How Tableside actually works.

Reasoning-Based Retrieval for Multi-Hop Question Answering

Raveesh Gupta, Ronn Chinowutthichai

PageIndex retrieval reaches ~92% Hits@10 on MultiHop-RAG by demoting the vector index to a coarse document filter and letting an LLM navigate each document's structure tree, then reranking with a recall net. Controlled ablations separate what actually moves multi-hop retrieval from folklore — and the error analysis shows the residual gap is temporal reasoning, not hop count.