Sparse embeddings - Benchmarks?

[paulmartrencharpro]

Hello,

I have done a RAG app with the hybrid retrieval with Qdrant & fastembed. I used the prithvida/Splade_PP_en_v1 model on my first implementation and it works very well. Much better than with a standard Qdrant Embedding Retriever.

With the latest version, there's now a second sparse embedding Qdrant/bm42-all-minilm-l6-v2-attentions. From my tests, I can see it's smaller and faster, but I can't quantify if it's better or worst. By testing my whole RAG system, I evaluated that the quality of the answers generated are similar, but that does not really grade the retrieval part of the process.

Is there any benchmarks that I could use to compare the different sparse embeddings that fastembed supports?

Thanks