RAG (Retrieval Augmented Generation) is an AI architecture that allows large language models (LLMs) to retrieve and use external and domain-specific data for their responses or outputs. In addition to being trained by the model’s knowledge, RAGs help generate more accurate, context-aware responses because they retrieve relevant documents or information from external sources like a vector database.

Why is RAG important in large language models and AI applications?

The fundamental reason to use a RAG architecture is to ground the LLM to real, enterprise-specific knowledge. This architecture helps with limitations of LLMs, such as hallucinations, during the generation process. 

The RAG can be used for more than retrieval, including query transformation, context formatting, and generation.

What are the components of a RAG? 

The RAG architecture has 4 core and 4 optional components:

1. Retriever 
2. Generator (LLM)
3. Embedding model
4. Indexing layer
5. Query rewriter (optional) 
6. Reranker (optional) 
7. Post-processor (optional) 
8. Memory (optional)

The following table summarizes the components and their purposes with examples.