Original Discord Post by darkladen | 2024-10-30 15:15:54
Hi, how can we strengthen the documentation we have uploaded and have it answer the questions well ?
Generally I have seen that when we upload the data (300 to 400 blocks of data, a block would be a row and 6 columns), the first records of the documentation is what the model can learn faster but going down in the document, you don’t recognize the information and also, in many situations, it takes a lot of days to learn well all the documentation.
Do you have any usage tips or LLM model recommendations that work best for learning this amount of data and answering everything?
Would it be possible to use another more structured data type such as JSON or similar?
Store Name: 5ASEC
Sector: Parking
Level: -1
Directions: Aisle behind me, reach the end of the aisle, turn right, and go down the escalator to level -1.
Location: Service Boulevard
Categories: home
Store Name: AUTOPISTA CENTRAL
Sector: Parking
Level: -1
Directions: Aisle behind me, reach the end of the aisle, turn right, and walk to Supermercado Lider, then go down the mechanical ramp to level -1.
Location: Service Boulevard
Categories: banks and financial services - highway
Why does it happen a lot that he does not know what he is asked, and after a lot of insistence in many ways he gives the correct information, but if I restart the chat, he forgets everything ? He does not remember the above and is not able to go deeper into the document until I insist him a lot to look for it ?
Moreover, when he starts to answer well, he answers practically everything in the same line, but when I restart the chat, he forgets it. This is very serious and as I said some time ago, it doesn’t even allow me to make a demo to the client in a convincing way.
Hi <@1023671043287699568> , sorry for the delay but we are on holidays.
I am with another account and the character ids are:
433f2f9e-915f-11ef-a874-42010a7be011
0e1a4762-96d3-11ef-8e96-42010a7be016
Both characters use the same documents from the knowledge bank.
By Wednesday we have a final demo with customer. In the previous versions of the character, I used shorter documents and after many days the character was able to answer 100% of the questions correctly, but now there is a lot more data but I don’t think it is too much for me to make so many mistakes.
The strangest thing is that after a lot of reinforcement and he finally starts to respond well almost 100%, after restarting the session he forgets all the reinforcement and responds badly again.
Maybe we need to spend more time reinforcing him with questions so that he finally learns to respond well but I’m not sure. This is why I have come back to ask you for some help, tips or best practices for this task.
Another important thing that happens to me is that when using gpt-4o-mini it generates more errors and takes longer to learn after reinforcement. Now I am using Claude-3-5-Sonnet and it is the one with which the reinforcement works better.
It is still not clear to me if after reinforcement with Calude, if I switch to Gpt it keeps the reinforcement or I have to start all over again.
uploading and updating files in the knowledge bank is not working. I have some files on hold for about 15min and they are still not finished processing and connecting.
uploading and updating files in the knowledge bank is not working. I have some files on hold for about 15min and they are still not finished processing and connecting.
Can you help me ?
Reply by sconvai | 2024-11-05 19:14:28
Can you please tell email-id of the account using which you are uploading the file. Also please confirm if the upload is complete or not?
Reply by darkladen | 2024-10-30 15:15:54
Hi, how can we strengthen the documentation we have uploaded and have it answer the questions well ?
Generally I have seen that when we upload the data (300 to 400 blocks of data, a block would be a row and 6 columns), the first records of the documentation is what the model can learn faster but going down in the document, you don’t recognize the information and also, in many situations, it takes a lot of days to learn well all the documentation.
Do you have any usage tips or LLM model recommendations that work best for learning this amount of data and answering everything?
Would it be possible to use another more structured data type such as JSON or similar?
Thanks.
Reply by sconvai | 2024-11-05 22:07:03
Is it possible for you to give an example document with a one or few sample questions to test this out. I am assuming this has more to do with document organization and can be improved.
Reply by sconvai | 2024-11-05 19:14:28
Can you please tell email-id of the account using which you are uploading the file. Also please confirm if the upload is complete or not?
Reply by darkladen | 2024-11-06 00:58:27
Hello, I had created another post dedicated to this problem and yes, they have already been uploaded too late.
Reply by sconvai | 2024-11-05 22:07:03
Is it possible for you to give an example document with a one or few sample questions to test this out. I am assuming this has more to do with document organization and can be improved.
Reply by darkladen | 2024-11-06 00:59:44
Hello, Here is a test document.
Thank you very much.
It was failing with almost any question asked in any form. It just wouldn’t give the correct answer or it would say it didn’t have the information, which is no longer the case.
Since the last big update I did, it is already working much better and almost at 100%. Sometimes it delivers the response well but it gets a little creative adding information that is not necessary but in general it is working well and the best thing is that the small changes I make, it is taking them very fast.
Thanks for your time and I hope that this work fine in the time.
One suggestion is to remove “---------------------” but keep two stores separated by two “\n\n”.
Also can you confirm if the character has improved. I have made some changes.