Knowledge conflict detection
Find the facts your chatbot's sources disagree on
When an old page says 14 days and a newer Q&A pair says 30, your agent could give a visitor either one. The Conflicts card on the Knowledge tab compares passages from different sources, lists the prices, dates, policies and contact details they disagree on, and shows both sides next to each other so you can fix the wrong one.
Visitor: How long do I have to send an order back?
Agent: You can return unworn items within 30 days of delivery. Start the return from the link in your order confirmation email and print the free label.
Visitor: A page on your site says 14 days, though.
Agent: Thanks for checking. Our returns policy gives you 30 days from delivery, so you are well within the window.
More sources, more chances to contradict yourself
A chatbot's knowledge comes from many places: website pages, uploaded price lists and manuals, Notion pages, Google Drive files, text snippets and Q&A pairs. Each is updated on its own schedule. A shipping fee changes in a Q&A pair while last year's PDF still has the old one; the opening hours move while a forgotten page keeps the winter times. The agent then finds two passages that disagree and has to choose, and it may choose the old one. Reading hundreds of sources by hand for every fact isn't realistic. Conflict detection does that reading for you and lists the pairs that disagree, so you fix the source instead of hearing about it from a confused customer.
Why check your knowledge for conflicts
Many wrong answers start in your own content, not in the AI.
Sources compared with each other
intoCHAT looks for passages from different sources of the same agent that cover the same thing, and asks an AI model whether they state different facts. Passages within one source and your instructions are not compared.
Focused on the facts that matter
It looks for disagreements about prices, fees and discounts, opening hours, dates and deadlines, policies such as returns, refunds, shipping, warranty and cancellation, numbers and limits, and contact details, addresses and names.
Both sides next to each other
Each finding shows Source A and Source B: the source and its type, the fact as that source states it and the passage it comes from, with a suggestion for what to change. A web page links to the page itself.
It runs by itself
A check runs when a training run finishes, again once a day for agents with new or changed sources, and whenever you click Check now. You do not need to remember to audit your knowledge.
Findings close when you fix them
When a passage in a finding changes and is trained, or one of the two sources is deleted, the finding is resolved without a click. The new text is compared again, so a fix that still disagrees shows up as a new finding.
How to find and fix conflicting facts
Train your agent
Add your sources on the Knowledge tab and click Train agent. On Pro and Enterprise, the sources a training run trained are compared with the rest of your knowledge as soon as the run finishes.
Open the Conflicts card
The card's title shows how many findings are open. Click Check now to start a check right away: findings appear within a minute or two, and the card updates while the check runs. Last checked shows when the last check ended.
Fix the source that is wrong
Decide which fact is right, then click Open source A or Open source B under the wrong one. Edit a text snippet or Q&A pair and train, update a web page and use Refresh from website, upload a corrected file under the same file name, or update a Notion page and import it again. A source you no longer need can simply be deleted.
Dismiss what isn't a conflict
If two facts don't really disagree, for example prices for two different plans, click Dismiss (not a conflict). The same two passages are never flagged again unless one of them changes. Mark resolved covers a fix you made some other way.
Conflict detection details
Three kinds of findings
Contradiction: both can't be true at once. Outdated: one passage looks like an older version of the other, and the finding says which source is older, from the sources' dates and any dates in the text. Inconsistent duplicate: both describe the same thing, such as a Q&A answer and a web page, but some details differ.
Different wording is not a conflict
The model is told to leave out passages that only differ in wording, where one simply gives more detail than the other, or that are about different products, plans, locations or periods.
Each pair is read once
A pair of passages is checked once, and again only when one of the two passages changes, so later checks only look at what is new. One check runs at a time per agent; if you train during a check, it goes round once more when it finishes.
Priority as a stopgap
Until you fix a conflict, mark the source with the right fact as priority with the star in the source list. It then wins when sources disagree.
Large knowledge bases, covered over time
A check reads up to 60 pairs of passages, and an agent up to 240 a day, the most recently trained sources first. Each passage is compared with its 4 closest passages from other sources, and only with passages that are very close in meaning.
No messages used
Checks don't use your plan's messages. Check now can be used 4 times an hour per agent, and the daily check starts around 02:30 UTC.
Visible to your whole team
Everyone on your team can see the card. Editors, admins and the owner can resolve, dismiss and use Check now. Open, Resolved and Dismissed each list the 100 most recent findings.
Frequently asked questions
What is knowledge conflict detection?
It is a check that reads a chatbot's knowledge for facts that disagree, so the bot doesn't give one visitor the old price and the next one the new price. In intoCHAT, the Conflicts card on the Knowledge tab compares passages from different sources of the same agent and lists contradictions, outdated passages and inconsistent duplicates, each with both sides and a suggestion.
Which plan do I need?
Conflict detection is part of Pro and Enterprise. On Free and Starter the card shows Included from Pro with an Upgrade plan button, and nothing is checked.
Does a check use my messages?
No. Checks don't count toward your plan's monthly messages. They have limits of their own: up to 60 pairs of passages per check, up to 240 per agent a day, and Check now up to 4 times an hour per agent.
Which source does my agent trust until I fix a conflict?
When passages disagree, the model is told to follow a fixed order: Q&A pairs first, then sources you marked as priority, text snippets, files, Notion pages, Google Drive files, Intercom articles, website pages and imported helpdesk tickets. Between sources of the same kind, it trusts the most recently updated one. Conflict detection shows you where your agent relies on that order, so you can fix the source instead.
Can it flag facts that don't really disagree?
Yes. The check is a helper, not a guarantee. It can flag two prices for different plans when a passage doesn't name the plan, for example. Dismiss these: they won't come back unless the text of one of the passages changes.
Can it miss conflicts?
Yes. It can miss a conflict when the two passages are worded so differently that they are not close in meaning, when both facts are in the same source, or when a passage has more than 4 close passages from other sources. To see which questions your knowledge does not answer well, use Questions to improve on the same tab.
What happens after I fix a source?
Once the changed passage is trained, the finding is resolved by itself and shows "Resolved: a passage changed". The new text is compared again, and if it still disagrees, a new finding appears. When one of the two sources is deleted, the finding shows "Resolved: a source was deleted".
Who reads my passages during a check?
The passages are read by the AI provider intoCHAT uses for answers, OpenAI, whichever model your agent answers visitors with. Only passages from your agent's own knowledge are compared; conversations are not part of the check.
Stop your sources from contradicting each other
Create your agent free and train it on your content, then upgrade to Pro and open the Conflicts card on the Knowledge tab.
Create your agent freeFree plan, no credit card needed.