Cleanup Studio
Profile a dataset, build a governed recipe, validate every record and apply it to metrics without changing source records.
Open Data+ → AI Cleanup Studio (/app/data-plus/cleanup-studio). Cleanup Studio is the only Data+ transformation workflow. An approved recipe changes the values stored metrics read. Source records are never modified.
Profile
Select a dataset to inspect up to 500 records. The profile shows inferred types, missing values, distinct counts, key candidates, duplicates and outliers. Sample counts are marked as samples and are never extrapolated.
Check cross-dataset references compares a sampled field with a key in another stored dataset you can read. If the reference dataset has more than 2,000 records, the check reports not evaluated rather than guessing at unmatched keys.
Build a recipe
A recipe holds up to 100 rules that run in order, and each rule reads the result of the one before it, so one field can have several steps.
| Group | Rules |
|---|---|
| Text | Trim, collapse spaces, uppercase, lowercase, normalize identifier, keep digits |
| Brazilian identity | CPF, CNPJ, CEP, phone |
| Contact | |
| Values | Map values, treat tokens as empty |
| Numbers and dates | Parse number, currency or percentage with explicit separators, parse date with an explicit format, BRL amount, Brazilian date |
| Checks | Validate a pattern |
| Columns | Split, merge, derive |
Most rules can write to a new column instead of replacing the original. Patterns must be anchored with ^ and $, and can't use groups, alternation or lookaround.
If any rule fails on a record, the whole record is quarantined and keeps its original values.
AI suggestions use field names and aggregate counts only, never source values, and a steward approves every rule. Rules that change identity, money, dates or meaning need a written reason.
Validate and publish
- Preview shows before and after values for the sample.
- Validate now checks every record of a stored dataset in batches of 200. You can pause and resume from the same point. No transformed row is stored.
- The run reports changed and quarantined counts, the reasons for quarantine, and quality before and after.
- Approve and publish creates an immutable version. Publication is refused if the source or the rules changed after validation.
Live sources have no stored rows to validate, so they are verified by a complete read of up to 10,000 records whenever a metric runs. Above that, the query is refused.
Rollback previews an earlier version and publishes it as a new version, after validating it the same way. History is never rewritten.
Access
Steward operations need the Data+ stewardship capability and read access to the source. Sources authorized by one user's own credentials can't be used here.
API and MCP
First-party steward operations:
GETandPOST /api/app/data-plus/cleanup-studio: datasets, state, preview, suggestions, publish and rollbackGETandPOST /api/app/data-plus/cleanup-studio/validation: run history,startandadvanceGET /api/app/data-plus/cleanup-studio/relationships: reference check
Public governed reads only: REST GET /api/v1/data-plus/cleanup-profile, POST /api/v1/data-plus/cleanup-preview, GET /api/v1/data-plus/dataset-profile?datasetId=<UUID>, and MCP blocchi.get_cleanup_profile, blocchi.preview_cleanup_recipe and blocchi.get_dataset_profile. The preview takes every rule except value mapping and returns counts only. Publishing stays in the studio. Raw rows and text examples are withheld.
See Metric readiness and Stewardship.
Last updated on