RAG

 

If you use LLMs like ChatGPT, you might find that on some days ChatGPT provides excellent answers with accuracy, but on other days churns out arbitrary answers with hallucination.  This is an inherent problem of data inconsistency in LLMs, which might be caused by out-of-date information that LLMs were trained on.

Enter Retrieval-Augmented Generation (RAG), a solution that connects LLMs to a data store and supplements LLMs with knowledge that can be either open-source like the Internet or closed-source like a collection of documents.

RAG was first proposed in Facebook's 2020 research paper “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” by Patrick Lewis et al., as an enhancing technique of the transformer model.  Some benefits of the RAG framework include:  (1) Avoid retraining LLMs, instead augmenting LLMs with updated information, (2) Data can be updated without incurring significant costs, and (3) Provide references for the information sources.

Let’s ask GPT-4 to describe the difference between RAG and a transformer-based language model.

Currently, RAG is still in its early phase.  It's a cost-effective way to enhance LLMs' capability for relevant, up-to-date information without retraining the model.  Implementing RAG can increase user trust (information sources can be verified) & improve user experience (more accurate answers and fewer hallucinations).  Can't wait to see how RAG will contribute to the evolution of AI in the years to come.