You Do Not Need a Vector Database (For RAG): Benchmarking IR Methods

[ylow]

The document retrieval problem for RAG is basically a case for information retrieval and there are simpler solutions to do so. Vector embeddings are still useful, but they should be used in a later stage of the IR pipeline and not as the first stage retrieval, for which there are simpler and more performant solutions.

Github link to notebook, and blog post: https://github.com/xetdata/RagIRBench/