DensePhrases

Why we included this project

DensePhrases, from Princeton NLP, takes a different approach to retrieval: rather than ranking whole documents or passages, it indexes billions of phrase-level vectors over Wikipedia, so a question can be answered directly with a short phrase. The same model can also return broader units like sentences or full passages when you need context for downstream work. The repo includes pre-trained models, pre-built phrase indexes, and examples that wire retrieved passages into a Fusion-in-Decoder QA model, entity linking, or knowledge-grounded dialogue. The ACL and EMNLP papers lay out the design in detail. It is a research implementation rather than a maintained production service, so treat it as a strong reference for building or studying dense retrieval, not as something to drop straight into a live stack.