Security & Read-Only Controls

Before connecting a source, agree on which data the reporting identity may read and how the returned results may be used. Datarelix’s controls restrict query execution; they do not decide which business information your organisation should expose.

Query controls

The model proposes a structured plan. Backend services validate it and execute the permitted steps. SQL is parsed for read-only validation; KQL and ES|QL have dedicated validators. Connection scope restricts the source objects available to reporting queries.

Connector defaults commonly include a 5,000-row result limit and a 30-second query timeout; exact limits and configuration vary by engine. A row limit does not cap scanned data or total compute cost. Consult the connector reference for the source in use.

A validation rejection is shown as SQL validation failed. It is different from a database permission failure. Troubleshooting

Credentials and model inputs

Credentials are encrypted at rest and used by backend connection services for testing connections, schema discovery and query execution. They are not supplied to the model. Rotate the secret at the source and update the connection when needed.

The model can receive questions, schema metadata, descriptions, conversation context and database error messages used to revise a query. AI schema discovery uses structure rather than sample rows. Follow-up context can include selected values from a single-row result, including text as well as numbers. Multirow result tables are not included through that follow-up mechanism. Do not put secrets in questions or descriptions.

Datarelix stores results and run artifacts so you can view them and reopen work. Review the privacy policy for service data handling, and the data-flow guide for examples.

Generated Python

Generated analysis runs in an isolated sandbox with network egress blocked and resource limits. The sandbox processes provided results; it does not receive the database connection credentials. These controls concern that analysis environment, not every flow of information in the service. Python analysis

Effective source access

Ask the database owner to use a dedicated reporting identity and review inherited roles, public privileges, callable functions and row/column policies. Read-only validation and a scoped source identity address different concerns. Neither makes an ambiguous measure correct or a sensitive read appropriate.

For example, a sales view may omit contact details that the report does not need. Use the minimum approved source scope for the work.

Saved work and requirements

Connections, runs, conversations and dashboards are private to their creator’s account. Account checks and database access policies enforce ownership. Team pools the subscription allowances; it does not provide shared connections or dashboards.

Run history is per account. Shared reporting, an organisation-wide audit log, SCIM provisioning and a selectable processing region are not standard features of this workflow. Confirm any required Enterprise identity or contractual arrangement with Datarelix.

For a vendor evaluation, use the security checklist. For a result check, compare the dates and definition, then have a SQL reviewer inspect calculations you cannot verify yourself.

Inspect the run record

A run connects the question to its plan, executed steps and returned artifacts. Reopen it from history to inspect the statement that actually ran, source tables, returned-row count and runtime. If a failed query was revised, review the successful version before relying on the result.

This record helps an analyst reproduce a calculation and helps a business user explain the source and period behind a chart. It is per-account history, not an organisation-wide administrative audit log. Re-executing a recorded query against changed source data may produce a different result.

Distinguish a validation rejection from a permission error

A SQL validation rejection means the generated request did not meet the service’s rules. A database permission error means the source identity was not permitted to perform the attempted read. Changing a password does not repair an out-of-scope query; broadening database grants is not the default response to a validation error.

Use the error message and the connector documentation to identify the cause; Troubleshooting lists each message. Keep sensitive connection details out of support examples.