RocketQA

Why we included this project

Most open dense retrievers are built around English, so a toolkit that ships pretrained models for both English and Chinese is a useful find. RocketQA pairs ERNIE-based dual encoders for candidate retrieval with cross encoders for re-ranking, giving you the two-stage setup that production retrieval systems usually run instead of a single embedding model. Dropping it into an existing stack is straightforward: load a model, encode queries and passages, and index the vectors with Faiss or JINA, with worked examples for both. The repo also carries the training code and baselines behind the published RocketQA papers, handy if you want to fine-tune on your own domain data or reproduce the reported results. Teams building Chinese knowledge bases, enterprise search, or retrieval-augmented generation get the most out of it, since strong open dense retrievers for that language are comparatively rare.