kandpal-2023-rag-3-bm25-paragraphs

IN premise — summaries/2026/08/24/kandpal-2023-long-tail-knowledge-s6-conclusion-and-future-work.md

Created 2026-08-25T02:58:07+00:00

Retrieval-augmented generation using 3 BM25-retrieved paragraphs prepended to the prompt substantially improves long-tail QA accuracy across all tested model sizes.

Summary

Giving a model three relevant paragraphs found by a simple keyword search, pasted right before the question, sharply improves its ability to answer rare or obscure questions, and the boost holds whether the model is small or large. In practical terms, this means you can fix long-tail accuracy cheaply through retrieval rather than by scaling up model size or switching architectures.