Examples of Custom Instructions for an RFI/RFP Response ChatBot

As presales for an eCommerce software solution, our job is to respond to countless RFIs/RFPs. Over time, we have generated a massive question-and-answer knowledge base. I want to use ChatGPT or GPT Builder to create an RFP Response Machine for future RFP questions.

I have 1000+ Questions and Answers in a CSV file that I can easily upload. I tried to write the custom instructions, but the results were not too good.

Could you share examples of custom instructions that you use to create an answering bot based only on the knowledge base provided?

Should I also upload the questions or only the answers to make the bot more accurate?

As a starting point, it is important to:

  • Respond only based on the knowledge base provided
  • Respond in the same format/style as provided
  • Provide as many reference links as possible, such as reference documentation, videos, web links, etc.

Usually giving orders and instructing in the second person worked for me. Preceding your instruction with commands, for example: ‘you must’, ‘always’ etc.
Don’t be afraid of slightly longer instructions, if they are worded well.

  • You must respond in exactly the same format as provided in your knowledge base
  • You must use word for word answers, as in your knowledge base
  • Always provide all reference links related to the answer

You can play with it…

For the first prompt you can tell it something like: If you cannot find answer in your knowledge base, reply ‘Sorry, I cannot give you answer for that.’

That way you will provide it with exact action of what to do when it cannot find the answer, rather than letting it make something up.

Hopefully, you should be able to get quite close to what you’re looking for.

As the last resort, you can still try to train it on these Q&As: https://platform.openai.com/docs/guides/fine-tuning

EDIT: The fine-tuning actually involves API, not the GPT builder, sorry for confusion.

What i do is I tag on an additional bot without chat history to read the messages first, draw from its sources, and handoff to the bot with chat history with instructions.

dunno if thats weird but… works pretty good for me.