Advanced RAG Technique — SubDoc Summary

Considering everything we’ve discussed and we’ve observed so far of the RAG frameworks, I think the best term to describe the techniques is simply:

“common sense”.

Having this attribute places you on a solid track to creating highly efficient and production-ready RAG systems.

Nevertheless, the community has been diligently working on ways to further enhance RAG performance. Learning from others’ “common sense” and experiencing your own “aha!” moments prove invaluable and very practical.

The above diagram is an illustration of a typical flow for a RAG system.

🚀Why Sub-Doc RAG?

The implementation of a RAG system introduces complexities such as the loss-in-the-middle problem and challenges with long document retrieval. We are well aware that naive chunking equates to poor RAG performance, as each chunk lacks global context awareness.

While there are various methods to improve RAG retrieval, today’s focus is on a particularly effective trick known as sub-document summaries — a context augmentation technique that significantly enhances RAG retrieval performance.

Curious to delve deeper into this?