Strucutre of sub questions in embedding

I have an exam. The exam has a long scenario. It then follows several questions. In the docs, it mentioned to use curie and have multiple questions under that one scenario. The user can might ask only one of those similar questions. The questions are also based on some domain knowledge.

How do I structure the questions and answers?
Is it like

content: "lorem ipsum... \n\nQuestion: [q1, q2, q3] \n\n Answer: [a1, a2, a3]"

something else?

And when we embed it, do I combined content+question+answer?

Honestly, you do not need embeddings for this type of task. These types of questions and answers are fixed strings and you can use simple string matching to generate the questions and to select the answer.

In other words, you do not need AI to accomplish this type of task.


I tried prompt and completion, it wasn’t great. The Q&A is not something like the Olympic wiki example, nor is it like product service bot.

The scenario goes something like “Billy wants to buy a MacBook for $1000. It is currently on sale for 20%. Sally, his sister, wants to share the MacBook with Billy. She will put in 40% of the cost. However, Sally does not have a lump sum cash and can only pay monthly at 5% to Billy per month.”

Question: how much will Sally have paid in month 3?

Something like this, think cpa exam but this is not cpa exam.

Edit: or question could be like create an amortizable table of payments. While I understand this question can be answered already in chatgpt, the data I have is very domain knowledge specific with very long scenarios.