Skip to main content

Customer panel

The panel is at app.datafala.com.br/login. This is where your tenant's administrator prepares everything the API uses to generate SQL. The schema and context go in only through the panel; the /v1 API does not change configuration.

ScreenWhat it is for
HomeSummary of the setup, questions from the last 30 days and end-user usage (active in the month, band and previous months).
SchemaDialect (PostgreSQL, MySQL or Firebird 2.5+) and DDL. Shows the recognized tables, columns and foreign keys; lets you describe and hide each one. For Firebird, paste the output of isql -x in SQL dialect 3.
ContextBusiness rules and glossary in free text.
ExamplesQuestion → SQL pairs (up to 200), validated by the SQL Guard when saved.
PlaygroundTest questions with the same pipeline as /v1/ask, seeing the rejection reasons without masking.
HistoryRecent questions with status, rejection reason, the user's rating and comment, the prompt sent to the LLM and the "Promote to example" button.
EvaluationUp to 50 question → expected SQL cases, run in batch to measure quality after changing the schema, the context or the model.
CredentialsCreate, list and revoke API keys.
SettingsYour LLM provider, model and key (with a test before saving), the customer name, the default language and the panel theme.

How to improve the answers​

  1. Describe what is not obvious. A column st with values A, C, P needs a description ("status: A = active, C = cancelled, P = pending").
  2. Hide what must not be used. Personal or internal data columns do not go into the prompt and are rejected by the Guard.
  3. Write rules in the context. "Revenue is the sum of valor_total for sales with status <> 'C'." Do not put secrets there.
  4. Add examples for the most common questions and for the ones that went wrong. A 👍 feedback in the History becomes an example with one click.
  5. Build an evaluation with the questions that matter and run it after every change.

Updating the schema​

Pasting a new DDL creates a new schema version. Descriptions and hidden markings are copied to the tables and columns with the same name. Answers already stored in the history are not revalidated against the new schema.