> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bubblav.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Self-Learning

> Let your chatbot learn new answers from real conversations — you review, it improves

**Self-Learning** closes the loop between your conversations and your knowledge base. When a chat ends in a way your bot couldn't fully handle — but the answer surfaced anyway — BubblaV drafts a new Q\&A entry for you and drops it in a review queue. Approve it, and your bot answers that question correctly next time.

<Frame>
  <img className="block dark:hidden" src="https://mintcdn.com/bubblav-e553cf80/nhelQxnCmiOq_LRR/images/knowledge-base.png?fit=max&auto=format&n=nhelQxnCmiOq_LRR&q=85&s=2cd5b14fc8a00d2e6d9f458acee32ca7" alt="Self-Learning Overview" width="1067" height="383" data-path="images/knowledge-base.png" />

  <img className="hidden dark:block" src="https://mintcdn.com/bubblav-e553cf80/nhelQxnCmiOq_LRR/images/knowledge-base-dark.png?fit=max&auto=format&n=nhelQxnCmiOq_LRR&q=85&s=b44059eb3c45b15e108d600c6eb00155" alt="Self-Learning Overview" width="1055" height="378" data-path="images/knowledge-base-dark.png" />
</Frame>

***

## Where learned answers come from

BubblaV watches three kinds of conversations for new knowledge:

<CardGroup cols={3}>
  <Card title="Human agent answers" icon="user">
    When a live-support chat is resolved, your agent's reply becomes a suggested Q\&A — exactly the answer a real customer needed.
  </Card>

  <Card title="Bot self-corrections" icon="rotate">
    Chats where the bot stumbled (low confidence or fallback replies) but recovered by the end become improvement candidates.
  </Card>

  <Card title="Thumbs-down threads" icon="thumbs-down">
    When a visitor rates an answer down, the thread is mined for what should have been said — by your agent, or by the bot itself.
  </Card>
</CardGroup>

Drafts are created by an AI model reading only the conversation transcript — no invented facts, capped length, and questions phrased the way visitors actually ask them. Questions you've dismissed in [Content Gaps](/user-guide/knowledge/content-gaps) are never re-proposed.

## Reviewing suggested answers

Learned entries land in **Knowledge → Q\&A** with a **Suggested** badge and a "Learned from conversation" note:

* **Edit** the question or answer first — useful for tightening an agent's phrasing.
* **Approve & publish** to make it live. The answer flows through the standard Q\&A pipeline, so it's immediately retrievable by your bot.
* **Discard** to delete a draft. Nothing goes live unless you approve it.

<Note>
  Publishing a learned answer counts toward your knowledge page limit, exactly like a manually added Q\&A entry.
</Note>

You'll get a bell notification when new suggestions are waiting. Manage it in [Account Settings → Notifications](/user-guide/account-settings).

## Auto-apply (optional)

By default every learned entry waits for your review. On the **Pro** plan you can enable **Auto-apply learned answers** in **Settings → General**:

* Only **strong-evidence** candidates are auto-published — typically answers written by a human agent on a resolved chat.
* Weak-evidence drafts (bot self-corrections, unanswered thumbs-down) still wait in the review queue.
* The toggle is **off by default**, and auto-published entries still respect your knowledge page limit.

***

## Plan availability

* **Review queue** (drafts + approve/discard): Pro and higher.
* **Auto-apply**: Pro and higher, off by default.

***

## Next Steps

<CardGroup cols={2}>
  <Card title="Q&A" icon="message-circle-question" href="/user-guide/knowledge/q-a">
    Manage suggested and manually added question/answer pairs.
  </Card>

  <Card title="Content Gaps" icon="sliders" href="/user-guide/knowledge/content-gaps">
    See what questions customers are asking that you haven't answered yet.
  </Card>
</CardGroup>