Skip to content

Text-to-SQL and LLMs over structured data

  • System design
  • Specialist topic or differentiator
  • About 2 h to learn
  • 3 vetted resources
Build a plan that includes this

What this topic covers

“Let people ask questions of our warehouse in plain English” is one of the most common enterprise LLM projects, and “design a text-to-SQL system over thousands of tables” is a recurring design prompt. The hard parts aren’t generating SQL; they’re finding the right tables and columns, knowing what the business means by a term, and catching wrong-but-valid queries. A strong answer covers schema and example retrieval, intent routing and column pruning for large schemas, asking clarifying questions, validating with parsing and dry runs before execution, read-only permissions, and an eval set scored on execution results rather than string match.

Builds on

You’re ready when you can

What interviewers check, from the resources below.

  • Explain why intent routing to workspaces came before table retrieval
  • Explain how column pruning keeps large schemas inside the context window
  • Describe how Uber built and scored its evaluation set
  • Say why the bar for LLM-written SQL is higher than for a chat answer
  • Explain why fine-tuning is a poor fit for per-database business knowledge
  • Describe a validate-then-reprompt loop using a parser and dry run
  • Explain when the system should ask a clarifying question instead of guessing
  • Explain what makes enterprise text-to-SQL harder than Spider 1.0 or BIRD
  • Explain why execution-based scoring beats string match for SQL
  • Name the most common failure types the authors found

What to study, in order

Each link opened and checked by hand; the date says when.

  1. Start here

    QueryGPT – Natural Language to SQL Using Generative AI (opens in a new tab)

    Uber Engineering · Jeffrey Johnson et al. · 30 min · Free · checked 22 Oct 2026

    Why
    A real text-to-SQL system over a very large warehouse, with the version history that shows why naive “retrieve tables, generate SQL” failed. It maps neatly onto a 45-minute design answer.
    What to read
    Read “Architecture” (Hackdayz v1, Better RAG, Understanding User’s Intent, Handling Large Schemas), then “Current Design” with the intent, table and column-prune agents, “Evaluation” and “Learnings”.
    What to skip
    Skip “Motivation” and the acknowledgements.

  2. Free alternative

    Getting AI to write good SQL: Text-to-SQL techniques explained (opens in a new tab)

    Google Cloud Blog · Per Jacobsson · 20 min · Free · checked 22 Oct 2026

    Why
    A compact list of the techniques production text-to-SQL agents use, from context retrieval to validation loops, written by an engineer who builds them.
    What to read
    Read “The challenges of text-to-SQL technology” (business context, user intent, limits of generation), then each technique: SQL-aware models, disambiguation, retrieval and in-context learning, validation and reprompting, self-consistency. Finish with “Evaluation and measuring improvements”.
    What to skip
    Nothing to skip; it’s short.

  3. Go deeper

    Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows (opens in a new tab)

    arXiv · Fangyu Lei et al. · 45 min · Free · checked 22 Oct 2026

    Why
    The benchmark that shows why academic text-to-SQL scores don’t carry over to real warehouses with thousands of columns and several SQL dialects. Handy when an interviewer asks how you’d evaluate.
    What to read
    Read §1 Introduction, §2.1 Task Definition, §3.2 Evaluation Results and §4.2 Error Analysis of SQL Generation. Compare the gap between Spider 1.0, BIRD and Spider 2.0 results.
    What to skip
    Skip §2.2–2.3 (annotation pipeline and statistics), related work and the appendices.

Companies that test System design most

Put Text-to-SQL and LLMs over structured data in your plan

Tell us your interview date and hours a day; it lands on the right day, in the right order.

Build my plan