Spotify built an AI data assistant called Vedder to help over 2,100 employees query 70,000+ datasets without needing SQL expertise. The key architectural insight is a 'cluster' model where domain experts curate the context the LLM sees — including relevant datasets with schema profiling, vetted question-SQL pairs, and business documentation. Automated generation of question-SQL pairs from query history was tested but rejected: domain experts accepted only 12.5% of auto-generated pairs, as most were noise. Each cluster has a health score tracking schema validity, coverage, and SQL reproducibility. Every conversation feeds back to cluster owners, enabling continuous improvement. The architecture encodes expert knowledge at scale rather than replacing experts.
