically make model inference local: the question and available context may go to the selected provider.

Paid plans cover administrative features, queries and enterprise support. Budget separately for the connected model and your database. For production data, establish who can see queries, where memory is stored and how each user’s permissions are checked. Start with reading; add modification only for a specific verified workflow with a separate confirmation step.

Suitable tasks

• Create a database assistant for questions
• Prepare and verify SQL queries
• Manage user rights of analytical agent