RAG without Vector-DB? - Yes

About:

This project introduces a novel approach to Cosmos data analysis, diverging from conventional techniques reliant on embeddings and vector databases. The methodology employed involves a TFID retriever, complemented by a meticulous long context reordering and a flash reranker. This innovative approach results in a substantial improvement in retrieval speed, particularly noteworthy with a parameter setting of k=8, while concurrently achieving accuracy levels closely aligned with those obtained through vector databases.

Check out the live demo on Hugging Face

Comparison: VectorDB vs. No VectorDB:

S.No		RAG with VectorDB	RAG without VectorDB
1.	Methodology	Vector Embeddings	TFID Retriever + LC Reorder + FlashReRanker
2.	Storage Requirements	Higher storage demand with additional computation	Minimal storage needed solely for the data
3.	Retrieval Speed	Rapid retrieval with VectorDB	Moderately fast retrieval without VectorDB
4.	Retrieval Accuracy	Achieves k <= 5	Attains k <= 8

Architecture Diagram:

Special thanks to Prithiviraj Damodaran for developing a light-weight and powerful re-ranker.