Measure and improve

Reviewing Conversations to find knowledge gaps

Reviewing conversations to find knowledge gaps

Every conversation your AI handles contains signals about what it knows well and where it falls short. A thumbs-down rating, a vague answer, or a question the AI cannot answer at all are all clues pointing to gaps in your training data. The Conversations feature in your Unless dashboard is where you turn those clues into concrete improvements, whether that means updating a Source, adding a new article, or creating a task for your team to fill the missing information.

Starting with the conversations list

The Conversations tab shows all conversations from the last 30 days, with the date and time of the last message, the number of responses, the user rating, and the conversation status. This list is your starting point for spotting patterns. Filter by rating to isolate negatively rated conversations, or filter for at least 2 or more responses to find longer conversations that tend to provide more insight into what users are asking.

Filters are persistent throughout your session. If you filter the list, open a conversation’s details, and return, your filtered view stays active. The Next conversation button also moves through your filtered list rather than the unfiltered one, which speeds up review when you are working through a specific subset. Filters reset when you navigate away from the Conversations section or refresh the page.

Reading the conversation details

Click Details on any conversation to open the full transcript. Each agent answer shows the sources used, any thumbs up or down ratings attached, and, where applicable, procedure runs and signal triggers. This is where you see exactly what the AI said and what it based its answer on.

Checking the sources

Click each listed source to open and review its content. If the AI answer does not match what you would expect from the source, you have found a problem. Remove wrong or outdated information and add updated versions when necessary. In the dashboard, you can also see all related sources, which include content the AI considered but did not include in its final answer. This gives you a fuller picture of what the AI was working with.

Flagging responses

If you believe a response has a technical bug, flag it and check the box to report it to the Unless team. You can also flag responses and leave notes for colleagues describing what the issue is and what the answer should have been. When you flag an answer, an analysis is generated on the spot to explain why the AI may have produced an incorrect or unsatisfactory response. This analysis stays available in the flag view even after you close and reopen it, so you can diagnose issues before contacting support.

Turning gaps into action

Once you have identified a knowledge gap, you have several ways to act on it. You can add the question to your Quality control center directly from the conversation details page. The question does not need to be perfectly structured at first, since you can edit the question and the control answer after saving. You can also create a knowledge suggestion task from the conversation details view, which lets you capture gaps without switching contexts. Tasks help your team track missing information and prioritize which knowledge gaps to address first based on impact and frequency.

Using the content library for deeper diagnosis

If a wrong answer persists, the Content library can help you find the offending content unit. The Engine reads every Source, breaks it into ideas, dedupes across Sources, and rewrites everything into one consistent voice. Every unit in the library traces back to the Source it came from, so you can follow the chain to the original content. Open a unit to see its source chain, and you will usually find that one of the upstream Sources is wrong, outdated, or contradicts another Source. Fix the Source and the library updates automatically. The Engine also flags conflicting content across Sources and topics where knowledge is thin, which appear in the Inbox as messages waiting for your judgment.

Conclusion

Reviewing conversations is not a one-time cleanup, it is a continuous loop. Check the negatively rated conversations regularly to spot recurring problems, prioritize improvements to your AI training, and ensure users receive accurate answers. Each conversation you review either confirms your AI is working or reveals exactly what needs to change. Feed those insights back into your Sources, your Quality control center, and your task list, and your AI gets better with every cycle.