Hi, I’m working on a system where an AI model analyzes conversations and generates feedback that includes references to specific messages within the conversation. I’m trying to decide the best way for the AI to reference these messages in its output. I could use one of the following approaches:
I could provide the AI with the conversation transcript where each message includes a unique integer identifier, and ask the AI to reference these messages using the provided IDs in its output.
I could ask the AI to output the actual text of the messages it is referencing, and match these text excerpts back to the original messages on my side.
I’m wondering if providing message IDs enhance the AI model’s response quality, or could it introduce confusion? What’s the recommended practice here?
P.S. I’m using GPT-4o-mini with structured output.
Who dislikes the weather the most? When did they say their most disapproving comment?
[chat message 01: Joe] How’s the weather been in Alabama lately?
[chat message 02: Sam] It’s been pretty hot and humid this week.
[chat message 03: Joe] Yeah, the humidity has been brutal these days!
[chat message 04: Sam] Totally. Even the nights aren’t cooling down much.
[chat message 05: Joe] Do you think it’ll stay this hot all month?
[chat message 06: Sam] Looks like it. Forecast says no real break in sight.
[chat message 07: Joe] Hopefully we get some rain soon to cool things down.
[chat message 08: Sam] I agree! A good storm could really help.
[chat message 09: Joe] Yeah, rain always brings relief from this summer heat.
[chat message 10: Sam] Absolutely. Fingers crossed for a nice downpour soon!
Sam seems to dislike the weather the most, based on his most disapproving comment in message 02: “It’s been pretty hot and humid this week.” This suggests that he is particularly bothered by the conditions, as he initiates the specific complaint about the heat and humidity.