# Introduction Large language models (LLMs) have become an essential element in the digital world. However, their ability to generate accurate and coherent responses can be limited by the lack of an external knowledge base. To overcome this challenge, researchers have developed a new RAG model that combines generation capabilities with search in a safe and self-improving manner. ## How the new model works The new RAG model uses a multi-stage validation structure to validate generated responses. This structure comprises three key components: * Entailment verification (NLI): verifies if the generated response is coherent with the input text. * Attribute checking: verifies if the generated response contains accurate and up-to-date information. * Novelty detection: verifies if the generated response contains new and interesting information. In this way, the RAG model can ensure that generated responses are accurate, coherent, and up-to-date. ## Results and applications The new RAG model has achieved impressive results on four datasets: Natural Questions, TriviaQA, HotpotQA, and Stack Overflow. Researchers have demonstrated that the model can achieve an accuracy of 40.58% in information retrieval, nearly doubling the performance of standard RAG models. Furthermore, the model has shown a more efficient corpus expansion capability compared to previous methods, with a 72% increase in document addition over naive write-back. ## Conclusion The new RAG model represents a significant step towards creating language models that can learn and improve over time. Thanks to multi-stage validation, the model ensures that generated responses are accurate, coherent, and up-to-date. In summary, the new RAG model is an example of how technology can be used to improve our ability to access information and generate accurate responses in a safe and self-improving manner.