Skip to main content
Delphina analyzes your warehouse — schemas, query history, expert users’ SQL, and uploaded documents — and documents your most important tables, metrics, and business rules automatically.

​
Prerequisites

​
Step 1: Configure sources

Tell Delphina where to look. Go to Context Layer > Sources and configure the raw inputs the agent will use to build your knowledge base. Tables — Select which warehouse tables should be in scope. Click Add from Catalog to browse your schema and check the tables you want documented, or use Bulk Edit to paste table names if you have many. Query Users — (Snowflake, BigQuery, and Databricks) Select the users whose query history the agent should analyze. Click Add from History to pick from recently active users. Pick your most experienced analysts — the agent learns join patterns, filter conventions, and metric definitions from their SQL. File Uploads — Upload reference documents that describe your data: dashboard definitions (LookML, Tableau), dbt project files, data dictionaries, or any other documentation. The agent extracts metric definitions, business rules, and table relationships from these. MCP Connections — Pull documentation that lives in external tools (Notion, GitHub, Slack, and others) via the Model Context Protocol. Set a knowledge pull prompt on the connection to scope what the agent ingests — for example, “Pull all pages under the Metrics space”. See MCP Connections.

​
Step 2: Update knowledge

Start a new chat and send:
Or, to scope the update to recent source changes:
Or steer it explicitly — name the tables, domains, and sources it should use:
The agent figures out the right flow on its own: with an empty knowledge base it plans a full onboarding; with an existing one it checks what changed in your sources and scopes the update to that. In the chat, the agent works through the update with you:
  1. Orients — it checks what is already in your knowledge base, then compares your warehouse schemas, uploaded files, query configuration, and MCP content against their state at the last update, so it knows what is new or changed.
  2. Proposes a plan — how many tables, metrics, dimensions, and rules it will document, and from which sources. Review the counts and tell the agent to adjust the scope before it builds. For very large warehouses the agent may suggest onboarding in multiple passes, one domain at a time.
  3. Builds the knowledge once you confirm — table documentation, metric definitions with tested SQL, and business rules.
Progress is visible on the update history page (View History on the Sources page) while the build runs; open a run to see per-step detail. Larger warehouses take a while — you can leave the chat and come back.

​
Step 3: Validate knowledge

Knowledge updates validate what they build as they go, but before creating evaluations it is worth running a standalone validation pass over the whole knowledge base — and you can re-run it anytime, not just during onboarding. Start a new chat and send:
This checks every document against the live warehouse — schema drift, broken metric SQL, broken cross-references, contradictory definitions, and coverage gaps. Knowledge stays read-only during validation: the agent reports what it finds in the chat without changing anything. To have critical findings filed as issues in Context Layer > Inbox — each with a suggested fix — ask for it:
You can then fix the findings with Delphina: open an issue and ask it to investigate from the issue page, or remediate several at once from a /knowledge chat:

​
Step 4: Create evaluations

Start a new chat and send:
This also starts a guided session: the agent plans a set of test cases from the knowledge it just built — questions paired with expected SQL and acceptance criteria — and builds them once the plan looks right. When adding to an existing suite, scope the prompt to what’s new:
Review the candidate evaluations with your team. These validate that the agent answers questions consistently with your documented definitions. See Evaluations & Quality.

​
Step 5: Run an evaluation

Go to Context Layer > Evaluations and click New Run. All test cases are selected by default (a full run) — narrow the selection to specific cases or domains for a quicker partial run. The run page shows live progress: each case gets a pass/fail verdict with the judge’s explanation, and failed cases are analyzed automatically and land as issues in Context Layer > Inbox. See Evaluations & Quality for reviewing results and fixing failures.

​
After onboarding

The agent immediately starts using your documented metrics, rules, and tables.
  • Use /knowledge in chat for ongoing refinement as schemas change. See Maintenance.
  • Re-run the update prompt from Step 2 when expanding to a new domain or after updating sources. The Sources page shows a Sources changed since last update banner when Delphina detects drift, and the agent scopes the update to what changed.
  • Re-run the validation prompt from Step 3 periodically, or after big warehouse changes.
  • Re-run your evaluation suite after knowledge changes to catch regressions.