Behind the curtain: Key learnings from building Genie

Chung Wu, Yi Liu16 Jun 2026

Problem

A general LLM does not know what "weekly active users" means in your company, which of roughly 400 tables with "revenue" in the name is the right one, or when your quarter starts. Teams define the same metric differently, and definitions change. Databricks Genie Spaces solved this with a subject-area expert who curates each agent by hand. That does not scale to every topic, and curated agents still miss context. Its new agent, Genie One, has to rebuild the relevant context on the fly for each question.

Example from the article: the question "what is my revenue last quarter?" needs a definition of revenue, a quarter start date, a choice among hundreds of candidate tables and views, and join and aggregation logic.

Solution in brief

Recreate the author's curation from platform data: recall through a semantic search index over assets and user conversations, precision through conflict resolution, and authority through a PageRank-like score over a graph of who creates and uses which asset. The agent then answers inside a harness that computes in code, cites sources and verifies inputs.

Questions this case answers
  • How to make an LLM data agent find the right business context and the one trusted source among many similar tables, dashboards and metric definitions without manual curation
  • How to structure an agent harness that computes in code, cites sources and verifies its inputs so that it does not invent numbers
  • How to make an LLM data agent find the right business context for a question without manual curation
  • How to decide which of many similar tables, dashboards and metric definitions an LLM agent should trust

Tools and methods

  • Semantic search index with distillation — finds assets and past conversations in compressed form.
  • Usage graph with PageRank-like propagation — ranks assets by trustworthiness.
  • Agentic asset search — iterative search with sub-agents.
  • Flat agent, lazy-loaded tools, non-blocking calls — limits latency and context size.
  • Grounding rules — code execution, citations, validation.
Continue reading — it's free

Google or a one-time email code. No password.