How can I get AI to know my own business terms?

Summary

• AI models don't know your internal terms, acronyms, or metric logic. Define them once in a governed place that people and AI both read, instead of re-explaining them in every prompt.
• On Databricks, Unity Catalog semantics holds that context: Pages for terms and acronyms (Beta), metric views for KPI logic, and certification flags for trusted assets.
• Genie Ontology (Public Preview) combines those definitions with context Genie infers from your assets and usage. Genie One and the Genie One MCP server use it to resolve business terms, and Unity Catalog permissions always apply.
• For policies, manuals, and wikis, retrieval-augmented generation (RAG) fetches relevant passages at question time, and Databricks AI Search indexes the source table.

How can I get AI to know my own business terms?

Give the AI a governed source of truth it can look up at question time, instead of hoping it guesses what you mean. Business context comes in two kinds. Structured definitions, such as what "active customer" means, how revenue is calculated, and which table is authoritative, belong in a semantic layer. Unstructured knowledge, such as policies, runbooks, and product manuals, belongs in a searchable index that the AI retrieves from.

Which approach fits which kind of business term?

What the AI needs to know	Approach	Databricks capability
Terms, acronyms, and entities	Governed definitions with synonyms	Pages in Unity Catalog semantics (Beta)
KPI and metric logic	Define once, query from many surfaces	Metric views
Which data to trust	Certified or deprecated flags	The system.certification_status tag
Table and column meaning for a single Genie Agent	Descriptions, synonyms, join relationships, SQL expressions, example queries	Genie Agent knowledge store and examples
Policies, manuals, and wikis	Retrieval-augmented generation	Databricks AI Search

How does Databricks give AI your business context?

Define it once. Unity Catalog semantics is "a collection of tools to help you define standard business metrics, terms, and organizational structures on top of your catalog data," so people and AI tools interpret data consistently under one governance model. A Page is a governed, authoritative definition of a business concept, with fields for synonyms, a description, an owner, and related assets. Genie Code can read documents you attach, extract terms, and propose Pages that you review before any are created. When Genie One answers a question about a term defined in a Page, it prioritizes the Page's definition over context it infers automatically and cites the Page. A metric view defines a KPI once, such as sum of revenue divided by distinct customer count, and users can group by any available field.
Tune one agent. Inside a single Genie Agent, the knowledge store adds agent-specific column descriptions, synonyms that match user language to column names, join relationships, and SQL expressions for measures, filters, and dimensions. These settings are scoped to the agent and do not change Unity Catalog metadata.
Resolve it everywhere. Genie Ontology (Public Preview) is the unified context layer for Genie. It brings together Unity Catalog semantics and inferred context: snippets that Genie extracts from metric views, dashboards, SQL queries, and Genie Agents. Each snippet gets an authority score based on where it came from, how often it is used, and how fresh it is, and Unity Catalog permissions gate which snippets Genie can use. A custom agent, including one built on Agent Bricks, can use the same context through the Genie One MCP server, which is generally available. The agent sends a natural-language question, and Genie answers from the ontology with links to the sources it cited.
Retrieve documents. Databricks AI Search builds an index from a source table and can sync automatically when that table changes. Hybrid keyword-similarity search suits source data with unique keywords such as SKUs or identifiers. In RAG, the request queries the index, the retrieved data is combined with the request into a prompt, and the model generates the response. The Databricks RAG guide lists proprietary knowledge, up-to-date information, source citation, and access control among the benefits.

What is a practical order of work?

1. Create a domain (a preview feature) for the business area, then author Pages for its key terms, or let Genie Code propose Pages from your existing documents.
2. Define KPIs as metric views so each metric has one definition.
3. Certify the assets your organization trusts and deprecate stale ones.
4. Index policies and manuals with Databricks AI Search for retrieval.
5. Ask real questions through Genie One or your agent, and check the citations.

FAQs

Can I put sensitive definitions in Pages?

No. Databricks warns that Page data does not support customer-managed key encryption and is stored in plain text that may be replicated globally. Keep PII, regulated data, and other sensitive information out of Page names, descriptions, synonyms, and body text.

Who can see my definitions?

A Page inherits access from the domain it belongs to, and anyone who can access the domain can read its published Pages. Unity Catalog permissions are always enforced for the Genie One MCP server, so users and agents can query only data they are allowed to access.

How do I keep definitions from going stale?

Mark assets as certified or deprecated, which steers Genie One toward the assets your organization vouches for. Genie Ontology also weighs how fresh each snippet is. For documents, set up the AI Search index to sync when the source table updates.

How do I check that answers are right?

Genie One cites the Page or sources behind an answer so users can verify it. For custom agents, trace each run and evaluate quality with LLM judges, custom scorers, and expert feedback.