[2401.15884] Corrective Retrieval Augmented Generation

View a PDF of the paper titled Corrective Retrieval Augmented Technology, by Shi-Qi Yan and three different authors

View PDF
HTML (experimental)

Summary:Giant language fashions (LLMs) inevitably exhibit hallucinations because the accuracy of generated texts can’t be secured solely by the parametric information they encapsulate. Though retrieval-augmented technology (RAG) is a practicable complement to LLMs, it depends closely on the relevance of retrieved paperwork, elevating considerations about how the mannequin behaves if retrieval goes fallacious. To this finish, we suggest the Corrective Retrieval Augmented Technology (CRAG) to enhance the robustness of technology. Particularly, a light-weight retrieval evaluator is designed to evaluate the general high quality of retrieved paperwork for a question, returning a confidence diploma primarily based on which totally different information retrieval actions might be triggered. Since retrieval from static and restricted corpora can solely return sub-optimal paperwork, large-scale net searches are utilized as an extension for augmenting the retrieval outcomes. In addition to, a decompose-then-recompose algorithm is designed for retrieved paperwork to selectively concentrate on key data and filter out irrelevant data in them. CRAG is plug-and-play and might be seamlessly coupled with varied RAG-based approaches. Experiments on 4 datasets protecting short- and long-form technology duties present that CRAG can considerably enhance the efficiency of RAG-based approaches.