# Text-to-SQL

> Text-to-SQL is turning a plain-language question into a SQL query a database can run, using a model grounded in your schema, documentation, and past queries.

Text-to-SQL turns a plain-language question into a runnable SQL query. Accuracy depends almost entirely on grounding: the model needs your DDL, your column descriptions, and known-correct example queries. It is a retrieval problem wearing a language-model costume, and the generated query is meant to be read before it is trusted.

**Text-to-SQL is the technique of turning a plain-language question into a SQL query a database can run, using a language model that has been grounded in your schema, documentation, and previous queries.** The output is a query, not an answer, which is what makes it reviewable.

The hard part is not the SQL. It is context. A model that has never seen your warehouse does not know that `rev_net_usd` is the revenue everyone means, that orders join to customers through an intermediate table, or that your fiscal year starts in February. Tools solve this by retrieving relevant context at query time: [Vanna](/tools/vanna) saves written documentation and known-correct question-and-SQL pairs to agent memory, embeds them, and pulls the most similar pieces into the prompt, which is [RAG](/glossary/rag) applied to a database. Enterprise platforms go further and route the question through a governed [semantic layer](/glossary/semantic-layer) so metric definitions are fixed rather than inferred.

Because the intermediate artifact is a query, text-to-SQL is one of the few AI features where verification is straightforward: read the SQL, check the joins and the filters, then run it. That is also the discipline the whole category depends on, and the reason the generated query should never be hidden from the person asking.

The Claude-specific walkthrough is [text-to-SQL with Claude](/guides/analytics/text-to-sql-with-claude), the tool-by-tool comparison is [the best text-to-SQL tools in 2026](/guides/comparisons/best-text-to-sql-tools-2026), and the wider workflow it fits into is [Claude for data analysis](/guides/analytics/claude-for-data-analysis).

---

_Source: https://agentscamp.com/glossary/text-to-sql — Term on AgentsCamp._