Learning Opportunities (AutoLearn)

AutoLearn creates a queue of questions that deserve human review because the knowledge-base search did not find relevant content. It helps you discover gaps, but it does not learn from them or publish answers on its own.

The queue is under Admin → Knowledge Bases → AutoLearn. The tab is always available to people who can access Knowledge Bases; when pending opportunities exist, a counter appears next to AutoLearn.

Before you start

In the agent editor, open Knowledge and confirm that:

1. at least one base is selected under Knowledge Bases;
2. Log unanswered questions is enabled under AutoLearn;
3. the changes were saved.

Search and logging work together. If no base is selected, both native tools are inactive. See Connecting Bases to Agents to configure and test retrieval.

When an opportunity is logged

The runtime instructs the agent to log an opportunity when Search Knowledge Base returns the exact no-relevant-results response or when the retrieved excerpts are clearly about another subject. The tool call still depends on the model’s decision, so not every insufficient answer will necessarily appear in the queue.

The record may contain:

• a generic, reusable question, which may differ from the conversation’s original wording;
• a suggested answer;
• a suggested base;
• an internal reason for the suggestion;
• the agent, associated conversations, occurrence count, and last occurrence.

The tool prompt asks the model to remove names, email addresses, phone numbers, Brazilian CPF identifiers, order numbers, and case-specific details. There is no second deterministic sanitizer before storage. Treat the question and answer as untrusted content and remove personal data before approving or exporting them.

How grouping works

While an opportunity is pending, another similar question from the same agent and organization may increment its count. Matching uses textual similarity above a threshold; it does not require identical text and does not combine opportunities from different agents or already resolved records.

Use the count to prioritize, not as an exact number of people. Concurrent calls can also produce duplicate rows. Compare the content before creating a new item.

Understand the queue

The screen shows pending opportunities only, ordered first by Occurrences and then by Last seen. Each row contains:

• Question;
• Agent;
• Occurrences;
• Last seen;
• Target base, when one was suggested;
• Ignore and Review and include.

Use the All agents, All bases, or No suggested base filters. Changing a filter clears the current selection. The reason field is stored by the product but does not appear in the table or review modal.

If the screen shows No pending learning opportunities, refresh the page and confirm connectivity before concluding that the queue is empty: a loading failure can also fall through to this empty state.

Review and include in a base

1. Select Review and include.
2. In the modal with the same name, verify the Target base. The suggestion is used when available; otherwise, the product fills in a base from the list. Change it when necessary.
3. Rewrite the Question as a short, generic, searchable title.
4. Validate the Answer. Correct facts and remove personal data, references to a specific conversation, and instructions that should not apply to other customers.
5. Select Create item in base.

Question, answer, and base are required. The text fields accept up to 2,000 characters, but the created item’s name uses only the first 200 characters of the question, cut at a word boundary. The interface does not warn about this truncation, so keep the question short.

The item starts with Pending status and enters embedding processing. The success toast provides an Edit item shortcut. Wait for the item to become Processed before expecting it to appear in the agent’s search.

If approval reports an error

Item creation and opportunity resolution are not a single transaction. Before retrying, open the selected base and look for the item: it may have been created even though the queue still shows the opportunity.

Ignore and undo

Use Ignore for noise, spam, out-of-scope subjects, sensitive content, or a question that should be addressed by improving an existing item. You can also select multiple rows and choose Ignore selected.

The confirmation says removal is permanent, but the current operation changes the status to dismissed. The toast offers Undo for five seconds. After that, the interface has no history view or restore control; dismissed and approved opportunities leave the pending queue.

Export the queue

Export CSV downloads only pending rows matching the current filters. The gaps-YYYY-MM-DD.csv file uses semicolons and includes the question, agent, occurrences, raw last-occurrence timestamp, and suggested base.

Before opening the file in Excel or another spreadsheet:

• treat it as untrusted data;
• remove personal data and customer-specific content;
• import columns as text when possible;
• do not distribute it as a complete backup, because resolved and dismissed records are not exported.

Values beginning with =, +, -, or @ are not neutralized by the exporter and may be interpreted as formulas by spreadsheet software.

Permissions and a safe review routine

The route requires View bases, but the current implementation may also expose controls that change the queue and create items to read-only profiles. Until authorization is corrected, limit triage to people authorized to edit content and keep an internal review step for every inclusion.

A safe routine is:

1. filter by agent and prioritize frequent occurrences;
2. search the base for an equivalent answer before creating content;
3. review facts, scope, privacy, and target base;
4. approve or ignore;
5. wait for Processed and test the question with the agent;
6. if the gap returns, improve the existing item, its terminology, or the search instruction before duplicating it.