Understanding Model Context Protocol (MCP), Retrieval-Augmented Generation (RAG), and Vector Databases for LLM Integration

MCP

MCP originated from Anthropic's November 25, 2024 article "Introducing the Model Context Protocol" and aims to achieve seamless integration of large language models (LLMs) with external data sources and tools, establishing a secure bidirectional link between models and data.

In simple terms, MCP's goal is to become the "HTTP protocol" of AI, driving standardization of LLM applications.

MCP Hosts : applications such as Claude Desktop, IDEs, or AI services that want to access data or tools via MCP.

MCP Clients : client components that connect one‑to‑one with MCP servers, analogous to database clients in traditional applications.

MCP Servers : programs that implement the MCP protocol to provide specific functionalities.

Local Data Sources : locally stored data accessed directly by MCP servers.

Remote Services : external services accessed by MCP servers, typically via API calls.

With MCP, LLM applications can interact with external resources through MCP Clients and MCP Servers, forming an architecture that links models to data and tools.

RAG

Retrieval‑Augmented Generation (RAG) combines information retrieval with text generation to improve the accuracy and reliability of large language models when answering specialized questions.

All AI models are fundamentally based on probabilistic mathematics; their outputs are the result of numerical computations, which can sometimes produce confident but incorrect statements, especially when the model lacks knowledge in a specific domain.

RAG addresses this "hallucination" problem by retrieving relevant external knowledge before generation, allowing the model to cite sources beyond its training data, thereby enhancing relevance, accuracy, and usefulness.

RAG has evolved through several stages, including Naïve RAG, Advanced RAG, Modular RAG, Graph RAG, and the recent Agentic RAG.

Vector Database

Vector databases are specialized systems designed to store, index, and retrieve high‑dimensional vector data efficiently, making them ideal for handling embeddings of unstructured content such as images, text, and audio.

A key example is Milvus, an open‑source vector database that provides high‑performance storage, indexing, and search for massive amounts of feature vectors, and is widely used in generative AI, recommendation systems, and multimodal retrieval.

Milvus employs a shared‑storage architecture with a clear separation of compute and storage, allowing horizontal scaling of compute nodes. Its architecture consists of four layers: Access Layer, Coordinator Service, Worker Nodes, and Storage Layer, each independently scalable and fault‑tolerant.

300万字！全网最全大数据学习面试社区等你来！

In RAG scenarios, such as searching a large collection of papers to answer a question like "What are the mainstream methods for COVID‑19 vaccines?", the workflow typically involves converting PDFs to text, splitting the text into passages, generating embeddings for each passage, storing them in a vector store, and then retrieving the most relevant passages to provide context for the LLM's final answer.

The vector store is the core component that holds both the vectors and the associated structured or unstructured data, and Milvus is often the database of choice for this purpose.