Curation in the Knowledge Catalog

Letting AI generate content and put it straight into use would be risky; reviewing everything in side spreadsheets is slow and leaves no trail. Curation solves that dilemma: every AI-generated asset — optimized document, FAQ and glossary — is born as a draft and only goes into use after a person compares the content with the source excerpts, edits it if needed and approves it, with an audit trail for each decision. You get the speed of automatic generation without giving up human control.

Curation is not an accessory: it is the functional safety gate of the whole cycle. A glossary term only starts influencing chat search after it is approved, and the export of content to external systems only serves what has already been approved.

To understand how assets are generated, see the knowledge preparation guide.

Where it lives

Curation lives inside the Knowledge Catalog: menu Descobrir › Knowledge Catalog, side tab Curadoria. The tab only shows up for users with the permission to view knowledge assets (KNOWLEDGE_ASSET_VIEW).

The breadcrumb shows Knowledge Catalog › Curadoria › sub-tab name, and each sub-tab has its own address — you can bookmark the one you use most: index.html?view=catalogo&tab=curadoria&sub=fila (replace fila with fontes or indicadores).

Old links keep working. Until September 2026 curation was a screen of its own, in the Governança menu group. Addresses saved back then (?view=curadoria-conhecimento, including with the tab at the end) open the Curadoria tab on the matching sub-tab; the old Glossário-tab link opens the catalog's Glossário tab already filtered to drafts. Workspaces that had the "Curadoria de Conhecimento" feature enabled now have "Knowledge Catalog".

The three sub-tabs

Glossary terms are not decided here: they have a single home, the Knowledge Catalog's own Glossário tab (see the Glossary section below). The opening text of the Curadoria tab carries the Glossário link, which takes you straight to the list of draft terms.

Reviewing an asset (side by side)

Click an item in the queue to open the review panel:

• Left: the generated content (sections, question-and-answer pairs or terms).
• Right: the cited source excerpts — only the excerpts of the original document that ground each item. This is where you check, without leaving the screen, whether the AI really relied on the document. On narrow screens the two views become the "Gerado" and "Fonte" tabs.

Editing before approving

The pencil icon opens a text editor per item. Aplicar (prévia) only updates the screen; Salvar edição actually saves — the asset stays a draft, is marked as human-edited and keeps the same version. An amber banner warns when there are unsaved edits, and the Aprovar button stays disabled until you save: you never approve something different from what you are looking at.

Editing goes through the same source validation as generation, only stricter: every item must keep at least one citation of a real excerpt from the document — an invalid citation rejects the edit with a message explaining why.

Approving or rejecting

• Aprovar — asks for confirmation in an in-app dialog; on confirmation, the notice "Ativo aprovado. Já elegível para uso no retrieval." appears.
• Rejeitar — the dialog requires a reason (a mandatory field that "fica registrado na auditoria"); the asset becomes REJECTED and leaves the queue.

Diff between versions

When an item has a previously approved version (a re-generation after a change in the source document), the Ver mudanças link appears. The comparison is computed on the server, item by item — section, question or term, matched by key: ADICIONADO / REMOVIDO / ALTERADO / IGUAL pills, and short changed items get a line-by-line comparison (green and red). You review only what changed, instead of rereading everything.

The Glossary-type asset

The generated asset of type Glossário (the set of terms extracted from a document) still shows up in the review queue like the others: approving it is what releases that set for export. Each individual term, on the other hand, is decided in the catalog's Glossário tab — approving the term is what makes the chat start resolving it.

Batch approval

Select items in the queue with the checkboxes — the batch bar appears with Aprovar / Rejeitar / Limpar. The confirmation dialog shows the exact count and warns: "Cada item gera um evento de auditoria individual.". Batch rejection also requires a reason, applied to all items.

• Up to 200 assets per batch operation.
• One item failing does not bring down the batch: the result reports "N com sucesso, M falha(s)" with the reason for each item.

Practical example — reviewing a batch of FAQs

1. A generation batch finished for the regulations domain (you followed it in the Indicadores sub-tab). Open Knowledge Catalog › Curadoria.
2. In the Fila de revisão sub-tab, filter by type FAQ. Twelve drafts of "Portaria 1.234/2026" appear.
3. Click the first one. On the left, the pair "Quem pode solicitar o benefício? / Servidores ativos com mais de 5 anos…"; on the right, the excerpt of the regulation that grounds the answer. It checks out.
4. In one of the pairs the answer is correct but wordy: click the pencil, trim the text, Salvar edição and then Aprovar.
5. Go back to the queue, select the remaining 10 pairs you already checked and use the batch bar: Aprovar → confirm. Each approval is audited individually.
6. Within minutes the approved FAQ is eligible for chat search and for external export.

Glossary

Glossary terms — the ones created by hand and the ones automatic generation extracted from documents — live in a single place: the Knowledge Catalog's Glossário tab. Generated terms arrive there as Rascunho.

Finding the terms to decide

The list shows terms of every status, and the status filter offers Rascunho (Draft), Aprovado (Approved), Rejeitado (Rejected) and Descontinuado (Deprecated). The Glossário link in the Curadoria tab opens the list already filtered to drafts — the same as the address index.html?view=catalogo&tab=glossary&status=DRAFT.

The detail of a term that came from automatic generation shows the line "Extraído pela IA do documento", followed by the code of the document it was born from.

Approving, rejecting and deprecating

• One term: in the detail, a draft term has the Aprovar and Rejeitar buttons; an approved term has Descontinuar.
• In batch: tick the terms' checkboxes — or use "Selecionar os N termo(s) da lista" to tick everything the filter shows — and the batch bar appears with Aprovar / Rejeitar / Descontinuar / Limpar.

Decision rules:

• Up to 500 terms per batch.
• Rejeitar requires a reason, recorded in the term's history and in the audit log.
• The transition is validated term by term: a draft becomes approved, rejected or deprecated; an approved term can only become deprecated. Each term keeps the record of the decision (previous and new status, who decided, when and the reason), and the batch is recorded in the audit log.
• One term failing does not bring down the batch: the result reports "N com sucesso, M falha(s)" with the reasons.
• Approved terms start being resolved by chat within 5 minutes — that is the cache window that serves them. Deprecated terms stop being resolved.

Editing a term

The edit form changes name, definition, domain and owner. It has no status selector: the status shows read-only, with the note "Muda pelos botões Aprovar, Rejeitar e Descontinuar." Editing the definition of an approved term keeps it approved.

Users with catalog view only see the list in read-only mode; editing, approving, rejecting and deprecating require the GLOSSARY_EDIT permission.

Source decisions

Documents that vanished from the monitored source are not deactivated automatically — the decision is yours. The sub-tab lists the documents flagged as out of source in the selected domain, for human triage.

Current state: the Desativar documento button is disabled ("Disponível em breve") — the action depends on the integration with the document validity cycle, not available yet. The listing already works and allows manual triage.

Indicators

The panel is aggregated on the server and refreshes itself every 60 seconds; recent approvals show up within 1 minute, because the 60-second cache is invalidated on every approval.

Freshness per monitored source shows as "em desenvolvimento" — that data is not populated yet and arrives with the source-monitoring integration.

Permissions per action

The interface hides or disables what you cannot do, reading the permissions of your own access, but the rule holds on the server: direct calls without permission are denied with a message in Portuguese and are audited.

Default profiles: administrators and stewards (ADVANCEDUSER) approve; analysts only view. Without `KNOWLEDGEASSET_APPROVE`, the Curadoria tab shows "Você está em modo de visualização.". These operations are also available through the API reference.

Behavior and accessibility

• Confirmations in in-app dialogs (never native browser windows); Escape closes them and focus is managed.
• Success and error notices in Portuguese, auto-dismissed after 5 seconds; network errors, permission errors and not-yet-available features have specific messages and a Tentar novamente button.
• Loading with skeletons; every sub-tab has guidance when it is empty.
• Responsive layout: on mobile, the queue stacks and the side-by-side review turns into tabs.
• The tab is instrumented with frontend telemetry, which follows each action and ties it to the corresponding call on the server.

Troubleshooting