A new retrieval method uses sparse token matching to extract local evidence from stored documents without additional training. By aggregating matching scores from query tokens, it improves retrieval performance across eight benchmarks.
HOW THIS AFFECTS YOU
●
builderYou can improve the accuracy of your RAG systems by leveraging existing contextual token representations.
●
researcherThis method demonstrates that token-level evidence can be successfully utilized for retrieval without re-training encoders.