아래는 7 Agentic RAG System Architectures에서 얘기한것인데 이해에 도움이 되어 정리해 봅니다.
- Agentic RAG Router
“Go through this door to find your answer.”
- Query Planning Agentic Rag
“Stay on the line, let me check in with a few experts before I can give you a complete answer.”
- Adaptive RAG
“Now that you ask the question this way, let me pull this special book from the library, along with other sources to answer you.”
- Agentic Corrective RAG
“Let me get rid of the information you don’t need before giving you a final answer”
- Self-Reflective RAG
“Let me check, does it even make sense? Hold on, let me see if anything needs correcting before answering you.”
- Speculative RAG
“I remember seeing this before from that source, but let me doublecheck before confirming for you.”
- Self Route Agentic RAG
“Sit back and relax, I know the best way to look for this answer. Be right back.”
how multi-agent agentic RAG systems work을 참조합니다.
-
single agent에 비하여 복잡한 workflow와 여러 종류의 query 를 지원
-
agent가 central orchestration을 수행하고 특정 목적에 최적화된 검색 agent들을 관리
Four Essential Agentic Design Patterns을 참조합니다.
4개의 중요 agent design 패턴들은 아래와 같습니다.
- Reflection Pattern
This pattern enables AI agents to learn from their experiences through:
- Initial response generation
- Self-analysis of outputs
- Continuous improvement through iteration
The reflection mechanism allows agents to enhance their decision-making capabilities over time, leading to more refined and accurate responses.
- Planning Pattern
This strategic approach helps agents handle complex tasks by:
- Breaking down large tasks into manageable components
- Setting clear objectives and goals
- Implementing dynamic strategy adjustments
- Replanning when necessary
This pattern excels in scenarios requiring systematic problem-solving and adaptability.
- Tool Use Pattern
The Tool Use Pattern extends AI capabilities by:
- Integrating with external tools and resources
- Managing multiple tool interactions
- Processing information from various sources
This framework allows agents to perform tasks beyond basic text generation, significantly expanding their practical applications.
- Multi-Agent Pattern
This collaborative pattern enables:
- Coordination between specialized agents
- Efficient task delegation
- Combined expertise in complex problems
Perfect for scenarios requiring diverse skills, like software development or project management.
Evolution of RAG: From Standard to Speculative Architectures을 참조합니다.
- Standard RAG: The Foundation
The conventional RAG approach follows a straightforward path: retrieve relevant documents and feed them directly into a large language model (LLM). While this laid the groundwork for knowledge-enhanced AI systems, it faces a significant challenge. The LLM must simultaneously process retrieved documents and generate coherent responses, often leading to inefficient resource utilization and potential accuracy issues.
- Self-Reflective RAG: Adding Metacognition
Building upon the standard architecture, Self-Reflective RAG introduces a crucial capability: self-awareness. The system learns to evaluate the relevance of retrieved documents, essentially developing a form of metacognition. This advancement allows for more refined document selection, though it requires additional instruction-tuning and computational resources to maintain this reflective capability.
- Corrective RAG: The Quality Control Layer
Corrective RAG takes document evaluation a step further by incorporating a dedicated Natural Language Inference (NLI) model. This external validator classifies documents as Correct, Ambiguous, or Incorrect before they reach the main LLM. While this enhanced quality control improves accuracy, it introduces additional processing overhead that can impact response times.
- Speculative RAG: The Game-Changer
Enter Speculative RAG, representing a paradigm shift in RAG architecture. Its innovative two-tier approach separates the knowledge-processing pipeline into distinct phases:
