Difference between TypeAdapter().json_schema and BaseModel.model_json_schema

[Mark90]

Initial Checks

• I confirm that I'm using Pydantic V2

Description

When presented with an annotated type with multiple Field objects that set json_schema_extra, TypeAdapter seems to combine them whereas BaseModel does not.

I was used to the behavior of BaseModel of not combining anything, which I can understand because merging json schema's is non-trivial in the case of conflicting keys. I would expect TypeAdapter to follow the same rules.

Below code example outputs the following

BaseModel.model_json_schema
{'properties': {'v': {'key2': 'value2', 'title': 'V', 'type': 'string'}},  # only contains key2
 'required': ['v'],
 'title': 'Foo',
 'type': 'object'}

TypeAdapter().json_schema
{'key1': 'value1', 'key2': 'value2', 'type': 'string'}  # contains key1 and key2

Example Code

from pprint import pprint
from typing import Annotated

from pydantic import TypeAdapter, BaseModel, Field

MyStr = Annotated[
  str, 
  Field(json_schema_extra={'key1': 'value1'}), 
  Field(json_schema_extra={'key2': 'value2'})
]
class Foo(BaseModel):
    v: MyStr


print("BaseModel.model_json_schema")
pprint(Foo.model_json_schema())
print("TypeAdapter().json_schema")
pprint(TypeAdapter(MyStr).json_schema())

Python, Pydantic & OS Version

pydantic version: 2.6.4
        pydantic-core version: 2.16.3
          pydantic-core build: profile=release pgo=true
                 install path: /Users/mark/.pyenv/versions/3.12.1/envs/312pydantic26/lib/python3.12/site-packages/pydantic
               python version: 3.12.1 (main, Jan  2 2024, 17:23:05) [Clang 15.0.0 (clang-1500.0.40.1)]
                     platform: macOS-14.4.1-arm64-arm-64bit
             related packages: typing_extensions-4.11.0
                       commit: unknown

[sydney-runkle]

@Mark90,

Thanks for reporting this. Definitely a bug - the schemas should match. I think we do actually want to merge the field infos like is done in the TypeAdapter case, but the answer should be somwhere in the fields.py or _fields.py.

Marking as a good first issue and adding to our next milestone!

[nix010]

@sydney-runkle I can work on this one.

[nix010]

@sydney-runkle While working on the issue I found something that I want to clarify.

from pprint import pprint
from typing import Annotated

from pydantic import TypeAdapter, BaseModel, Field


def _set_key2(s):
    s['key2'] = 'value2'


CustomField = Annotated[
    str,
    Field(json_schema_extra={'key1': 'value1'}),
    Field(json_schema_extra=_set_key2),
    Field(json_schema_extra={'key3': 'value3'}),
]


class Foo(BaseModel):
    field: CustomField


print("-1. BaseModel.model_json_schema")
pprint(Foo.model_json_schema())
print("-2. TypeAdapter().json_schema")
pprint(TypeAdapter(CustomField).json_schema())

Result

-1. BaseModel.model_json_schema
fields --  {'root': FieldInfo(annotation=~RootModelRootType, required=True)}
{'properties': {'field': {'key3': 'value3',
                          'title': 'Field',
                          'type': 'string'}},
 'required': ['field'],
 'title': 'Foo',
 'type': 'object'}
-2. TypeAdapter().json_schema
{'key1': 'value1', 'key2': 'value2', 'key3': 'value3', 'type': 'string'}

Note that key2 is set by a callable json_schema_extra

• The 1. result should be the same as 2. (have all keys key1, key2, key3 ) right ?

[sydney-runkle]

@nix010,

Great question. Seems the problem is even worse than I imagined 😂. I'd direct you to this issue to understand some more context behind modifying json schema when merging field info: #7686.

Even though I like the approach of the TypeAdapter more, I do think that changing this would be a breaking change. Let me chat with @dmontagu on Monday to see what he thinks the best next step is.

[sydney-runkle]

Alright, let's go with the TypeAdapter approach. This should also fix #7686 that I've mentioned above.

We'll want to document this as a breaking change, and illustrate to users how they can delete keys via callable json schemas!