Knowledge Base Report

This article covers Insights > Knowledge Base. The page lists every time a flow or an AI agent searched a knowledge base: what was asked, what answer was given, and which documents the retrieved chunks came from. To build and fill a knowledge base, see Knowledge Base. The date range, remembered filters, paging and data retention work the same way on every report page and are described in Reports.

📝 Note: The page is available only when AI agents are available for your tenant, and your role needs the Knowledge Base Reporting permission.

The examples use ApplianceBrands, a knowledge base holding support documents for two appliance brands, Vantera and Solmar. A customer asked "My refrigerator is not cooling. How do I reset it?"

Finding an answer

Open the Insights menu and select Knowledge Base.

1. Date: Filters on when the answer was produced. Default: Last 30 days.

2. Conversation/Session ID: Takes an exact ID and lists that conversation's knowledge base calls. It ignores the other filters.

3. Knowledge Base: One or more knowledge bases.

4. Channel: One or more channels.

5. Status: Found or Not Found.

6. Results: One row per knowledge base call, newest first.

The page has no export and no column options.

To find the example answer:

1. In Knowledge Base, select ApplianceBrands and click Filter.
2. Find the row whose User Query reads "My refrigerator is not cooling. How do I reset it?".
3. Click the row. Its detail panel opens beside the list.

💡 Tip: When you start from a conversation, copy its Session ID in Conversation Details and paste it into Conversation/Session ID. The list then shows every knowledge base call made in that conversation.

Reading an answer

1. Header: The knowledge base name, the response time in seconds and the Status.

2. User Query: The query in full.

3. Standalone Question: The query rewritten by the RAG Agent node so that it makes sense on its own, without the rest of the conversation. Shown only when there is one.

4. Response: The generated answer. A Cached Answer badge means the answer was served from cache.

5. Retrieved Chunks: The average score, then one card for each chunk, showing its rank, source document, score and full text. A chunk's score is green when it is at or above this answer's average, and amber when it is below.

In the example, the Standalone Question is "How do I reset my refrigerator?" and the average score is 0.72. Chunk 1 comes from vantera-refrigerator-support.txt with a score of 0.58, shown in amber because it is below the average. Reading a chunk's full text next to the Response shows which passage the answer was written from.

💡 Tip: If a chunk comes from the wrong document, such as a Solmar file behind an answer to a Vantera customer, tags can narrow what a flow retrieves. See Knowledge Base Tags.

Where rows come from

Each row is one knowledge base call, not one customer turn, so a single customer message can produce several rows. Rows come from three places:

Only RAG Agent rows can have either status. The status depends on whether the average chunk score met the RAG Agent node's search threshold. Calls that failed outright are not recorded.

⚠️ Warning: Rows from the Knowledge Retrieval node and from an AI agent's Knowledge Retrieval tool always show Found, so for those rows Found does not mean that relevant chunks came back. Filtering on Not Found shows RAG Agent rows only.

📝 Note: You cannot open the detail panel of a Not Found row. To see that conversation, use the row's trailing icon.