##What is RAG?
RAG (Retrieval Augmented Generation) is a technology architecture that combines information retrieval with language model generation capabilities. Its core idea is to retrieve relevant documents from an external knowledge base before allowing the large model to answer questions, and provide these documents as context to the model, thereby generating more accurate and evidence-based answers.
##The three core components of RAG
**1. Index construction**
Firstly, it is necessary to divide the knowledge document into paragraph blocks (Chunks), convert them into vectors through embedding models, and store them in a vector database. Commonly used vector databases include Milvus, Pinecone, Chroma, etc.
**2. Search strategy**
After the user asks a question, the system also converts the question into a vector and retrieves the most similar document fragment from the vector database. Retrieval can adopt various strategies: semantic retrieval, keyword retrieval, or mixed retrieval.
**3. Generate fusion**
Combine the retrieved document fragments with user questions to form an Augmented Prompt, and input it into a large language model to generate the final answer. In this way, the model no longer relies solely on knowledge from its own parameters, but answers based on the actual documents retrieved.
##The Value and Limitations of RAG
RAG effectively addresses two core issues of large models: knowledge deadline (inability to obtain new information after training) and illusion (fabrication of facts). However, RAG also faces challenges such as unstable retrieval quality and long context processing. The paper "Retrieval Augmented Generation for Knowledge Intensive NLP Tasks" published by Lewis et al. in 2020 laid the theoretical foundation for RAG.
[Reference source] Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", NeurIPS 2020； Official documentation of related open source projects.