Text-to-SQL systems through the papers and tutorials people actually use.

Course 6: Text-To-SQL Systems

One-hour objective: defend a text-to-SQL architecture using LLM-era sources that are either highly cited or practically canonical: prompt design, decomposition, enterprise tool flow, constrained decoding, and permission-safe BigQuery access.

Course 6 Reading List

Use three required sources: two LLM-era text-to-SQL papers plus one practical enterprise tutorial. Citation counts below are Semantic Scholar snapshots from this update and should be treated as popularity signals, not exact permanent numbers.

Readiness Checklist

You are ready for the interview version of this topic when you can connect the canonical benchmark story to a practical enterprise architecture.

• You can explain DAIL-SQL's LLM-era prompt design lessons: question representation, example selection, example organization, token efficiency, and execution accuracy.
• You can use DIN-SQL's decomposition pattern to separate intent classification, schema linking, query decomposition, SQL generation, and self-correction.
• You can translate the AWS Bedrock Agents tutorial into a BigQuery architecture with schema tools, query tools, auth boundaries, validation, and error handling.
• You can explain where constrained decoding or constrained validation fits before a warehouse call, using PICARD as the reference point.
• You can design a bounded repair loop that distinguishes parser failures, dry-run failures, runtime failures, semantic wrong-result failures, and uncertainty that needs human escalation.
• You can propose evals for execution accuracy, schema-linking accuracy, tool-call precision, query cost limits, permission denial, repair success, and answer provenance.

Interview Drill: AI Infra System Design

Prompt: design a BigQuery text-to-SQL assistant using the popular-source stack from this course: DAIL-SQL-style prompting, DIN-SQL-style decomposition, AWS-style enterprise tools, and PICARD-style validation.

• Start with the contract: accept natural language, user identity, workspace, candidate datasets, and clarification state; return SQL, validation receipts, result summary, and provenance.
• Design prompting with DAIL-SQL vocabulary: choose schema representation, select high-similarity examples, control token budget, and keep cross-domain examples from polluting the prompt.
• Use DIN-SQL's control flow: classify the question, decompose hard requests, select relevant schema context, generate SQL, and run a bounded self-correction loop.
• Borrow the AWS tutorial's production shape: expose schema discovery and query execution as tools, keep credentials outside the model, validate generated SQL, and capture structured errors for repair.
• Add PICARD-style safety: reject invalid SQL structures early, then dry-run for syntax and cost before any execution against customer data.
• Close with evals and operations: golden text-to-SQL tasks, execution accuracy, schema-linking accuracy, permission probes, cost-limit tests, trace receipts, and human review for high-impact queries.