Pydantic models should include discriminated tags to make defining configs easier

[illeatmyhat]

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

LangChain uses Pydantic models for practically every aspect. This is actually great. It makes it really easy to write configs.
But for example, what if you don't know which source of embeddings you want to use for a vector db ahead of time?
Pydantic doesn't know how to deserialize a field without knowing ahead of time which class to use.
To do this, it needs a field to discriminate between different options. This is often called a tag, and the technique is called Discriminated Unions.

Motivation

Currently we have to do something like this, and use a bunch of validation to ensure we actually selected correctly.
And it takes up a lot of vertical space.

class ChromaIndex(BaseModel):
    embedding_type: Literal['huggingface', 'openai']
    huggingface_embeddings: HuggingFaceEmbeddings | None
    openai_embeddings: OpenAIEmbeddings | None
    ...

Proposal (If applicable)

For example, HuggingFaceEmbeddings should include a embeddings_type field

class HuggingFaceEmbeddings(BaseModel, Embeddings):
    embeddings_type: Literal['huggingface_embeddings']
    ...

That way, in our config models, we can simply do

class ChromaIndex(BaseModel):
    embeddings: HuggingFaceEmbeddings | OpenAIEmbeddings = Field(discriminator='embeddings_type')
#  or even
    embeddings: Embeddings.__subclasses__() = Field(discriminator='embeddings_type')
    ...