Demystifying AIGC, Agents, RAG, and MCP: Core AI Concepts Explained

1. AIGC

AIGC (AI‑Generated Content) refers to the automatic creation of text, images, audio, or video using large models such as GPT‑4, Stable Diffusion, or DALL‑E. The release of ChatGPT in November 2022 sparked a surge of interest in AIGC.

1.1 Multimodal technology

Single‑modal models handle only one data type (e.g., pure text). Multimodal models can process two or more modalities simultaneously, enabling scenarios such as text‑to‑image, text‑to‑video, image‑to‑text, and combined audio‑visual generation.

Text‑to‑image: DALL‑E, Imagen, Stable Diffusion, 腾讯混元文生图

Text‑to‑video: Sora, Stable Video Diffusion

Image‑to‑text (image understanding): GPT‑4V, Gemini, Qwen‑VL

Image‑to‑video: Runway Gen‑2, Stable Video Diffusion

Video‑to‑text: Gemini 1.5, Gemini Pro Vision

1.2 RAG (Retrieval‑Augmented Generation)

RAG combines information retrieval with large language model (LLM) generation. When answering a query, the LLM first fetches relevant passages from an external knowledge base, then generates a response based on both the retrieved context and the original prompt, reducing hallucinations and keeping information up‑to‑date.

Knowledge limitation / outdated data

Hallucinations

Lack of source traceability

Insufficient domain‑specific knowledge

RAG was created to address these shortcomings.

2. Intelligent Agent

An agent is a software entity that perceives its environment, makes autonomous decisions, and takes actions to achieve specific goals. Unlike pure AIGC systems, agents can orchestrate multiple tools via function‑calling, turning generation capabilities into general‑purpose problem solving.

2.1 Function Call model

Function Calling enables LLMs to invoke external tools (e.g., weather APIs, calculators) by generating structured JSON parameters. OpenAI introduced this capability in June 2023 on GPT‑4, and many subsequent models have followed.

GPT‑4 (OpenAI)

Claude‑3 (Anthropic)

Gemini‑2.0 (Google)

DeepSeek‑R1 (DeepSeek)

{
  "name": "get_current_weather",
  "description": "获取指定城市的天气",
  "parameters": {
    "type": "object",
    "properties": {
      "city": {"type": "string", "description": "城市名称"},
      "unit": {"enum": ["celsius", "fahrenheit"]}
    },
    "required": ["city"]
  }
}

Typical three‑step workflow:

Define the function schema (name, description, parameters).

The model decides which function to call and generates the JSON arguments.

Execute the function, return the result to the model, and let it produce the final answer.

2.2 Agent workflow

Agents repeatedly invoke function‑calling models, potentially chaining multiple calls. Example: a travel‑planning agent uses weather, driving, public‑transport, and map tools to generate a complete itinerary.

Platforms such as Coze, Dify, and Tencent Cloud Agent Development Platform allow developers to configure prompts, select plugins, and publish agents without writing code.

3. MCP (Model Context Protocol)

MCP, released by Anthropic on 24 Nov 2024, standardizes communication between LLMs and external tools, turning the traditional M×N integration problem into an M+N model. It acts as a “USB‑C” for AI, enabling reusable, plug‑and‑play tool access.

Key advantages over traditional approaches:

Integration cost: one‑time development, reusable across the ecosystem.

Functionality: supports multi‑tool coordinated task chains.

Openness: open‑source protocol encourages community‑driven tool libraries.

Security: data stays on‑premise with fine‑grained permission control.

Since its release, major cloud providers (AWS, Google, Microsoft, Tencent, Alibaba, Baidu) have adopted MCP, creating a de‑facto industry standard and spawning services such as mcp.so and mcpmarket.

4. Summary

Agents orchestrate core AI primitives—AIGC, RAG, function‑calling, and MCP—to build sophisticated applications that go beyond single‑task generation, turning large language models into versatile, executable assistants.