Chat completion or completion endpoint for multi turns?

API
1

Hi Everyone!

This post aims to discuss how to overcome the lost memory for multi turns of QA with the ability to get external context.

There are two ways of doing this:

  1. Chat Completion API, where you send the API with the list of messages. Each user message will have {Context} and {Question}.
  2. Completion API, where you store the messages as concatenations of message in string, like you can format like History : {Role} : {Content}.

For my scenario, I went to 1 first, as it seems to be newer. But I would suffer from lost memory base on the wording of my the latest question. For example, if I say summarize the conversation, gpt remembers the conversation. But If I say, what was my last message, it didn’t know.

I do not intend to get the best formula here, just curious to know if anyone has come across before, what is your opinion?

2

for more details: (Don’t know why I could not add links here…)

semantic-kernel
issues
4707