Advanced RAG Concept: Improving RAG with Multi-stage Document Reranking

[@stillness0072]

Very well explained, thank you

[@_Han-xk1zv]

Are you familiar with the reasons for conducting re-ranking step? Specifically, given the premise of extracting relevant document candidates using only DPR, I'm curious about your perspective on why we'd need to conduct re-ranking using a cross-encoder, in addition to extracting relevant document candidates by computing cosine similarity with a bi-encoder.

[@TechVizTheDataScienceGuy]

Bi-encoder doesn’t really takes into account inter-attention calculation between sentences whereas cross attention does. Ideally one should use cross attention for first step as well, it’s just that it gets really expensive to do so. So get top k using fast method then re-arrange then with better method. Thanks

[@_Han-xk1zv]

@@TechVizTheDataScienceGuy Thanks for your constructive feedback!

[@faiqkhan7545]

Hi amazing video, can you make one to demonstrate a practical approach to this ? like a semantic search tool kind of prototype

[@TechVizTheDataScienceGuy]

Sure, let’s see. Have put in the idea list 👍

[@giridharreddy7011]

Does any library or framework do this out of the box

[@TechVizTheDataScienceGuy]

Did you check llamaindex, haystack or langchain? The should have some logic of reranking, if not this one exactly.

[@himanshukumarsharma9098]

Awesome.... Thank you so much

[@TechVizTheDataScienceGuy]

You’re welcome!