I am facing problem with chat completion API. It is not analysing the user reply properly. As a tutor It asked a question related addition and the user have replied correctly but it gave the user reply is wrong and gave the correct answer but the user reply and the assistant reply is same. I have…

It is not analysing the user reply properly and randomly choosing the answer but the explanation is correct

{ "model" => "gpt-4o-mini", "messages" =>[ { "role": "system" "content": You are a tutor. Ask question after each reply. Give a chance for student to reply Acting as a teacher to a private student and in conversation mode, provide the information below to the student. Give the student an…

Where’s the rest of the prompt? Where is the user and/or assistant prompt? Are you using vision? Your temperature is probably way too high btw. Mine succeeded with temp 1.0 :stuck_out_tongue_winking_eye:

If you want the highest resilience, you would use a specialized AI call that doesn’t involve the user chat, which verifies that the generated question is logical and topical, can be understood, and has a correct answer and remaining answers are incorrect. It then creates an answer key, directly inf…

User chat is fine. See my example. This is easy for 4o. Something else is wrong.

One “looks good to me” is not sufficient. Use model gpt-4o-mini as described to perform a tutoring session. Give such a confusing developer message. Follow the format described in the first post, generating LaTeX questions in addition to lessons and Socratic method. Get 30 chat exchanges deep… T…

Chat is fine. [image] _j: Get 30 chat exchanges deep… That’s not good design though? If all you want to do is use AI to answer individual questions one by one. Reduce the look-behind to a small window so prompt is focussed, preferably only on the current question. And that would save you …

Tried this dozens of times Fully formatted example, no function calling: [A forum post asks what rac{3}{4} + rac{1}{4} equals, with the options given, and a user responds with "B) 1". (Captioned by AI)]

Now - look closer at the original post. What ‘merefield’ needs to be saying in your screenshot is not “question 1”, but rather, “B”. A suggestion “do nothing” “it works fine” is dismissive of the problem, whereas I present changes that should improve such an application and reported symptom.

Ah got it. Misread. Let me explore further.

Chat completion API giving wrong answer

API

merefield February 14, 2025, 3:05pm 13

Tried this dozens of times

Fully formatted example, no function calling:

Topic		Replies	Views
GPT thinks math answers are wrong, even when it says they are right API gpt-4 , api , response_format , gpt-4o-mini	2	212	October 27, 2024
Gpt-5-chat-latest, can't solve: 5.9 = x + 5.11 API	16	3472	August 13, 2025
Gpt-4o-mini (even gpt-3.5-turbo) works but gpt-4o doesn't Bugs gpt-4o	7	437	November 24, 2024
Fine tuning gpt-3.5 turbo with mutiple user, system and assistant prompts in a message API gpt-35-turbo , api	2	1133	December 20, 2023
Response for mathematical request Prompting	9	3547	October 29, 2024

Chat completion API giving wrong answer

Related topics