Original Discord Post by darkladen | 2024-09-04 20:04:40
Hello, Im preparing a new avatar with new bank knowledge document but after modification from web document editor, saved the document but got stuck learning about this document and has never finished learning, so I can’t use it.
Thanks.
Char Id: 92ae0640-3f9d-11ef-860b-42010a7be011
Reply by darkladen | 2024-09-04 21:38:07
Today I uploaded a new file for the knowledge bank with information from a store and it works horrible. It doesn’t match almost anything in the document. We are already in conversation with you to take a premium plan but in the meantime I can’t upload any document for nothing works. The system gets stuck in “learning” mode and when it finishes (after a long time of waiting) it answers anything that does not fit what I need.
What can we do ? I need to move forward with the development but I can’t do it this way because it doesn’t work.
Reply by k3kalinix | 2024-09-04 22:24:47
Hello <@305834434178056192>,
Reply by k3kalinix | 2024-09-04 22:24:59
Could you please share Character ID?
Reply by k3kalinix | 2024-09-04 22:25:19
And did you talk with my teammate?
Reply by darkladen | 2024-09-04 22:54:05
Hello <@1023671043287699568> ,
Char Id: 92ae0640-3f9d-11ef-860b-42010a7be011
I haven’t talked to anyone else, except for my other post, but this time it’s more complex because it’s almost not working for me, let’s say it’s below 30%.
I would be happy if this document format would work as it is the only thing that is giving me problems.
Thanks
Reply by k3kalinix | 2024-09-04 23:13:19
Okay, thank you for details.
I have informed my teammate.
Reply by sherlock6147 | 2024-09-05 07:41:02
Hi <@305834434178056192>
To help improve the accuracy of the responses, could you please try copying the content of the file from the knowledge bank editor into a new text file? Once you’ve done that, use this new text file for the knowledge bank and upload it.
This should help ensure better formatting and accuracy.
Thanks
Reply by darkladen | 2024-09-05 14:00:18
Hello <@759094701038370877>,
I have uploaded a new file with some modifications and deleted the old one. It also has a different name if with this I can force it not to use the old file in case of some kind of cache memory.
Despite the above it does not respond at all as it should and I also get an error:
Something went wrong!
grpc: received message larger than max (4396793 vs. 4194304)
Thanks
Reply by darkladen | 2024-09-05 15:00:06
I have uploaded a new file (KB_Sodimac_Catalog_v3.txt) to the knowledge bank, set the GPT4-o usage and temeperature to 0.3 since the lower it goes, the more it is supposed to cling to the document.
For now I think it works better but I will continue to do more tests to see how it works.
Question 1: if we load a document, the model will improve its memory if we add pieces of documents that complement some that are already loaded or is it better to upload again the document with all the improvements/modifications?
Question 2: If I change the LLM I want to use (Mistral, GPT, etc) is the change instantaneous or do I have to wait for the change to take effect after a short time?
Thanks.
Replying to darkladen’s Message
Reply by darkladen | 2024-09-05 14:00:18
Hello <@759094701038370877>,
I have uploaded a new file with some modifications and deleted the old one. It also has a different name if with this I can force it not to use the old file in case of some kind of cache memory.
Despite the above it does not respond at all as it should and I also get an error:
Something went wrong!
grpc: received message larger than max (4396793 vs. 4194304)
Thanks
Reply by sherlock6147 | 2024-09-06 07:30:42
to fix the grpc: received message larger than max (4396793 vs. 4194304) errors I would recommend adding ‘respond in small sentences’ or something similar to the character’s backstory.
This error happens when the response received is too large
Replying to darkladen’s Message
Reply by darkladen | 2024-09-05 15:00:06
I have uploaded a new file (KB_Sodimac_Catalog_v3.txt) to the knowledge bank, set the GPT4-o usage and temeperature to 0.3 since the lower it goes, the more it is supposed to cling to the document.
For now I think it works better but I will continue to do more tests to see how it works.
Question 1: if we load a document, the model will improve its memory if we add pieces of documents that complement some that are already loaded or is it better to upload again the document with all the improvements/modifications?
Question 2: If I change the LLM I want to use (Mistral, GPT, etc) is the change instantaneous or do I have to wait for the change to take effect after a short time?
Thanks.
Reply by sherlock6147 | 2024-09-06 07:36:15
question 1: it will be best to test out both implementations to compare depending on priority of improving the accuracy
question 2: the change in LLM is instantaneous
Reply by darkladen | 2024-09-06 13:52:43
Hello, <@759094701038370877> We appreciate the information and willingness.
I will be doing a lot of testing and will advise of anything that occurs.
Thank you very much.
Reply by darkladen | 2024-09-06 19:02:08
Hi, I have noticed that when I make changes editing a knowledge bank document from the file editor via the admin web, the model becomes very erroneous responding anything and the accuracy is almost zero, but if I disconnect the file and upload the text document with all the modifications, it works and responds again.
It is quite strange but it may help you to perhaps resolve important issues.
Regards.
Reply by sherlock6147 | 2024-09-09 09:59:59
Thanks for informing us about the issue, Yeah we have noticed the same issue and are working on a patch for the file editor.
Will let you know once its fixed
Reply by darkladen | 2024-09-12 14:41:34
Hello,
I have uploaded a document to the knowledge bank, with a format that I have already discussed elsewhere and I leave a sample below. The strange thing is that when I start to ask questions to the avatar, I have noticed that if for each category I have 20 products, in most cases it does not recognize the products beyond the number 6 to 8. In some cases it does recognize them but only in some categories. I have tested the models gpt-4o and gpt-4o mini, I have changed the temperature of the tested model from 3 to 4, and well sometimes it improves a little but it is still not so accurate with such a short document thinking that at this moment I have 14 sample categories (in total 280 products) and the final document should be 150 thousand products and a little more. Here is a sample of the document.
[number of category]. [category name] category ([name sector] sector)
Ex:
Reply by darkladen | 2024-09-12 14:43:18
Thanks !!
Reply by darkladen | 2024-09-25 14:09:31
Hi, well we are still having this problem.
We are about to lose a big client because another player came in with something much more limited and only in 2D with the only functionality being that the model responds accordingly based on their document.
Anyway, today I see that the demo I am preparing is working better, although it is still not 100% in all the answers.
Is it possible to have an example to handle this kind of documents and make them work 100% ?
I am still testing using the other models you offer to see if they respond better.
Thank you.
Reply by k3kalinix | 2024-09-25 14:13:32
Did you talk to my teammate? I remember I shared a Calendly link with you.
Reply by darkladen | 2024-09-25 16:27:57
yes we have already talked to the teammate by calendly but we still only get some tips but not a 100% working solution but well, I guess the idea is that we go to the next level but if we don’t win the client yet, how can we get a really working demo project ? Clearly after the customer accepts it we can move 100% to the next level of subscription.
I think we should look at some kind of demo support to get functional projects created.
Are we still looking at this here or elsewhere?
Thanks.