Tabular Retrieval Tools

Answer structured-data questions via text-to-SQL, execution, and generation — treating schema selection and SQL runners as agent tools.

Tabular questions ("units of Fruity Fedora sold in the last 7 days") are not solved by dumping table text into context. The workflow is: (1) text-to-SQL from query plus schema, (2) SQL execution, (3) natural-language generation from the result. If many schemas cannot fit in context, first predict which tables to use.

This is the bridge from classic RAG into the agentic pattern: retrievers and SQL executors are tools in the tool inventory. Write-capable SQL raises indirect prompt injection and human approval gates concerns — treat generated SQL like any other write action under agent system-level defenses.

---

Sources

• AI Engineering (Chip Huyen), ch. 6