I need advice on accepting a JSON list of strings as an input to an API. For example:

["ABC", "DEF"]

Rather than an object:

{
  "item_arr": [
    "ABC",
    "DEF"
  ]
}

I'm upgrading from BentoML 1.2.x where the old syntax supported a list of inputs by specifying a JSON() input in the decorator:

@svc.api(input=JSON(), output=JSON(pydantic_model=ResponseV1), route="/predict")

However, as I try to use the new syntax with a class, I'm unsure of what the

input_spec

should be.

@bentoml.api(route="/2.0/predict", input_spec=???)
    def predict(self, input_arr: List[str]):

The only examples I see of lists as inputs are using a numpy ndarray, which I don't need here. I needed to upgrade versions partially because I needed to enforce the process count -- which didn't seem to be something I could do with the older version of bentoml:

@bentoml.service(
    resources={"cpu": "1"}
)

Maybe I overlooked the mechanism?

Does this works for you? https://docs.bentoml.com/en/latest/build-with-bentoml/iotypes.html#pydantic

Thanks for trying, but it does not. As far as I can tell, it's impossible to make bentoml 1.3.x take input that looks like:

["ABC", "DEF"]

Those examples would only work for:

{
  "item_arr": [
    "ABC",
    "DEF"
  ]
}

For the moment, I've worked around the issue, but it appears to be a limitation that I'll need to work with in the future if we are going to upgrade.

@Frost Ming what is your suggestion here?

not supported in 1.2 style io descriptors for now

Is there a way to do it with 1.3? (we have it working in 1.2 by describing it as

JSON()

and not specifying whether it is an array, object, or whatever.)

You have to implement it in a FastAPI and mount it until we support it officially

thanks