Learned Q&A library
The AI gets smarter from your past conversations. High-confidence Q+A pairs get promoted into a library it can lean on.
Learned Q&A library
The Learned Q&A library is how Ochre's AI gets smarter from your real work. High-confidence question-and-answer pairs from past conversations and reviewer feedback get promoted into a library that future tickets can draw from.
It sits next to The brain: KB graph. The brain is your authored docs. The library is your earned answers.
Pairs land in kb_learned_qa and pass through a quarantine workflow before the AI uses them on live tickets — a typo in a promoted answer cannot leak into a customer reply before a human sees it.
What gets promoted
A Q+A pair is eligible for promotion when:
- The customer's question and the agent's reply are clear and self-contained.
- The customer was satisfied (positive CSAT, no follow-up reopen).
- A reviewer in Quality assurance review marked it high quality, OR the AI's draft was sent with no human edit and got positive CSAT.
Pairs that fail any of those checks are not promoted.
How promotion works
Two paths:
- Automatic. A scheduled job scans past conversations. Conversations meeting the bar are surfaced as candidates with the question, the answer, and the topic.
- Manual. A reviewer can promote any conversation directly from the QA review screen.
You see the candidates on AI → Brain → Learned Q&A under Pending. Approve or reject each one. Approved pairs go into quarantine for a short hold, then become live.
Auto-approval is available but off by default. Most teams prefer human review.
How the library is used
When a new ticket arrives, the AI pulls from two sources:
- The brain (your articles).
- The Learned Q&A library (relevant past pairs that have cleared quarantine).
Both are weighed for relevance. If a learned pair is more relevant than any article, the AI uses it. The reply still gets shown with a citation, in this case to the originating conversation.
Why this matters
A lot of support knowledge never makes it into a doc. Edge cases, workarounds, "actually you have to click X first" tips. The library captures those without you writing 200 articles for them.
It also means your AI improves passively. The first month is rougher. By month three, it has a meaningful library of your team's real answers.
Editing and removing pairs
The library is editable. You can:
- Edit the answer to clarify language.
- Add tags.
- Remove a pair entirely (if it became wrong, e.g. a feature was removed).
- Promote a pair to a full KB article.
A library pair that you find yourself referencing often is a sign you should write a real article. Library pairs are a stopgap. The brain is the canonical source.
When to remove a pair
- The product changed.
- The customer's situation was unusual and the answer should not generalize.
- Multiple newer pairs contradict it.
- A reviewer flags it.
Removing a pair is hot. It stops appearing on the next AI request.
What the library is not
- A replacement for your KB. The brain is still your foundation. The library augments.
- A training step. We do not retrain models on your data. We retrieve from your data.
- A free-form note system. Internal notes, runbooks, and team wikis live elsewhere.
Privacy
Pairs in the library are scoped to your workspace. They never leak into another customer's AI. The original conversation context is preserved (you can click through to the source thread from any pair).
If a customer asks for deletion under GDPR, pairs derived from their conversations are deleted too.
How it interacts with the brain
When the AI cites a learned pair on a reply, the receipt shows that. It looks like a normal article citation, but with a "learned" badge.
The graph view in The brain: KB graph does not include learned pairs by default; the library has its own list view on the same surface.
Watching library quality
A few things worth tracking:
- Library coverage. Percent of AI replies that drew on at least one learned pair. Higher = more leverage from your past work.
- Library citation rate. Of those, how often the cited pair was used as the primary answer.
- Edit rate on drafts that cited a learned pair. High edit rate means the pair is stale or wrong; remove it.
Recommended setup
- Auto-promotion: candidates only (off by default).
- Review the pending queue weekly.
- Approve aggressively. You can always edit or remove later.
- Promote frequently-referenced library pairs to real articles every quarter.
Related
Was this article helpful?