DataFrame IODescriptor Information - OpenAI swagger

[gaspardc-met]

Hello bentoml team,

My query starts from a somewhat broad concern: how can the user/consumer of the my service know the name/order of the expected input data from the different model API endpoints ? You can stop me right there if the concern is unfounded and a better pattern exists 👍

From a human standpoint, this is possible thanks to the great PandasDataFrame IODescriptor:

• in service.py : @service.api(input=PandasDataFrame(columns=FEATURES), output=PandasDataFrame(columns=[TARGET]))

• in the swagger UI we get a nice display of the columns from the input and output (dummy example):
  

However, programmatically, when one tries to read the docs.json attached to the swagger, the columns are not stored in a list but in a text description under summary or description :

And the schema from the 200 response is pretty uninformative (in this use case) :

Would it be possible to store the columns in a manner that is easy to extract from the docs.json ?
I know the format is constrained, but maybe the unused fields like examples (

BentoML/bentoml/_internal/service/openapi.py

Line 109 in 03ee3e4

# examples=None,

) could be used towards that end ?

Bonus: Anyone else having a different solution to a similar use case ?

Thanks in advance !

Note: currently not impossible, albeit not ideal

        input_features = (
            response.json()["paths"]["/predict"]["post"]["summary"]
            .split("[")[1]
            .split("]")[0]
            .replace(" ", "")
            .replace("'", "")
            .split(",")
        )