I let a model suggest Postgres indexes, then made the database mark its work

A developer has created a tool to validate AI-generated Postgres index suggestions by testing them directly within the database. Recognizing that LLMs often propose plausible but ineffective indexes, the author built a system that creates suggested indexes within a transaction, runs EXPLAIN (ANALYZE, BUFFERS) to verify if the query planner actually utilizes them, and then rolls back the changes. The experiment, which tested nine different models across eight common queries, revealed that while models are effective at identifying standard optimization opportunities, they often fail to recognize when a query cannot be improved by indexing. The findings suggest that automated verification is essential, as models frequently propose redundant or useless indexes. The tool, 'pg-index-referee', is available on GitHub and allows developers to safely benchmark AI suggestions without permanently altering their database schema.