Assistants API not retrieving files while using GPT-3.5

Prompting
, ,
1

Hi!

I’ve been trying to use the assistants API to control my smart home (through Home Assistant). I’ve gotten the function calling all working and it works when given the exact entity ID.

However, I would like to be able to say something like “Turn on the living room lights” and have it automatically fetch the ID - so I created a markdown file with a list of all the entity IDs in it, and prompted the AI to retrieve data from it. But it never seems to call Retrieval no matter what I do, and will just make up it’s own entity ID.

I have gotten this working with GPT-4 (gpt-4-1106-preview), however it is still too expensive for my needs and I’d like to get GPT-3.5 working if possible. (gpt-3.5-turbo-1106)

My current prompt:

“You are a smart home assistant, programmed to control the user’s smart devices. The uploaded ‘devices.md’ file contains the entity IDs for every device - check there before calling a function. Do not assume any values or parameters, such as entity IDs - if something is unknown, call Retrieval or ask the user for clarification.”

Thanks in advance!

3 Likes
2

Can anyone else confirm if this is true? It seems like 3.5-turbo-1106 is unable to retrieve the files I’ve uploaded but 4-turbo-1106 can. I haven’t found any confirmation that it can or can’t but I haven’t been able to get it to. RAG was specified for 4 turbo on some website talking about azures API, but it didn’t say anything about 3.5. Can someone else test it? I could just be doing something wrong.

2 Likes
3

I am facing the same problem with the Retrieval tool using Assistants. It works fine with Gpt 4 preview but gives issues like - slower response, failure of Run etc with GPT 3.5 turbo!

4

I can confirm GPT 3.5-turbo-1106 constantly reporting some unspecified error while trying to read the file attached in the assistant settings.

5

I can confirm this issue.
It works well, and fast, with gpt-4-1106-preview.
But with gpt-3.5-turbo-1106, it takes time and fails.
Since gpt-4 is 10 times more expensive this is a considerable misfortune.
I wonder if it really doesn’t work or if it has something to do with the size or number of files.

6

I dont know if this is related. I uploaded a pdf to assistant gpt-3.5-turbo-1106 and if i just wording slightly, i will get this error

I can't access the file you uploaded. Please provide me with the necessary information or content from the file so that I can assist you effectively. For example, if there are any tracking details, delivery estimates, or company policies related to package deliveries, that would be helpful.

But if i ask a question directly verbatim as in the pdf, then assistant api gives me the info i need. Is this something that can be fixed with custom instructions?

1 Like
7

Same here. GPT 3.5 turbo 1106 crashes when a retrieval had to be done. Also, random function calls are attempted when there’a no need to make them.

8

I tried in the Playground to upload the file while using gpt-4-1106-preview saved, than changed to gpt-3.5-turbo-1106 and saved the assistant. Now it seem that the retrieval works with gpt-3.5-turbo-1106. I don’t know it is a coincidence or you can resolve this bug this way.

9

Looks like gpt-3.5-turbo can only access less than one files at a time.

10

Did you try explicitly adding file id in instruction?

client.beta.threads.runs.create(
  thread_id=thread.id,
  assistant_id=assistant.id,
  instructions = f"Use file with if: {file.id} to answer questions." # This helps to force the model to access the file!
)
11

I’m facing the same issue too ! can’t read from the file using gpt 3.5