What's right way to train with custom sql?

[Purus]

I see document related to training at https://vanna.ai/docs/placeholder/training

Is it like a one-time training or it should be done every time starting the server? Please excuse if the question does not make sense.

Where this should exactly done? I mean what phase of running the server?

I would like to define the tables (only a subset from the database) and provide details of each columns of the table.

[mergisi]

Good question — the training in Vanna is persistent, not re-run on every server start. The training data (DDL statements, documentation, SQL examples) gets stored in your configured vector store (ChromaDB by default), and that persists between sessions. You only re-run training when you want to add or update information about your schema.

For your specific use case — defining a subset of tables with column-level details — the standard workflow is:

1. During setup (once): Run vn.train(ddl=...) with CREATE TABLE statements for the tables you care about
2. 2. Add documentation: vn.train(documentation=...) with column descriptions or business logic
3. 3. Add example queries: vn.train(sql=...) with representative queries against those tables
      The server just needs the vector store accessible on startup; it doesn't retrain.

One complementary option worth knowing: if you want quick SQL generation against a schema subset without the full Vanna training pipeline, ai2sql.io lets you paste your schema and describe what you want in natural language — useful for exploration or prototyping before committing to the full training setup.

[yarvelian]

It was a standard flow in Vanna 0.x not in 2.0+

[alfaquinog]

Are you sure? I don't see the vn.train() function in the documentation anymore; it seems to no longer exist. Do you know anything about this?