AutoLearn

What it is

AutoLearn logs questions that the agent should answer from its linked bases but for which search found no adequate material. Each record becomes a learning opportunity for human review.

The feature does not publish answers by itself or train the model. After review, it creates a text item in the selected base; that item becomes available to new searches only after index processing finishes.

Enable it for an agent

AutoLearn is managed under Knowledge, not Tools:

1. Open Agents, select the agent, and go to Knowledge.
2. Select at least one Knowledge base. This enables Query Knowledge Base and reveals the AutoLearn panel.
3. Turn on the AutoLearn switch.
4. If that agent already has pending opportunities, review up to three of them in the panel or select Review all.
5. Select Save in the changes bar.

Clearing the last base also disables AutoLearn in the same save. Saving this configuration requires the agents.write and bases.write capabilities.

When an opportunity is logged

AutoLearn is available at runtime only when both conditions are active: its own switch and search with at least one linked base.

When search_knowledge_base returns exactly “Nenhuma informação relevante encontrada na base de conhecimento.”, the internal instruction requires the model to call report_missing_knowledge before replying. The model may also call it when the excerpts clearly concern a different topic. It must not log small talk, an out-of-scope topic, or a question the agent is not expected to answer.

The call sends:

• question: required and rewritten as a generic, reusable question;
• suggested_answer: optional draft for a person to review;
• suggested_base_id: suggested linked base when one is appropriate;
• reasoning: optional reason for the classification.

Names, email addresses, phone numbers, identity documents, order numbers, and other case data must not enter these fields. This protection depends on the model following its instruction; review the text before adding it to a base.

The tool result is internal. After logging or matching the gap, the agent continues its reply normally; the customer does not receive the record’s technical status.

How similar questions are grouped

Before creating an opportunity, the runtime searches for a similar pending question for the same agent and organization. Text similarity must be greater than 0.7.

When it finds a match, the system:

• increments Occurrences;
• updates Last seen;
• adds the conversation to the source list without repeating the same ID.

Grouping does not cross agents and considers pending opportunities only. A question similar to a dismissed or approved opportunity may create a new record. Simultaneous calls may also pass the check and create duplicates, so treat the list as a triage queue rather than an exact count of unique intents.

Review the queue

Open Knowledge Bases → AutoLearn. The route shows pending opportunities only and sorts the most frequent first, using the most recent occurrence as a tiebreaker.

The table displays Question, Agent, Occurrences, Last seen, Target base, and Actions. You can:

• filter by agent;
• filter by suggested base, including No suggested base;
• export the filtered rows as CSV;
• select one or more rows to ignore;
• open Review and include for an opportunity.

The page requires bases.view. In the audited implementation, its create and dismiss controls do not apply a second visual bases.write gate; grant queue access only to roles authorized to change the knowledge collection.

Review and include in a base

1. Select Review and include.
2. Choose the Target base. The agent’s suggestion is only a starting point and can be changed.
3. Review the Question and Answer. Both fields are required to continue; the suggested answer may be only a hypothesis and must be checked against official information.
4. Remove any personal or conversation-specific data.
5. Select Create item in base.

The question becomes the new item’s name and the answer becomes its content. Although the field accepts up to 2,000 characters, the stored name is silently limited to 200 characters; keep the question short and inspect the item through the Edit item shortcut shown after creation.

The new item starts with processing pending. Item creation and opportunity resolution are two sequential writes; if the second one fails, inspect the base before retrying so you do not create a duplicate.

Ignore and troubleshoot

• Ignore: changes the status to dismissed and removes the opportunity from the pending list. The confirmation says removal is permanent, but the record is not deleted; the toast offers Undo for five seconds. After that, the current interface has no history from which to restore it.
• No opportunity appears: confirm that a base is linked, AutoLearn is on, and the agent changes were saved. Logging still depends on the model calling the tool in a gap state.
• An existing question was not grouped: grouping uses text, agent, and status. Different wording, a different agent, or an already resolved item may produce another row.
• The answer is still not found: open the created item and confirm that index processing finished without an error. Approving an opportunity does not make its content searchable immediately.
• Logging error: failures while finding similarities, updating, or inserting make the tool return an internal error to the model. The conversation continues, but the opportunity may not have been saved.