How we built Text-to-SQL at Pinterest

Pinterest built a Text-to-SQL feature using Large Language Models (LLMs) to help users write analytical queries more easily. They implemented a system that retrieves table schemas, incorporates low-cardinality columns, uses WebSocket for response streaming, and integrates Retrieval Augmented Generation (RAG) for table selection. The second iteration includes offline vector index creation, NLP table search, table re-selection, and query summarization. Areas for further development include metadata enhancement, scheduled or real-time index updates, scoring strategy revision, query validation, and user feedback.