How should a data agent retrieve definitions while working?

Retrieve when a decision needs meaning

Loading an entire business glossary into every prompt wastes space and still leaves the relevant definition ambiguous. Detect the metric, entity, or join choice that the task requires. Use a stable identifier when available; otherwise collect candidate definitions with their domains and owners. The agent should not generate an answer from the first semantic hit merely because the text sounds familiar.

Scope identity by domain and effective time

Two groups can use “active customer” with different inclusion rules. A definition is selected by more than a name: domain, owner, validity interval, permitted audience, and source revision all matter. Retrieve these fields together. If the user asks about last year’s cohort, a current definition may be the wrong one even when it is the newest and most popular result.

Two teams share a metric label

In a constructed company, the sales team counts an account active after a signed contract, while the product team requires recent usage. A user asks for active accounts in an adoption report. A naive retriever returns the sales definition because its name is an exact match. The agent compares task domain and report purpose, then selects the product definition or asks which population the user intends.

Run a retrieval checklist before query construction

The checklist is an interface between search and decision.

Check	Required evidence
Identity	Metric or entity ID, not only label
Scope	Domain, population, and permitted user
Time	Effective interval for the question
Authority	Owner and approved revision
Conflict	Competing definitions and reason
Use	Query field mapping and citation

Return conflicts as data

When two governed definitions fit, provide their distinguishing rules and ask the user which decision they are making. When a source is missing, say that the selection cannot be verified. A retrieval score ranks candidate text, not business authority. Keep the result linked to its original record so corrections and version changes can be traced into later agent runs.

Check the answer against the chosen rule

Inspect the generated query’s grain, filters, time zone, and joins against the retrieved definition. Execute a boundary case where competing definitions yield different results. Log the selected semantic revision with the task, while protecting sensitive query data. A successful query execution establishes syntax and access, not that the chosen meaning matches the user’s decision.

Evidence boundary for semantic retrieval

• dbt semantic models: dbt documents semantic-model elements that can be retrieved and applied. It does not resolve this fictional organization’s conflicting labels.
• W3C PROV-O: PROV-O defines relations for entities, activities, and attribution. Provenance records do not determine which definition the user intended.

The competing active-customer definitions are invented. Retrieval ranking and authorization need local implementation evidence.