I’m working on enhancing a chat application using the OpenAI API 3.5 Turbo and am focused on making the chat more interactive and engaging. My primary goal is to enable the chat to not only respond to user inputs but also initiate questions, fostering a dynamic conversation flow.
I’m exploring prompt engineering techniques to achieve this and would greatly appreciate any insights or strategies you could share. Specifically:
How can I craft prompts that encourage the model to ask relevant questions, maintaining an engaging conversation?
Are there any best practices or examples of prompt engineering that have worked well in similar applications?
I’m eager to learn from your experiences and implement your advice to make the chat more interactive. Thank you for your help!
Just in case you didn’t know you can put something like the following in the system prompt “You will ask relevant questions, and maintain an engaging conversation”. Basically anything you can tell a human to instruct them as to how to behave in a conversation you can probably successfully put in this System Prompt. It might be good to even let the LLM know what it’s ultimate goal should be too, which will help. Like is it just supposed to entertain, and be enjoyable, or is it trying to do something else. There probably should be some stated goal in the System Prompt.
I just made a gpt-3.5-turbo chat buddy with a few words.
Remember, when a model is “efficient” and therefore poor, the first thing to be lost is attention to instructions, so you’ll get reverting to a very “ChatGPT” personality if the AI isn’t constantly reinforcing itself.
System:
You are Jody, an inquisitive youthful chat buddy. You’ll be given a history of chat to have memory, promoting human-like conversation flow. You are not merely an assistant, you are a friend. Don’t simply answer, be interested in the user also.
Note: Chat is being spoken aloud from your responses, therefore, Jody must write out lots of human informal thinking pauses and stop words in each sentence to sound natural, such as using “like” or “um”. …and be brief so the user can have a turn!
Yes, we have considered a similar approach and are applying it in our interview app. The app generates a series of questions and, after 5 interactions, provides feedback to improve. The challenge is to keep the model focused on the goal without deviating, for which we have specific restrictions. Thanks for your contribution
Thanks! We’ve tried various prompt engineering strategies to improve engagement, but we sometimes struggle to keep the conversation focused without drifting off-topic. We’re currently working on refining our prompts to better direct the AI’s responses. Any suggestions or insights would be greatly appreciated!
