@Marvin provide examples on how to use one of your own tooling utilities, namingly

tool_from_model

found in module

marvin.utilities.tools

I couldn't find any specific examples or documentation related to the

tool_from_model

utility in the

marvin.utilities.tools

module on GitHub or in the Prefect-related repositories. It might not be part of publicly available Prefect documentation or issues. To get more information, you might want to check the source code directly if you have access to it, or contact the maintainers if this is part of a private or less-documented library. If you have more context or specifics about where this utility is used, I can try another approach or offer guidance based on general software development practices.

😄 marvin doesn't have its own source code in the vectorstore, just prefect stuff

F

if what does?

Yes that's what I was asking - didn't know OpenAI itself followed the OpenAPI schema 😄

Are functions ran in OpenAI's engine itself?

nope, its basically like marvin lib: hey openai, a user just asked what the weather is in chicago and i have a

get_weather

tool that accepts a

city

param, here's the openapi schema - how should I call the function? openai: cool yeah i'd say you should call

get_weather("chicago")

marvin lib: ok the result of

get_weather("chicago")

is "sunny and 60 degrees" openai: great, producing the final chat message now based on that tool result

Basically only the function signature is passed

yep, represented as an openapi schema

great, gonna experiment creating a shell tool and see how that goes - sharing a thought further on this, I recall not all models support engine-driven function calling (though that's rapidly evolving and more models do). I want to assume imperative function calling implementations isn't and won't be on Marvin roadmap given its simplistic nature, hence leaving such "DIY" use cases to users' own implementations

candidly, there is actually nothing on marvin's roadmap, i am passively maintaining it controlflow is the spiritual successor to the marvin experiment and has broader LLM provider support, if you're going to dive into something, I would pick controlflow marvin is great if you want native types from unstructured inputs without a lot of overhead, but for the full blown "make your own agent with tools" I would recommend controlflow. you can still do structured outputs in controlflow as well

In [1]: import controlflow as cf

In [2]: cf.run("i ate 3 bagels with about 16 oz of cream cheese", result_type=list[int])

Out[2]: [3, 16]

and fwiw im not familiar with any LLM providers that run arbitrary user code server side, I think most frameworks will run user code (that assistants use as tools) client side

That is correct, it's more of a sugar coat whereas instead of having OpenAI engine "invoking" your function, you imperatively invoke it yourself, with a prompt like:

These are your choices:
{{ *functions }}
Which one should be invoked, based on {{ context }}

...and the client goes by to invoke the function manually. No schemas involved. That's the DYI I was referring to

yeah I think this was just how you wrote system prompts before "function calling" existed (which is what we were doing in

marvin

in early 2023), but now frameworks like controlflow or langchain just let you pass the native python function and abstract away the fact that the API gets an openapi schema, which the model is now fine-tuned by the LLM provider to respect

I was hesitant on trying out controlflow because it was not clear how it differs or integrates with prefect, which is the core of my backend

controlflow is itself backed by prefect. so if you run some controlflow example, and then do

prefect dashboard open

you'll see the work happening on your server / workspace