Yuval Belfer argues that agentic search has not eliminated the need to prepare data for retrievalRetrieval is the process of selecting relevant stored information and returning it to an AI system for the current task.. Conventional RAG systems choose chunk sizes during indexing, before they know the future queries, so very large chunks can blur useful details while very small chunks can lose wider context.
Yuval Belfer describes tests that index the same material at multiple chunk sizes across meeting-transcript, narrative-question-answering and other datasets. Focused and context-heavy questions favor different chunk sizes, and an oracle comparison illustrates the recall left on the table by committing to one fixed size.
Yuval Belfer proposes multi-scale indexing: make several differently chunked indexes, retrieve from each, promote matches to document-level rankingsRanking orders candidate results by a chosen measure of usefulness or relevance, helping AI systems present the most suitable items first. and merge those rankings with reciprocal rank fusion. He reports stronger recall across the tested datasets than individual fixed-size baselinesEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria., while acknowledging extra memory use and open questions about how many sizes to retain and whether a better fusion method exists.
Watch on YouTube




