princeton-nlp/DensePhrases

Enables multi-granularity retrieval across phrases, sentences, passages, and documents using dense vectors indexed over billions of Wikipedia phrases, supporting downstream tasks like entity linking and knowledge-grounded dialogue. Built on transformer-based encoders with pre-trained models available via Hugging Face, it indexes phrase-level representations for real-time retrieval at scale while maintaining flexibility to aggregate results at different semantic levels. The system integrates with open-domain QA pipelines and demonstrates effectiveness in specialized applications including slot filling and document retrieval without requiring task-specific fine-tuning.

606 stars. No commits in the last 6 months.