Understanding Retrieval-Augmented Generation (RAG): Concepts, Evolution, and Types

Retrieval‑Augmented Generation (RAG) is a technique that enhances large language models (LLMs) by retrieving external knowledge before generation, thereby addressing two major LLM issues: knowledge cutoff (the model only knows what was in its training data) and hallucination (producing plausible‑looking but factually incorrect text).

RAG works by first retrieving relevant documents from external sources such as knowledge bases, vector databases, or web search APIs, and then feeding these documents to the LLM to produce more accurate and reliable answers. This process can be seen as giving the model a "reference book" or a "second brain" to consult when it lacks sufficient knowledge.

The development of RAG can be divided into several stages:

Naive RAG : The simplest form, using a single retrieval method (full‑text or vector search) to fetch documents that are directly fed to the LLM. It suffers from limited semantic understanding, poor output quality, and difficulty optimizing performance.

Advanced RAG : Improves the three phases of retrieval – pre‑retrieval (document quality enhancement, index optimization, query rewriting), retrieval (domain‑fine‑tuned embeddings), and post‑retrieval (reranking and context compression) – to achieve more accurate and relevant results.

Modular RAG : Decomposes retrieval and generation into reusable components, allowing mixed retrieval strategies, tool/API integration, and flexible engineering for specific domains.

Graph RAG : Incorporates graph‑structured indexes to enable multi‑hop reasoning and richer context, especially useful for tasks requiring relationship understanding.

Agentic RAG : Employs LLM‑based agents that can dynamically decide whether and how to retrieve, invoke tools such as search engines or calculators, and iteratively refine results, making the system more intelligent and adaptable.

The table below summarizes the characteristics and advantages of each RAG type:

RAG Type

Features

Advantages

Naive RAG

- Single index (TF‑IDF, BM25, vector search)

- Simple to implement

- Mitigates hallucination

Advanced RAG

- Document enhancement

- Index optimization

- Query rewriting

- Reranking

- More accurate retrieval

- Enhanced relevance

Modular RAG

- Mixed retrieval

- Tool/API integration

- Modular engineering

- Greater flexibility

- Adapts to diverse scenarios

Graph RAG

- Graph‑based index

- Multi‑hop reasoning

- Context enrichment via graph nodes

- Relational reasoning

- Suited for structured data

Agentic RAG

- LLM‑based agents

- Dynamic decision‑making

- Automatic workflow optimization

- Higher retrieval accuracy

- Handles complex, multi‑domain tasks

Future RAG research is expected to focus on two main directions: (1) increasing intelligence, where more sophisticated agentic capabilities will make RAG an even better partner for LLMs, and (2) supporting diversified data modalities (text, graphs, code, images) within a unified retrieval framework.

Original author: Zhang Wenjun.