Serious problems with the Knowledge Base system

Core Features Knowledge Bank
1

Hello Convai team,

I have been working with several avatars and testing different document structures in the knowledge base, but the system is not responding correctly to the uploaded content.

This problem also occurred a long time ago, but then it worked very well, using Markdown structures, it was always accurate in its responses.

Chat ID: 7685b42a-84f4-11f0-bd70-42010a7be01f

:magnifying_glass_tilted_left: Main problem

When the document includes structured information—for example, a program with daily schedules—the model mixes information from different parts of the text.

If I ask it something specific like “What is Monday’s program?”, the AI responds with a confusing combination of data from Monday, Tuesday, or Wednesday, and even makes up parts.

:puzzle_piece: Details observed

I have tried multiple formats: plain text, Markdown, lists, block separations, clear headings, etc.

Regardless of the format, the model does not distinguish sections well or correctly associate contextual data.

The result is inconsistent, mixed, or incorrect answers, even to very simple questions.

This makes the system unusable for structured or sequential information such as schedules, programs, or event listings.

:page_facing_up: Example

Question: “What activities are there on Monday?”

Expected response: only those on Monday.

Actual response: mix of Monday + Tuesday + Wednesday.

:warning: Request

Could you review how the internal document indexing or segmentation system is working?

It would be very helpful.

2

I forgot to mention that all the information uploaded as “brief background” in Core Description works fine, but not the information uploaded as separate documents in the Knowledge Base. For the latter, any related question found in these documents is retrieved poorly and ultimately responds with anything.

3

It’s indisputable!!!

I have adjusted the Backstory information (made it a little more concise) to add other types of information, such as a program/schedule, which has never worked well for me when I ask questions and it responds terribly, but when I add it to Backstory, it works perfectly.

The problem is that I can’t add any more information because I’ve reached the limit, and it doesn’t work well at all in the knowledge base documents.

4

Now it’s working even worse. I uploaded two very small files to the knowledge base (one with 219 words and the other with 436 words) and it’s unable to retrieve information from these documents. It doesn’t respond to any of them.

This is quite serious. Whenever we have to show the product to a customer, we encounter the same problem. The only major issue is that it doesn’t work well with the knowledge base.

Will we have a solution by this Sunday?
I hope we can get some help/a solution.

5

Thanks for flagging this. I understand how disruptive this is during client demos. Our engineering team can review this during the week. In the meantime, please try the following:

  1. Create a new character.

  2. Write a clear, focused character description that does not conflict with the knowledge base.

  3. Re-upload your knowledge in strict Question–Answer pairs

Please also share:

  • Character ID and recent Session IDs

  • A sample question that fails, answer and the expected answer

6

OK, I will create and upload the files again following the instructions provided. However, although I have a few question/answer pairs, I couldn’t create many because most of them are answers from the general data uploaded. I would have to create hundreds of pairs to cover most of the information.

I will be back soon with more data.

Thank you for the information.

PS: I forgot to mention that I repeatedly get an error message on the dashboard when I try to upload/assign files.
image

7

Here is some information about the new character and the session that repeats the mistakes of the other character and its knowledge base.

Char ID: 41c57c66-ac54-11f0-ab6d-42010a7be025
Sessions ID: d4511a38980174cef92589088fde646f - b19b7c4a210a0d132d64c1c285d855a2 - debdbc83fae5ab42c72d7a2608455d7d

The only difference is that in one document, he managed to respond to information that is in the first few lines.

8

I tried changing the model in Core AI and using Claude-4.1-opus, but since it is not a good conversational model, it responds with literal information, adding Markdown characters that are in the document, and listening to them is very uncomfortable, not to mention that the responses are very slow (10+ seconds). Although the response is complete, it is not the same as ChatGPT.

In general, almost all models do not work well with the files in the knowledge base.

9

What language are your documents in?

10

In Spanish. Although I am translating everything into English now to do more testing.

I should also mention that everything is working much better in the new character, using the Gemini-2.5-flash model.

With regard to the old character, everything is still very bad, but since the new one responds better, I will stick with the new one.

And I thank you in advance for your help.

11

This should be in English. Could you try it with the Add knowledge section?

12

Sure, I’ll run some quick tests, but I’d rather create a new character because tomorrow we’re installing this Avatar project for a client, and to avoid it failing as much or more than the other avatar, I’d rather run new tests on another character.

Thank you.

13

Today I am testing the newly created character (the one that was working a couple of hours ago). Now it works intermittently (in its responses), as if it doesn’t know the answer sometimes and other times it does. I hope it doesn’t have any mental lapses today when I’m at the event.

The way it works is too strange, and the end result will not be good for us.

New – It is also happening that in some cases the audio of the spoken response is not received.

I hope it gets sorted out in time.