Extend ExampleSelector add_example functionality with the Indexing API

[lmitlaender]

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

The idea is to use the existing Indexing - Langchain API when adding examples to an existing example selector or something similar, to achieve deduplication and potentially clean up of examples added to the vector store.

Motivation

Even when only supporting "None" it would at least already make use of the deduplication feature of the Indexing API.
Extending it further to support the other cleanup Modes would be even better of course.

Currently there is, at least to my knowledge, no good way to make sure the examples you add to an example selector vector store collection aren't duplicated in the future directly integrated into LangChain. It is not good practice to keep recreating the collection completely when there isn't a need for it when something like the Indexing API does exist already.

Proposal (If applicable)

A simple way of integrating the Indexing API, at least for the None cleanup option, would be this:

from langchain_core.documents.base import Document
from langchain.indexes.base import RecordManager
from langchain.indexes import index

class SemanticSimilarityExampleSelector(BaseExampleSelector, BaseModel):

...

    def add_examples_indexed(self, examples: List[Dict[str, str]], record_manager: RecordManager):
        """Add examples to selector vector store using indexing"""
        documents: List[Document] = []
        for example in examples:
            if self.input_keys:
                string_example = " ".join(
                    sorted_values({key: example[key] for key in self.input_keys})
                )
            else:
                string_example = " ".join(sorted_values(example))
            documents.append(Document(page_content=string_example, metadata))
        
        index(docs_source=documents, record_manager=record_manager, vector_store=self.vectorstore, cleanup=None)

...

Of course this isn't refined in any big way at all, but is at least something that could already make use of the Indexing API while also allowing to add multiple examples at a time to an already existing example selector.