Skip to main content

What's in a context repository

TL;DR

A Context Engineering Studio repository holds linked assets that reference your governed Atlan context, and a semantic model that says how those assets answer questions. This page covers both, and how to scope a repository to one use case.

Connect

A context repository holds two things: the Atlan assets it points to, and the semantic model that says how those assets answer questions.

Everything else an agent needs already lives in Atlan. The repository doesn't copy it.

This page covers both parts and how to scope a repository. For why agents need business context and how the workflow runs, see What is Context Engineering Studio?.

Scope to one use case​

Scope a repository to the job an agent has to do, not to a team or a domain.

"Answer pipeline questions for the sales ops team" is a use case. "Sales" is not. The second sounds tidier and produces a worse agent: more tables than any one question needs, more competing definitions of the same word, and an evaluation set too broad to tell you what's broken.

A narrow repository is easier to test, easier to correct when a question fails, and easier to reason about when two teams define revenue differently.

When a new set of questions comes up, such as marketing asking about campaign performance, build a second repository for it. Don't add those tables to the first.

Linked assets​

Linked assets are the Atlan assets the repository points to: tables, columns, and glossary terms. A linked asset can also be a skill, a reusable capability that lives in Atlan like any other asset, for the rarer use case that needs a procedure or a deterministic calculation rather than just a definition. Skills aren't carried into a deployment, because Snowflake semantic views and Databricks Genie don't support them. The Claude harness can retrieve them through the Atlan MCP when a use case needs them.

Each one carries a Reference badge. The repository links to the asset in Atlan for provenance rather than storing a copy, which has two consequences worth knowing:

  • The source stays governed where it lives. Editing a description or glossary term in Atlan is still the way to change it.
  • Regenerating the model pulls the current state of each linked asset and keeps the edits you made by hand. A change at the source flows in when you regenerate, not before.

Context Agents generate and maintain this layer across your estate. CES assembles the slice one use case needs. See Context Agents Studio.

Semantic model​

The semantic model is what CES assembles on top of those references. It's the part that doesn't exist anywhere else, and the part you edit.

Tables Each table in scope, with the dimensions, time dimensions, and facts an agent can query.

Relationships How tables join, so a question spanning two of them resolves correctly.

Derived metrics Measures built on top of your facts and glossary terms, such as weighted pipeline value. Some engines can't accept every metric you define. See Metrics that aren't emitted.

Named filters Default conditions to apply, such as excluding test accounts or internal orders. A filter not yet attached to a table shows as unassigned.

Verified queries Question-and-SQL pairs that ground the engine in answers you've confirmed are correct.

Custom instructions Business rules and SQL-generation guidance in your own words, alongside the model description. This is where knowledge that isn't a column or a metric goes: which source to prefer, what a fiscal quarter means, when to escalate rather than answer.

How agents use it​

Ask "What was net revenue last quarter?" and the agent resolves the question against the repository:

  1. Net revenue and last quarter resolve through the semantic model, which supplies the measure, the joins, and the default filters. The linked glossary term supplies the certified definition.
  2. Your custom instructions apply the rules that aren't in the model, such as excluding intra-company transfers.
  3. The query runs on the connected engine, scoped to the tables in this repository.

Because the repository defines all of this in one place, every engine reading it answers the same way.

See also​