Problem
Spotify has more than 70,000 datasets, and demand for data answers outgrew what domain experts could handle one Slack message at a time. Putting all schemas into an LLM does not work. A million-token context cannot hold the whole warehouse, and a schema does not say what an "active user" means or that small values in an INT64 column are legacy test data. Given many similar tables, a model confidently picks the wrong one.
Example from the article: an INT64 column does not say that its values below 100 are legacy test data.
Solution in brief
Split the warehouse into domain "clusters", each owned by a named expert team. A cluster holds tables with profiling, expert-approved question-and-SQL pairs, and business docs. The agent works only from this curated context. A per-cluster health score tells owners where to spend curation time, and logged conversations feed back to them.
Questions this case answers
- How to make a natural-language data assistant give trustworthy answers when the warehouse is too large to fit into the model's context and schemas do not carry business meaning
- How to keep expert-curated context for an LLM agent accurate as the underlying data changes
- How to make an LLM data assistant pick the right tables and write trustworthy SQL over a warehouse with tens of thousands of datasets?
- How to scale the knowledge of a few data experts so that non-SQL users can self-serve data questions?
Tools and methods
- Domain clusters (datasets, pairs, docs) — the unit of curated context, owned by a named expert team.
- Few-shot question-SQL pairs — expert-approved examples that teach the model how to query the data.
- ReAct loop — the agent reasons and calls tools in steps, adjusting to each result.
- Cluster health score — continuous signals that show which cluster context has degraded.
- Slack bot, MCP server, web UI — the surfaces where users reach the agent.
Google or a one-time email code. No password.