Following on this thread and I am working on the similar case - enforce GPT to follow the guideline to conduct diagnosis. It seems like two major routes:
Agent/Prompt: explore way to coordinate different prompts/ agent and state to follow the guidelines.
Fine-tune model: fine-turn model on specific conversation data which I am not quite sure how to train such data which is not a Q&A pair rather a conversation.
It might be possible to do this much more cheaply using multi-agent RAGs … not sure a custom tuned LLM is needed (which would be much more expensive to build).
+1
There is some very interesting discussion and further information in an exchange on Linkedin including a newsletter article from @bill.french about CustomGPT from @alden and a nice additional blueprint from Bill that people might be interested in.
See Bill’s draft: The Hidden Side of Conversational AI · GTC
Basically a set of recipes for making a business.
Is there a reason that you have to use system? I love the idea of doing interview by AI . It might be a good alternative to survey monkey, anyone agree?
Absolutely – surveymonkey is nice when the questionnaire is very structured and rigid.
But when the interview is open ended and diagnostic for example, one can’t have rigid forms. That is where generative AI can really help understand “language” and get behind the meaning of things.
Thanks. Will ping you when the team is ready for early adopter program.
Have you tried this by chaining agent + voice and if yes, are you seeing issues with latency?
Nope – didn’t do the chaining yet. But I did just get the basic interviewer agent working. Hoping to release the github code and streamlit app soon (along with a medium article)
@alden did you have any luck with the interviewer agent? Would love to see what you built!
A somewhat good solution will be combining a traditional bot with an LLM. For the Question flow + more natural conversation.
For anyone still struggling with this, you can check out this post.