We are building an account setup agent which takes in user’s name, date of birth and phone number for registration, passes the information back to the functions for processing.
What we notice is that most of the time the date of birth and phone number passed by the realtime api to the function is incorrect.
- We see correct transcript text for both the information, but when its sent to the function it changes.
- Also it makes up few information and tries to complete the information, example if I give a 9 digit number, it adds some random number and makes it a ten digit number. If we only give the month and year, it makes up the date and sends.
3 Likes
Seeing the exact same issue with order numbers. The AI will even confirm “the order number is 1234” but then enter something else into the function call, e.g. 123456.
Currently this is a total blocker for using the Realtime API.
1 Like
Same problem posted here: /t/why-is-realtime-model-so-bad-at-understanding-sequences-of-numbers/995062
“Glad” to see it was not happening just to me. There is a clear issue with number sequences recognition
1 Like
So here’s a question: Is a large language model a suitable tool to handle critical data such as order numbers, account IDs or financial figures which just have to be correct?
Would you be better off using standard software code to get and handle this type of critical information, use the LLM to write the text you want to generate around the numbers and then combine the two at output time using normal software code?
The fundamental way LLMs work means they are never going to be 100% reliable when it comes to numbers or other items of critical data.
Well, we are in production with the regular Chat Completions API and I have basically never seen this issue with that product. This problem is specific to the Realtime API and feels more like a legitimate bug than an LLM reliability issue.
IT IS a bug 100%
I’ve debugging it for the past 3 days and the issue comes from OpenAI, to be honest it could be just the speech to text model they are using instead of the LLM
@domingosl
Did you solve this problem? If yes, could you please share the details?
@stevenou
Hi, are you able to solve this problem? Could you please share the details?
same issue here. The transcript is correct but the AI answer is not aligned with transcript info …
same issue too. Correct transcription but wrong input argument.