Implemented a retriever for R2R

[RamXX]

Hello team,

This PR implements a retriever for R2R (RAG-to-Riches) as a new retriever option in DSPy.

R2R is a flexible RAG solution that combines vector search, hybrid search, and knowledge graph capabilities, specifically implementing GraphRAG as described in the original Microsoft paper, offering enhanced retrieval options for DSPy users.

Features

The R2R features exposed in this PR are:

• Very flexible implementation, capable of receiving parameters at instantiation or later during forward()
• Vector and hybrid search support with configurable parameters, including kwargs for full control of the input
• Knowledge graph-based retrieval with local search options (global being refactored by R2R)
• Dual operation modes: direct search and RAG generation by the R2R server
• Collection management with name/UUID resolution
• Comprehensive error handling and validation everywhere
• Full type hints and docstrings documentation

Example

    >>> rm = R2RRetrieve(collection_name="My Collection")
    >>> dspy.configure(lm=lm, rm=rm) # lm previously defined
    >>> retriever = dspy.Retrieve(k=10)
    >>> query="What is machine learning?"
    >>> topK_passages = retriever(query).passages

Currently out of scope

• Agent searches
• Ingestion or any non-query operation
• Metrics and assessments of previous searches (can be implemented by the user)
• Chunk metadata included in R2R responses (can be implemented in future releases)

I'm not affiliated with this team, but I am using their product in at least one project and I believe our community can benefit from having this opportunity available. I'm open to any suggestions and willing to keep maintaining it as needed.

Thank you for your consideration.

--Ramiro Salas, a.k.a. DrWho? (DSPy & R2R Discord)