When developing a GPT agent, that environment will allow you to call external functions that are described with OpenAPI YAML format. That works fine with my GPT. But trying to get that same functionality fails in the Assistants playground when working with the OpenAI API framework. First, to use code interpreter you have to convert it from YAML to JSON. Then I don’t see how to call an external api. Is the code interpreter ONLY to be used to call uploaded snippets of text that I provide?
Same question here. I want to upload the OpenAPI spec for my external API just like GPTs.
I’m trying to do the same thing with all the GPT-4 limits I keep hitting. It really is ridiculous how quickly I hit the limit due to the GPT not following the openapi specification. (POST with requestBody parameters)
From my reading, I don’t see how the playground would work with a custom function call. The docs state that when a function call needs to occur it pauses execution.
The function calls also don’t seem to support any openapi spec. It appears as a simple function definition, but it seems like some external interaction is required to handle the function call then provide the output back to the assistant.
I really wish the assistant and custom GPT was more aligned and you could convert a GPT to an assistant.
Follow up: I was able to get something in the Assistants playground kind of working. You have to define the function as just a simple function with input parameters defined, but it doesn’t seem to support the full openapi specification that I can tell. I wan’t able to find anywhere that actually documents what it does support and there isn’t a direct correlation to how the ChatGPT plugins were defined either.
So, what was an api call i turned into a simple function call. In the assistant playground when it called the function, it pauses execution and waits for you to provide the input data manually. So, I took the input parameters and used postman to get the data. I manually copied it into the chat prompt area in the playground and added a property of
"success": true to the data that was returned from postman.
Then I submitted the data and it proceeded on with the process.
What I haven’t tried yet, but was wondering if it was possible, is whether the assistants can use the code interpreter to execute an api call. Again, I don’t see much documentation on what the code interpreter can actually do. It talks about reading in data files and doing data analysis, but never addresses specifically what libraries it has access to.