Convi's assistant doesn't ship knowing your store. It learns automatically the moment you install it, and you keep teaching it as it runs. This article maps the full learning loop — what's imported automatically, what you add yourself, and how the system improves itself from real shopper conversations.
In this article
The five sources of knowledge
Auto-imported sources (read-only)
Sources you author (editable)
Sources Convi generates from real conversations
The Train section, at a glance
The improvement loop
A few rules the AI follows
The five sources of knowledge
Convi's brain is the union of several knowledge sources. Each one shows up under Train → Content with its own row in the Source column:
Source | Origin | Who controls it |
Policies | Auto-imported from Shopify (Returns, Shipping, Privacy, Terms of service) | Shopify (edit there) |
Pages | Auto-imported from your Shopify pages | Shopify (edit there) |
Blogs | Auto-imported from your blog articles | Shopify (edit there) |
Shop info | Auto-imported on install — store name, location, currencies, shipping destinations, payment methods, capabilities | Shopify (edit there) |
Custom FAQs | Q&A you author in Convi | You |
Imported from URL | Public pages Convi extracts and saves into your KB | You |
Content gap | Real shopper questions Convi couldn't answer well | System (you turn each into an FAQ) |
Enhanced chat | Chats Convi processed into reusable knowledge | System |
Handed over | Conversations that escalated — used as a learning signal | System |
Auto-imported sources (read-only)
Four sources are managed by Shopify: Shop info, Blogs, Pages, Policies. You see them in Convi but you cannot edit them in Convi. Each row in the table gives you an Update in Shopify button that deep-links straight to the right Shopify settings screen (e.g. for a policy, Settings → Legal).
Why read-only? Because Convi re-syncs them from Shopify on a schedule. If you edited them inside Convi, your edits would be overwritten on the next sync. Editing them at the source means the change actually sticks.
Convi refreshes auto-imported content on a daily background job — you don't have to trigger anything.
Sources you author (editable)
Two sources are yours to author:
Custom FAQs — short Q&A pairs you create under Train → Content (or via the Add answer button on a knowledge gap in Analyze → Recommendations).
Imported from URL — paste a public URL and Convi extracts the content into a knowledge entry you can edit before saving.
Use these for anything your imported Shopify content can't cover: bundle-specific refunds, brand-specific FAQs, custom shipping rules, return-window exceptions.
Sources Convi generates from real conversations
Three sources are populated automatically as your assistant runs:
Content gap — every time the AI couldn't answer a shopper question confidently, the question is logged here. These are the questions you'll see surfaced under Analyze → Recommendations with one-click "Add answer".
Enhanced chat — processed snippets of past conversations that became reusable knowledge.
Handed over — conversations that escalated to a human, used as a learning signal so the AI knows where its current edge is.
You don't author these. You react to them — the most useful one is Content gap, because every entry there is a real shopper question you should answer.
The Train section, at a glance
The whole Train area is the merchant-facing surface of the learning system. Five tabs:
Tab | What it teaches Convi |
Intro | Your training health overview — coverage, freshness, quality, custom sources. |
Content | The knowledge base itself — every source, every row, with Add knowledge and Import from URL actions. |
Persona | Identity (name + avatar), voice (tone + answer length), and free-form advanced instructions. |
Abilities | The capabilities Convi can use during a conversation (search, order tracking, handover, etc.). |
Rules | Plain-English rules Convi follows in every chat. |
The improvement loop
The loop you'll fall into once Convi is live looks like this:
Watch knowledge gaps. Open Analyze → Recommendations weekly. Each item is a real question Convi couldn't answer.
Add an FAQ. Click Add answer — Convi pre-drafts an answer you can edit before saving.
Test it. Click Test to open the question in Evaluate → Playground and verify the new FAQ actually fixes the answer.
Rules for behavior, FAQs for facts. If the gap was about how Convi should respond (tone, escalation, conditions) rather than what it should say, add a rule under Train → Rules instead.
Re-check Pre-flight. Open Evaluate → Pre-flight and re-run any failing tests after you've added knowledge.
That's the full loop. Most teams run it once a week.
A few rules the AI follows
These are non-negotiable, regardless of how you configure Convi:
No fabrication. If Convi doesn't have the answer in your knowledge base or in a tool it can call, it says so — and offers handover.
No invented prices, stock, shipping times, or URLs. Prices come from Shopify in real time.
Mutating actions need verification. Order cancellations and shipping-address edits go through a confirmation step before anything changes in Shopify.
Customer-supplied content is treated as untrusted. Anything sourced from a shopper (reviews, notes) is sandboxed when handed to the model — it can't smuggle instructions to the AI.