I’m making a project with GPT that makes recommendations based on the context. I’m trying to tell the bot to ask some relevant questions, which works perfectly. My problem is that after the questions, I want the bot to create a summary of the user’s answers, instead of trying to recommend something. I want to use the summary to create an embedding and then ask the bot again to recommend using the context. The problem is that the bot most of the times ignores that part of the prompt.
I’m using gpt-3.5-turbo.
What I have in the prompt:
“Never make a recommendation”
“Before making a recommendation you must answer only with [FINISH] - summary of user’s answers”
Consider creating a “summary of the answers” in the app and not using the chat API?
You can simply store this data in your app and create your own summary, or create a string with all the answers and prompt another OpenAI model to create a summary for you and use that summary.
Thank you! I asked to just answer with [FINISH] after asking the questions, and I made the summary after and sent it to embeddings, and now it’s working perfectly.
I just have another unrelated question, do you know how to add negative prompts? The bot is always asking for a location, even if add “Do NOT ask for a location”.
Hi @ze_t , Thanks for your question! Could I know how you evaluate if the Assistance has collected enough information for the following recommendation ?
I found the “Shop” app have implemented an assistant based on chatgpt api and curious
how it make it. Introducing ChatGPT and Whisper APIs
