What is Retrieval Augmented Generation (RAG)?

Retrieval-Augmented Generation (RAG) is an AI framework that enhances the capabilities of a Large Language Model (LLM) by connecting it to an external, authoritative knowledge base. The process involves two main stages: first, an information retrieval system fetches relevant documents or data snippets from the knowledge base based on the user's query. Second, the LLM (the generator) uses this retrieved context, along with the original query, to synthesize a more accurate, detailed, and verifiable response. This methodology grounds the model's output in factual data, reducing hallucinations and allowing it to use information it was not originally trained on.

A. This describes fine-tuning or retraining, a process that modifies the model's internal weights, whereas RAG augments the model's knowledge at inference time without altering the model itself.

C. This is a generic description of what many Transformer-based LLMs do. It is not specific enough to define RAG, which is characterized by its distinct retrieval component.

D. This incorrectly equates RAG with fine-tuning. RAG is an alternative approach to fine-tuning for incorporating domain-specific knowledge, operating during inference rather than training.

1. NVIDIA Developer Blog. (2023). What Is Retrieval-Augmented Generation? NVIDIA Corporation. "Retrieval-augmented generation (RAG) is an AI framework for retrieving facts from an external knowledge base... The process has two phases: retrieval and content generation." [Link: https://developer.nvidia.com/blog/what-is-retrieval-augmented-generation/]

2. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems 33 (NeurIPS 2020). (Section 2: Methods, p. 3). "RAG models combine a pre-trained parametric memory, a sequence-to-sequence model, with a non-parametric memory, a dense vector index of Wikipedia, accessed with a pre-trained neural retriever." [Link: https://proceedings.neurips.cc/paper/2020/file/6b493230205f780e1bc26945df7481e5-Paper.pdf]

3. Stanford University. (2023). CS224N - Lecture 12: Retrieval and Question Answering. Stanford NLP Group. The lecture describes RAG as a system that first "retrieves a set of relevant documents" and then uses a "Reader/Generator" to produce an answer based on those documents. (Slide 53, "Retrieval-Augmented Generation (RAG)").