Support Pydantic json schema manipulation for tool schemas

[bengladwell]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Add a GenerateJsonSchema field to StructuredTool to allow control over tool schema generation.

Motivation

We have had a lot of difficulty working with OpenAI's Responses API because of Responses API's stricter rules regarding JSON schemas. We have encountered this common problem, which required us modify how we specify optional tool parameters.

We also had problems with tool parameters that are a discriminated union Pydantic type. After much back and forth with OpenAI support, it was determined that we need to remove default values from the discriminated field so that those fields are included in the JSON spec required array.

Rather than make our Pydantic models non-standard, it would be nice to just control the tool schema output. Of course, it is possible to simply generate the JSON schema and pass it directly to SturcturedTool.from_function as the args_schema argument. However, this bypasses the processing of tool args performed by Langchain, such as the removal of injected variables. We would like to have both: Langchain tool arg processing and control over how the schema is generated.

Proposal (If applicable)

Add a field like:

schema_generator: GenerateJsonSpec = GenerateJsonSpec

to langchain_core.tools.structured.StructuredTool.

Then, in StructuredTool.args, generate the json schema like:

json_schema = input_schema.model_json_schema(schema_generator=self.schema_generator)