Unified Context-Intent Embeddings for Scalable Text-to-SQL

Pinterest evolved its Text-to-SQL system into a production Analytics Agent using unified context-intent embeddings and governance-aware ranking to serve 100,000+ tables and 2,500+ analysts.

• •A three-step pipeline converts historical SQL queries into semantic embeddings: domain context injection (PinCat glossary terms, metric definitions), LLM-based SQL-to-text, then text-to-embedding
• •SQL-to-text generates three outputs per query: a high-level business summary, analytical questions the query can answer, and a detailed logic breakdown
• •Analytical questions act as a retrieval bridge, matching new user queries against known intents across different phrasings and table structures
• •Table governance reduced the data warehouse from 400K to 100K tables via a Tier 1–3 system, providing quality signals for ranking