It seems that I have recently encountered this issue, which I definitely didn’t have at the beginning. I don’t know exactly when it started, as I haven’t analyzed it closely, but I’ve noticed that the avatar doesn’t seem to know much from the information I uploaded to the knowledge bank.
I placed a series of clear data there, and later, when I ask something, it says it cannot process the uploaded files. What could be the problem?
Her answer:
“Oh, really? I’m sorry, but I can’t access or process uploaded files. However, if you have any questions or information you can share in text, I’d be more than happy to help!”
Pretty sure some time ago worked well. Also, randomly the answer is far from true, like I ask a price from that list which lets say is 255, and answer is 452, I don’t even know from where…
BTW: I replicated the same thing on openAI chatgpt 4o and had no issues with receive a correct answer …
a. Currently we do not support KB in language other than English. It could work however we do not officially support it.
b. The organization of file is not favorable for KB. Currently we treat anything separated by two “\n” (or Enter) as separate paragraph. This is resulting in two separate chunks being constructed, one with only header (Pret energie 01 ianuarie 2025) and the rest of the table in second chunk. I suggest you change that to single “\n” and upload the file again. E.g.
Change This:
Pret energie 01 ianuarie 2025
ora 01, XXX ron, volum YYYY MWh
To this:
Pret energie 01 ianuarie 2025
ora 01, XXX ron, volum YYYY MWh
Apparently, it doesn’t work. I switched to English, broke it into smaller parts, but it still doesn’t seem to know the information in detail. I’ll try using general data to see if it learns them.
However, the core API is ChatGPT-4o, and when using 4o separately, it works perfectly.
The truth is, if it can’t remember and reproduce exact numbers, it’s quite problematic for my project.
After extensive testing, text reformatting for the Knowledge Bank, and dozens of questions between Convai and OpenAI, on Convai, I had around 20% positive responses, 40%-50% negative responses, and the rest were incorrect answers. On OpenAI, I had 100% accurate positive responses, regardless of how I phrased the question—whether in English, Romanian, mixed words, etc.
Another thing I noticed was that I didn’t see any difference when selecting different AI models or adjusting the temperature setting. The responses were not structured differently.
Hi, honestly I haven’t pushed too much on this issue lately. I understood that the data structure I want the AI to understand is hard to process in a plain text format. Unfortunately, it would need to be in PDF or CSV. Working with text is a real headache.
I’m waiting for the moment when I’ll be able to upload other types of files to Convai as well, just like I can already do with ChatGPT, Kimi, etc. At the moment, I only have these limitations with Convai.
@Florin_Danilov : Can you please confirm if you are still facing the issue. We have a new feature which should help RAG/AI better understand the document structure. If you are willing, we can enable the feature for you to try.
Hi there, thanks for the update. During this period, I’ve been focused on another module of my application, but the issue with how my data is formatted in the text file still persists.
What kind of test feature or tool could I use to better understand how to upload data in a way that’s correctly interpreted? Other format than txt?
