Show HN: RAG Architecture for optimizing retrieval volume/relevancy tradeoff
github.comWe propose a RAG architecture that uses hierarchical semantic chunking and graph-based context exclusion to maximize relevant information while minimizing the total volume of retrieved context.
The system recursively splits documents into a hierarchical tree structure and dynamically selects the most optimally-sized chunk from each branch by identifying and excluding redundant ancestors and descendants during the search process.
This approach ensures a higher relevant-to-total information ratio by retrieving diverse segments from across the document without including overlapping or nested chunks Interesting approach. How does this architecture handle the 'lost in the middle' phenomenon when the retrieval volume increases? I've found that simply increasing volume often degrades reasoning quality unless re-ranking is extremely aggressive. we consider this a separate issue that should be addressed using special approaches. however, the proposed technique generally tends to decrease the probability of something being lost in the middle by minimizing the total size of the retrieved documents.