Skip to main content

Create your context repository Private Preview

Connect docs via MCP

Describe your domain in the Context Engineering Studio (CES) chat window and the agent does the rest: it searches your Atlan data graph, identifies relevant assets, and bootstraps the context repository, including a semantic model, skills, and verified queries. You can refine the result through chat at any point.

Before you begin​

  • Context Engineering Studio is enabled on your workspace and your team has the CES persona. See Enable Context Engineering Studio.
  • You have identified a business domain and a use case to start with. Keep the scope narrow. A good first repository covers a single dashboard or a small set of closely related datasets. Expand to adjacent domains as separate repositories once you've validated the first.

Create context repository​

  1. In the Context Engineering Studio chat window, describe your domain and use case. The more specific you are, the better the output. A good prompt covers:

    • Tables to include or exclude
    • The types of questions the agent must answer
    • Who uses it, such as a finance analyst or sales ops team
    • Deployment target: MCP, Snowflake Cortex Analyst, or Databricks Genie

    Example 1, high-level:

    Create context for agents supporting the sales domain.

    Example 2, specific:

    Create context for a finance analyst to answer revenue and budget questions
    using the invoice, budget_actuals, and cost_center tables. Exclude
    staging tables. Deploy to Databricks Genie.

    CES searches your Atlan data graph, selects the most relevant assets, and starts building the repository.

  2. Once CES generates an initial repository, review what was included in the right pane and refine through the chat window based on your use case.

    Example:

    Add the regional_performance_summary table. Exclude archived_transactions.

    You can refine at any time: add or remove assets, update metric definitions, rename dimensions, or ask the agent to rewrite business logic.

  3. Once complete, your repository appears in the right pane with the following structure:

    • skills/: Agent skill files (for example, SKILL.md) containing the domain description, key metrics, and agent instructions.
    • semantic_models/: The core semantic model in YAML, including metrics, filters, joins, and relationships inferred from your catalog, BI lineage, and query history.
    • verified_queries/: Ground-truth question-to-SQL pairs that give agents a reference set of correct answers.
    • quality-report.md: Eval scores per dimension with specific fix suggestions. Review this before deploying.
    • simulation_config.json: A test question set generated from your data, used to run evals.

    Deployment artifacts (Snowflake Cortex semantic views, Databricks Metric Views) are added to the repository when you deploy.

Next steps​