Training & knowledge gaps

Every store has questions the AI can’t confidently answer on day one. The Knowledge Gaps inbox is the workflow that closes those holes: Zubby flags weak-answer moments as they happen, you draft a fix with one click, and the fix becomes part of the knowledge base once you approve it and its embedding job completes. Improvement still has to be verified with representative questions and citations.

Where it lives

The AI Training section of the sidebar has two surfaces that matter here:

AI Quality (AI Training → AI Quality) - the readiness overview: products and variants indexed, knowledge docs loaded, descriptions missing, failed syncs, open gaps, feedback items, and captured leads, plus automation recommendations for the next highest-leverage training move.
Knowledge Gaps (AI Training → Knowledge Gaps) - the triage inbox this page is about.

The same section also holds Locked Facts (verbatim merchant facts), FAQ, Enrichment (the AI’s offline product research), Product Tuning, and the AI Improver inbox.

What lands in the gaps inbox

Detected knowledge gaps - shopper prompts the live AI flagged as likely weak-answer moments, each with its page context (page type, URL, and the product the shopper was viewing).
Merchant feedback - replies your team labeled wrong, incomplete, or needs handoff while reviewing conversations. The inbox counts these separately so you can see policy problems vs. coverage problems at a glance.

The triage loop

This is the whole self-learning cycle - gap, draft, review, embed:

Open AI Training → Knowledge Gaps. Each gap card shows the shopper’s question, when it happened, and the page and product context it happened on. Why it matters: these are real questions from real shoppers - fixing them beats guessing what content to write.
Click Draft answer with AI. Zubby writes a knowledge-base draft grounded in your existing documents - it is saved as suggested, not live, so nothing reaches shoppers yet. (Click Open conversation first if you want the full thread before drafting.)
Go to Catalog → Knowledge. Drafts appear in the AI-suggested articles panel at the top.
Approve publishes the article and queues its embedding job. After that job completes, test a relevant question and inspect its citations before assuming the article is being retrieved. Discard deletes the draft. Do not approve an answer that still needs a factual correction. If the draft cannot be edited before approval, discard it and create a reviewed Knowledge article instead; approving first can queue it for indexing.

Nothing ships without you

AI-drafted answers are never used live until a human approves them. The draft is grounded in your existing knowledge - if your documents don’t contain the answer, expect a thin draft; that’s your cue to write the entry yourself.

Reading the AI Quality metrics

The AI Quality page tracks the pipeline end to end: product import (products and variants available to the AI), knowledge grounding (documents loaded), and shopper learning (recommendation events feeding outcome data back into ranking). A rising Knowledge gaps count means the inbox is backing up - open it and clear the top items. Descriptions missing is the best worklist for the AI Catalog Improver.

Best-in-class workflow

Triage the gaps inbox once a day for the first two weeks. After that, a weekly cadence is enough.
Approve only answers supported by authoritative merchant sources. A partial or speculative answer can be worse than a clear handoff, and the AI can’t cite content that doesn’t exist.
Watch the merchant feedback queue for wrong labels - those are usually a stale document to fix, not a missing one to write.
Promote hard facts (return windows, warranty terms) to Locked Facts so the AI quotes them verbatim instead of paraphrasing.

Adjacent

Example: repair a missing shipping answer

A shopper asks about weekend dispatch and the draft guesses “daily.” Check the actual dispatch policy. If it says weekdays only, reject the unsupported draft and add a reviewed article with that rule. Wait for indexing, then ask both a weekday and a weekend question. Confirm the answer and citation before marking the gap solved. An approved article is evidence to test, not proof that every future answer is correct.