Use a library to validate basic data types about guardrails HOT 10 CLOSED

One possible way might be to allow users to specify an optional pydantic model in the tag.
The model would need to have fields that matches the field names, and then all validation can be done in the pydantic model instead of in the XML

    <object name="bank_run" pydantic="BankRun">
        <string
            name="explanation"
            description="A paragraph about what a bank run is."
            on-fail-length="reask"
        />
        <url
            name="follow_up_url"
            description="A web URL where I can read more about bank runs."
            on-fail-valid-url="filter"
        />
    </object>

and then

from pydantic import BaseModel
class BankRun(BaseModel):
    explanation: str
    follow_up_url: str

There's many advantages to this, we can specify default values, optional fields, all the other stuff pydantic supports, and the wider python ecosystem in general since pydantic validation is very extensible, people can use this way to use the validators lib to plug into the pydantic schema too.

NOTE: this would also make using the @register_validator redundant

One other effect of this is that users won't need to know arbitrary tags, eg. like <url> or <string>, it'll just all be <field> or similar, all the data types will live in python.
This does loosely tie the RAIL format to Python which may not be preferable, but would be much more natural to work with in Python IMO.

from guardrails.

@cktang88 Nice, I think this would be super useful! I like this framing of the solution and the idea of adding field as a data type.

I do think that supporting primitives like url, code etc. is nice so that the spec isn't too reliant on Python. I plan to add support for rail in other languages in the future. However, to make the support for pydantic work, adding field as a datatype along with the existing data types makes a lot of sense.

I'll get a PR started to add this.

from guardrails.

Started a PR #35 that starts to address this issue (see details there)

from guardrails.

@devenbhooshan this is a pretty neat suggestion!

I'll think some more about what the best interface for integrating existing python validator library could be. One way to do this could be that for supported types (e.g. email, IP address, etc.), validators provided in the python validator lib is supported by default. Wdyt?

from guardrails.

would be cool to integrate w/ Pydantic directly https://docs.pydantic.dev/ which seems to be one of the most common Python schema validation libs

from guardrails.

Pydantic support would be awesome. +1

from guardrails.

@cktang88 @oliverbj could definitely add Pydantic support for schema validation. The Validator classes in Guardrails took inspiration from Pydantic's Validators.

Do you have an idea of what an integration with Pydantic's schema validation would be like within Guardrails?

from guardrails.

Could we generate a models.py file with code describing pydantic models from a guardrails spec?

from guardrails.

@irgolic I think so! I'll have to double check what the edge cases are around nested structures, but it should be doable.

That would make it pretty easy for users so that they don't have to make double effort to first write the spec and then the pydantic models.

from guardrails.

This was added in #35 (docs on how to use Pydantic). I'll close this issue out -- thanks for the suggestion everyone!

from guardrails.