[FEATURE] Integration with late chunking model

[heemin32]

Is your feature request related to a problem?

Contextual text chunking maintain the entire context of the document in each chunk so that the retrieval accuracy can be improved. However, because we have to make a call to LLM to generate a context aware chunking, it is not scalable with large dataset. The late chunking model might be able to overcome the scaling issue while maintaining the context of entire document for each chunk.

What solution would you like?

Seamless integration of late chunking model in neural search plugin

What alternatives have you considered?

Customer can generate the embedding outside of OpenSearch

Do you have any additional context?

https://jina.ai/news/late-chunking-in-long-context-embedding-models/