Core Concepts
Datarelix answers a question by writing a read-only query against one of your connections, running it, and showing you the result together with the query that produced it. The concepts below explain what each part is, what you can change, and where the limits are.
Connection
A connection is a saved, read-only link to one database, warehouse or set of files. It holds the details you entered in the connection form; credentials are encrypted when stored and are never sent to the AI model.
- Every conversation and dashboard uses exactly one connection; questions in it run against that connection only.
- Connections are private to your account.
- You can pause a connection to free a slot under your plan without deleting it; see Billing & plans.
Engine-specific fields, permissions and network requirements are in the connection references.
Schema context
Schema context is what Datarelix knows about your tables: names, column types, keys, relationships and descriptions. It is how a question such as “paid orders by region” gets matched to the right tables and columns.
- Discovery reads the structure of the schemas you selected. It is free and you can rerun it whenever the database changes.
- AI schema analysis adds descriptions and suggests relationships. It counts against the monthly allowance in your plan.
- On the connection’s schema page you can edit descriptions, mark key tables, confirm or dismiss relationships, and choose which tables and columns are in scope.
Better descriptions help Datarelix interpret a question; they do not guarantee the right calculation.
Conversation
A conversation is a thread of questions on one connection. Follow-up questions can refer to earlier answers (“now split that by month”). You can edit or rerun one of your recent messages; the conversation continues from that point and later turns are replaced. The question guide covers follow-ups in detail.
What each answer shows
For every question that runs, Datarelix shows:
- the result table;
- the query that ran (SQL, or KQL / ES|QL for Kusto and Elasticsearch);
- the source tables it read;
- the row count and runtime.
Use these to check the answer: confirm the dates, filters and definition, and compare a total you already know. When a question needs a calculation SQL cannot express well, the answer can include a Python analysis step working on the query results.
Before a query runs, Datarelix checks that it only reads data. If a generated query fails, Datarelix revises it and retries a small number of times before showing you the error. Troubleshooting explains each error message.
Charts and dashboards
Ask for a chart in the question, or ask for one on an existing answer. Charts are interactive and can be restyled in the conversation or with the style editor. Add a chart to a dashboard to keep it: dashboards are private to your account, use one connection, and refresh when you choose. See Dashboards.
Limits on each question
Every question runs within fixed limits:
| Limit | Value |
|---|---|
| Query run time | 30 seconds |
| Rows returned per query | Up to 5,000 |
| Python analysis run time | 90 seconds |
| Python analysis memory | 512 MB |
A row limit caps what is returned, not what the database scans. Some engines apply their own limits too; check the connection reference for yours. Monthly allowances — questions, active connections and AI schema analyses — depend on your plan and are described in Billing & plans.