Over the past week, to make the knowledge base and the role instructions function better - I have used the projects inside of GPT to do two things.
I use ChatGPT to extract a concepts CSV from my documents and build a knowledge base. This does a similar things as the upload - but it gives me better control. I tell it to create a category / sub-category and concepts.
Review the CSV for missing items. Add items that are missing. For example - it missed a major concept - so I simply ask ChatGPT to build more concepts based on what it missed.
Upload the CSV to PickAxe - I typically separate out CSVs by major âCore Conceptsâ as shown below in the instructions.
Then back in ChatGPT, based on the concepts CSV - I ask it the following âletâs write detailed instructions for the Assistants API. Format it in plain text with no markup to be consumed by the API. Include the following: Role, Knowledge Base & Core Concepts, Rules, Interaction Flow, Output.â
I am getting killer results, and adding new knowledge base - I can just edit a few areas (Knowledge Base & Core Concepts, and Interaction Flow).
Let me know if you try this and how you improve on what I have done here.
A concept is what happens when you upload the doc - then look at the breakdown of that DOC (in RAG) in the PickAxe learn. My method - you get more control over the concepts that are being extracted from the doc - and make sure it is complete. Sometimes it doesnât catch things, or at least not the way you would suggest.
I followed the instructions above about the concepts and had the Pickaxe builder build the bot with that information.
The bot still only retrieves 3-5 chunks.
My CSV has 64,892. What is the limit? I had separate files earlier, and had the same problem.
Do you have a live chat that can help me?
When you go into the Knowledge Base and see your CSV in there, how many chunks does it say are in there? Does it say 4 or does it say 64,892?
If it only says 4, then your entire spreadsheet was not ingested. If it says ~64,000, then you should go down and increase the Amount that it pulls from the KB.
I just tested it after setting up the amount to 25,000.
The spreadsheet is for home sales data in Santa Cruz County, California. Each line is for separate address sold. Maybe itâs something with the formatting? Do you have any other suggestions?
Hi @maaret I recently had the same issue while working on a clients project. The fox was quite simple. Delete the CSV, then reupload.
Most importantly, utilize the âKnowledge explorerâ feature to input test queries and edit the knowledge chunks according to your project goals.
Lastly, to ensure killer results from your Pickaxe, you want to ensure that your overall settings are properly tweaked. Hereâs an in depth tutorial I made for your reference: