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