DPR

Dense Passage Retrieval (DPR) uses dual-encoder architectures to outperform traditional BM25 keyword matching in open-domain question answering.

DPR replaces sparse vector lookups with dense embeddings generated by two BERT-based encoders (one for queries, one for contexts). By mapping text into a continuous 768-dimensional vector space, it captures semantic relationships that keyword-based systems miss. Facebook Research demonstrated this efficacy on the Natural Questions dataset, where DPR achieved a Top-20 retrieval accuracy of 79.4 percent (a 20 point jump over BM25). It serves as the standard retrieval backbone for modern Retrieval-Augmented Generation (RAG) pipelines.

https://github.com/facebookresearch/DPR