19
Finally got GPT to stop hallucinating my data with one weird trick
I was pulling customer records from a CSV into a chat session for a market analysis, and it kept inventing sales figures. Then I tried adding a strict "only use provided data" instruction at the end of every prompt, right before hitting send, and it worked perfectly. Has anyone else found a simple prompt tweak that fixed a major hallucination issue?
3 comments
Log in to join the discussion
Log In3 Comments
grant_allen851h ago
Yeah that putting the instruction at the end thing is legit. I had a spreadsheet of shipping costs and it kept adding fake handling fees until I started slapping "dont add any numbers that arent in the csv" right before hitting enter. @grace_allen you're totally right about the exact numbers thing too, I had to do the same with inventory counts or it would round everything to the nearest hundred. It's like the model gets lazy if you dont spell out "no creative filling in the blanks" right at the end of your message. Straight up just paste the same warning every time and it works.
5
grace_allen18d ago
My CSV had dates formatted wrong and it kept making up quarterly totals until I added "use the exact numbers from the CSV, do not round or estimate" at the start of every query. That single line cut my error rate down to almost nothing. Now I just paste that boilerplate instruction before each prompt and it stays honest.
3