9 Essential RAG Design Patterns You NEED to Know

Key Concepts

• Retrieval Augmented Generation (RAG)
• Naive RAG
• Query Transformation with RAG Fusion
• Iterative Retrieval
• Adaptive Retrieval
• Agentic RAG
• Hybrid RAG
• Multi-Agent Systems (Supervisor/Sub-Agent, Sequential Agents)
• Agentic Routing
• Deterministic Designs
• Agentic Designs
• Vector Stores
• Knowledge Graphs
• SQL Databases
• Large Language Models (LLMs)
• Context Window

Deterministic RAG Designs

This section outlines RAG architectures that follow predefined logic without dynamic decision-making by the LLM.

1. Naive RAG:

   • Description: This is presented as the foundational or baseline RAG architecture.
   • Application: Suitable for straightforward scenarios like FAQs.
   • Limitation: Often insufficient for production-level AI systems, implying a need for more sophisticated approaches.

2. Query Transformation with RAG Fusion:

   • Methodology: Involves decomposing and expanding the user's original query.
   • Objective: To fetch a broader and more diverse set of relevant document chunks. This aims to improve the comprehensiveness of the retrieved information.

3. Iterative Retrieval:

   • Methodology: Employs multiple retrieval passes.
   • Application: Specifically designed for handling complex user questions that may require progressively refined or additional information. This contrasts with single-pass retrieval.

4. Adaptive Retrieval:

   • Methodology: Incorporates a decision-making component to determine the necessity and type of retrieval.
   • Decision Points:
     • Whether retrieval is needed at all.
     • If retrieval is required, whether a simple or a multi-stage retrieval process is appropriate.
   • Benefit: Optimizes resource usage and retrieval strategy based on query complexity.

Agentic RAG Designs

This section focuses on RAG architectures where the LLM is empowered with decision-making capabilities.

5. Agentic RAG:

   • Description: The LLM is granted decision-making authority.

   • Capabilities: The LLM can autonomously decide:

     • What information to retrieve.
     • How many times retrieval operations should be performed before generating an answer.

   • Benefit: Allows for dynamic and context-aware retrieval strategies.

   • Hybrid RAG (Variant of Agentic RAG):

     • Description: An extension of Agentic RAG.
     • Capabilities: The agent can retrieve information not only from vector stores but also from knowledge graphs and SQL databases.
     • Benefit: Enables access to structured and semi-structured data in addition to unstructured text, providing a richer information base.

6. Multi-Agent Systems:

   • Architecture: Involves a main agent and multiple sub-agents.
   • Benefits:
     • Cognitive Load Distribution: Distributes complex tasks among specialized sub-agents.
     • Context Window Protection: Prevents the main agent's context window from being overwhelmed, allowing for more efficient processing of user questions.
   • Variants:
     • Supervisor/Sub-Agent: A hierarchical structure where a main agent delegates tasks to sub-agents.
     • Sequential Agents: Agents are arranged in a sequence, allowing for a pipeline of specialized processing steps. This enables multiple specialist AI agents to work collaboratively in a defined order.

7. Agentic Routing:

   • Methodology: Intelligently directs a user's query to the most suitable agent for processing.
   • Objective: To ensure that each query is handled by the agent with the most relevant expertise or capabilities.
   • Principle: Emphasizes that there is no universal RAG pattern; selection is crucial.

Conclusion

The video presents nine RAG patterns categorized into deterministic and agentic designs. Deterministic patterns like Naive RAG, Query Transformation with RAG Fusion, Iterative Retrieval, and Adaptive Retrieval offer structured approaches. Agentic patterns, including Agentic RAG, Hybrid RAG, Multi-Agent Systems (Supervisor/Sub-Agent and Sequential Agents), and Agentic Routing, empower LLMs with decision-making capabilities and sophisticated task distribution. The core takeaway is that the choice of RAG pattern is highly dependent on the specific use case, with no single solution fitting all requirements.