Additional Instructions per run are ignored

API
1

Hi. I’m trying to provide additional_instructions with each thread run to ensure the assistant has the most relevant context within the app (state of the NPC character). But it seems that additional instructions are getting ignored often. I tried to use Playground to see what was going on. I see Add Message and a Run requests

{
"role":"user"
"content":"Hi"
}

{
"assistant_id":"asst_7Shq9JxBeQGEXfTxmBH5M5Cx"
"additional_instructions":"Make sure to add 'TEST INSTRUCTION' at the end of your reply"
}

but the answer is

Hello! How can I assist you today?

Model is set to gpt-3.5-turbo

Instructions field is set to

You are a helpful assistant helping to develop and debug API integration. End your answer with information about instructions and additional instructions that you were given.

Why it is not working correctly?
It seems that switching to GPT-4-Turbo and back to 3.5 makes it work. But it is very inconsistent.

Maybe there is another way of providing relevant information to the assistant for each run?

2

Just seeing the same thing here.

Using:

assistant_response = openai.beta.threads.runs.create_and_poll(
        thread_id=thread_id,
        assistant_id=assistant_id,
        additional_instructions=additional_instructions,

And if I print the assistant_response … There is no sign of the additional_instructions…

Looking at one of the examples

They simply override the instruction…

3

And looking at the latest source of that function openai-python/src/openai/resources/beta/threads/runs/runs.py at 906ece0e45235cb170bc062902182eb8b55698b9 · openai/openai-python · GitHub

They do seem to pass it to the create function that sends that additional_instructions to the backend… So not sure what happens next…

4

This is still broken – For now I just prepend any additional_instructions to the message, anyone has a better solution? Thanks!

5

Broken — additional_instructions doesn’t seem to do anything.

6

This was all posted 2 minutes ago?

7

Can confirm this is still broken.