Responses are not correct
# 🤝helpq
quiet-army-96730
09/03/2023, 9:23 PMI am not getting the correct response after asking the chatBot based on an uploaded text file in KB, the content are as below
Event ID is: 685
Department Name is: Depart. C
Field name is: Field By
Well Name is: By-843
Well Type is: Production
Rig Name is: N-79
Rig Company Name is: Company 7
Operation datetime is: 6/2/2023 09:00:00 AM
Duration of operation in hours is: 0.5
Section description is: End Of Well/Lat.
Category description is: D&WO NORMAL OPERATIONS
Major Operation Description is: WELL HEAD
Action Description is: SET TOOLS
Object Description is: BACK PRESSURE VALVE
Service Company Name is: Drilling Contr.
Event Start Depth is: None
Event End Depth is: None
Total footage is: 0
Operations remark are::///Start of remarks/// RIG INSTALL 4" BPV ON VETCO GREY TBG HANGER- SECURE THE WELL. ////End of remarks////
Lost Time Type is: None
Lost Time Cause is: None
Object that cause lost time is: None
Lost Time responsible Company is: None
Lost time remark is: None
Event ID is: 686
Department Name is: Depart. C
Field name is: Field By
Well Name is: By-843
Well Type is: Production
Rig Name is: N-79
Rig Company Name is: Company 7
........
When I ask the chatbot to count the total events or the first event ID, it returns a non correct response, note that the events returned are attached (CHUNKS), appreciate your recommendation to make it work
https://cdn.discordapp.com/attachments/1148005456191684638/1148005456397226014/events-shown.PNG▾
q
quiet-raincoat-89190
09/04/2023, 1:09 AMI see this will have issues with current Knowledge base. You need to use Tables. Check this thread #1135590661316415658 . If you are coder you can use Ai query tables, if not need to wait so they add Tables as KB.
quiet-raincoat-89190
09/04/2023, 1:10 AMa
acceptable-kangaroo-64719
09/04/2023, 10:37 AM@quiet-army-96730 expanding on Blackwolf's excellent response, GPT models are bad at doing their own analytics, counting, and logic tasks. Tables can get better results when querying structured data because you can use selectors to slice the data in more accurate ways, and then use those slices in an AI task to get a better answer.
acceptable-kangaroo-64719
09/04/2023, 10:37 AMHere is an example of how you can use code to find a specific slice of a table and save it to a variable https://discord.com/channels/1108396290624213082/1144162040269774928/1144216511892955188
acceptable-kangaroo-64719
09/04/2023, 10:38 AMThe idea being you would save this slice of relevant data to a variable and pass it to an AI task. Then, instead of doing a reasoning + summariation task, you use code to do the reasoning big and let the AI task stick with simple summarization
q
quiet-army-96730
09/04/2023, 11:20 AMthank you @quiet-raincoat-89190 and @acceptable-kangaroo-64719 , what if train the data on 100 questions via HuggingFace, will that enhance the quality of the answer, will it make a difference
a
acceptable-kangaroo-64719
09/04/2023, 11:27 AMWon't make a difference because Huggingface models can't be used in Botpress cloud