Behind the curtain: Key learnings from building Genie

For teams evaluating agentic analytics with AI/BI and Genie, the talk outlines the design tradeoffs that make an agent production ready.

Large language models are improving quickly, yet they do not know an organization's definitions, trusted sources, or preferred workflows. Genie One's core problem is to close that gap automatically and at scale, so a question like "What was revenue last quarter?" resolves to the right definitions, tables, joins, and time boundaries without hand holding.

Rebuilding enterprise context on demand

Genie Agents, formerly Genie Spaces, demonstrated that human-curated instructions, examples, and definitions yield high recall, precision, and authority, but curation does not scale to every domain. Genie One flips the model by assembling context dynamically at query time.

Unity Catalog remains the governed source of truth, while most knowledge lives in day-to-day work on the Databricks Platform. Genie One indexes dashboards, notebooks, pipelines, tables, columns, views, and prior agent conversations, enriches them with generated summaries and keywords, and distills results so the agent does not drown in tokens. Lineage and access controls ensure users only see context derived from assets they are allowed to access.

Ranking authority with Genie Ontology

Not all assets are equally trustworthy. Genie Ontology, a system for governed and discovered knowledge, combines Unity Catalog governance with authority signals learned from platform activity. Popularity helps, but expertise matters more.

A PageRank-style algorithm propagates authoritativeness across a usage graph. Signals include expert authorship, expert consumption, and asset popularity, which together elevate certified, relied-upon artifacts over ad hoc ones. Conflicting or evolving definitions can be resolved by prioritizing higher authority context for a given topic and team.

Harness design for speed and breadth

Genie One favors a flat agent architecture for most tasks. Flat traces are faster, simpler to debug, and more token efficient, especially with prompt caching, and they avoid lossy handoffs between sub-agents. Hierarchy appears selectively when task complexity warrants isolation or when an existing Genie Agent can be reused as a specialized sub-agent.

Skills and tools are bundled into groups and loaded only when needed, keeping the default prompt lean while enabling deep capabilities on demand. Asset discovery blends parallel exploration across asset types, type-specific ranking, iterative agentic search, and usage-graph relationships. Non-blocking parallelism returns fast results immediately while slower SQL jobs continue in the background, which keeps explorations responsive and reduces end-to-end latency.

Grounding and personalization to keep answers reliable

Genie One forces computation into code, not the model's head. Aggregations, joins, filters, and numerical reasoning execute as SQL or Python so results are reproducible and auditable. Citations back to queries, tables, and retrieved evidence are mandatory, user inputs and identifiers are validated against real data, and untrusted content is sanitized against injection. Authority signals weight certified and widely used assets more than ad hoc ones.

Personalization sharpens relevance. Recent activity guides asset discovery, upcoming personal memory features will reuse prior queries and workflows, and personal MCP connections extend context into tools like docs, email, calendar, chat, and tickets, while respecting access controls.

Production AI means context first, models second

An AI coworker becomes valuable when it assembles the right enterprise context, executes computation verifiably, and stays fast under real workload. Genie One's ontology-driven ranking, flat-first harness, deferred skill loading, parallel exploration, and strict grounding form a practical blueprint for agentic analytics in production.