Three level of RAG

1. Basic RAG (Naive RAG)

Core Workflow:

Document Loading → Text Splitting → Text Embedding →

Knowledge Retrieval → Prompt Construction → LLM Q&A

• Characteristics: The most classic and straightforward implementation, with a linear and simple process.

• Advantages: Low implementation cost, ideal for quickly building a prototype for validation.

• Limitations: Weak ability to handle complex or ambiguous queries, prone to issues like "retrieving irrelevant information" or "missing key information".

---

2. Advanced RAG

Core Upgrades: Introduces advanced processing techniques on top of the basic workflow, mainly including:

• Query Rewrite: Optimizes vague, colloquial user questions into more precise formulations better suited for retrieval.

• Intent Recognition: Determines the user's true intent behind a question to select different retrieval or generation strategies.

• Rerank: Reorders initially retrieved results by relevance to improve the quality of information fed to the LLM.

• Prompt Engineering: Refines the prompts input to the LLM, guiding it to better utilize retrieved information.

• Query Expansion: Expands the original query with synonyms and related concepts to broaden the retrieval scope.

• Advantages: Significantly improves retrieval precision and final answer accuracy, enabling it to handle more complex business scenarios.

---

3. Modular RAG

Core Concept: Decouples each step of the RAG workflow (e.g., splitting, embedding, retrieval, reranking, generation) into independent modules for separate research and optimization.

• Characteristics:

  • Flexible Composition: Applications can select the optimal solution for each step based on requirements (e.g., using RecursiveCharacterTextSplitter for splitting, bge-m3 for embedding, CrossEncoder for reranking).

  • Easy Iteration: Upgrades to a single module (e.g., replacing with a better embedding model) do not affect other modules, facilitating continuous optimization.

  • High Customization: Dedicated modules can be designed for specific domains (e.g., law, healthcare).

• Representative Directions: Approaches like DSP and ITER-RETGEN mentioned in the diagram are concrete practices of the modular philosophy.