Basically, I’d like to be able to create assistants that can ingest files by using tools rather than by making users upload them directly or by uploading them at assistant-creation-time.
A simple example would be an assistant with access to the file system that would have ls, cd, read_file, etc. tools, with the read_file tool adding a new file to the thread when it is run.
I know I could add files to the assistant when executing the tool, but I’d like to scope the files to the specific thread (as if the user uploaded it in a message).
Agreed, I don’t want to pass back a huge JSON string when it just needs to be loaded into a CSV and used in Code Interpreter
Totally agree, was just about to make the exact same post but saw yours. Another thing that I also would find useful is being able to somehow update the tools parameter when required_action.type is “submit_tool_outputs”. Would make it so much easier to implement custom tool retrieval for assistants and guide the workflow.
The solution I’ve found is create a function that relieves the file and when called uploads the file to the client, it responds with the file_id then needs to be passed back and update the assistant with the file_id and then continue the run
@bmorrow35 how did you make the assistant understand that the “tools output” was successful and that he should expect the results as file(s) in following messages to resume his processing?
I can’t figure out which kind of output, function definition or instructions I should put in place for this to work. The assistant always gets confused in one way or another.
I initialize the assistant with an empty file_id parameter , then use the modify assistant function after the tool output and pass the file_id that was just generated. Probably not the most ideal solution, but I have no coding background and this is what I found works. I create a new assistant each time I start a thread. I believe there’s a way to update the file_ids in the run but I already had this coded in when I came across that, so I’ve stuck to it. Will probably try to move to updating the run instead in the future
from openai import OpenAI
client = OpenAI()
my_updated_assistant = client.beta.assistants.update(
“asst_abc123”,
instructions=“You are an HR bot, and you have access to files to answer employee questions about company policies. Always response with info from either of the files.”,
name=“HR Helper”,
tools=[{“type”: “retrieval”}],
model=“gpt-4”,
file_ids=[“file-abc123”, “file-abc456”],
)
print(my_updated_assistant)
Ah okay, thanks for the reply @bmorrow35. Restarting the thread or re-creating the assistant is not a viable option for what I’d like to achieve.
It is indeed possible tu upload new files and make reference to them in following runs but I can’t figure out a way to properly set the instructions for the assistant to make the leap of waiting for a following run with new files references to continue his initial processing.
AFAICT this is a serious shortcoming for the “functions-calling” tool, it might be by design or not – only OpenAI could answer that.
Are you deleting the assistant at the end of the action, or just accumulating a bunch of unused assistants? Are essentially ‘containerized assistants’ a thing?
How I enabled files as tool outputs without modifying the assistant:
1- Cancel the run client.beta.threads.runs.cancel when tool is invoked.
2- Introduce a new message to indicate that the tool has been activated, specifying its name and parameters.
3- Follow up with an additional message containing the file_ids , confirming successful execution. Include a description of the file and its schema.
4- Initiate a new run within the thread, incorporating these updates.
been deleting them. have a trigger to delete all assistants and files when i ‘terminate’