다단계 및 에이전트 기반 RAG 패턴

Welcome to Advanced RAG! So far, we've explored basic Retrieval Augmented Generation (RAG) where an LLM answers a query using a single set of retrieved documents.

However, real-world questions can be complex. They might involve multiple parts, require reasoning, or need to interact with different systems.

Simple RAG struggles with these challenges. That's why we need more sophisticated approaches like Multi-stage RAG and Agentic RAG.

Multi-stage RAG is an advanced pattern designed to handle complex queries by breaking them down into smaller, manageable parts. Instead of one big retrieval, it performs several targeted retrievals.

Think of it like a detective solving a case: they don't just look for one clue; they follow leads, gather more information, and piece it all together step-by-step.

The first step in Multi-stage RAG often involves query decomposition. Here, an LLM analyzes your original complex question and reformulates it into multiple, simpler sub-queries or specific search terms.

This allows the system to retrieve documents that are highly relevant to each distinct part of your overall question, improving accuracy.

Once the original query is decomposed, the RAG system performs iterative retrieval. This means:

• Each sub-query is sent to the retriever.
• Relevant documents are fetched for each sub-query.
• The results from one retrieval step might even inform or refine subsequent sub-queries.

Finally, all retrieved information is combined and sent to the LLM for a comprehensive answer.

Imagine you ask: "What are the main causes of climate change, and what are some recent technological solutions being developed to combat it?"

A Multi-stage RAG system might:

1. Decompose: "Main causes of climate change" and "Recent tech solutions for climate change."
2. Retrieve 1: Find documents on causes.
3. Retrieve 2: Find documents on solutions.
4. Synthesize: Combine info to answer both parts fully.

While Multi-stage RAG improves retrieval, Agentic RAG takes it a step further. An LLM agent is a system where an LLM acts as a 'brain' to reason, plan, and execute actions using a variety of tools.

Instead of just retrieving, an agent can decide what to do next based on the user's query and the available tools. RAG becomes one of its powerful tools!

The core of an agentic system is an LLM that can:

• Understand: Interpret the user's goal.
• Plan: Break down the goal into steps.
• Act: Choose and use appropriate tools for each step.
• Observe: Evaluate tool outputs and decide next steps.

Tools can include web search, calculators, external APIs, and, crucially, a RAG retriever for your knowledge base.

In an Agentic RAG system, the RAG component isn't the whole pipeline; it's a specialized tool the agent can invoke. The agent decides when to use RAG.

For example, if a user asks a question that requires factual information from your specific document store, the agent will 'decide' to use its RAG tool to fetch that context.

Consider the query: "What was our company's Q3 revenue last year, and how does it compare to the industry average for that period?"

An LLM agent might:

1. Use RAG Tool: Retrieve internal Q3 revenue report.
2. Use Web Search Tool: Find industry average Q3 revenue.
3. Use Calculator Tool: Compare the two figures.
4. Synthesize: Provide a comprehensive answer.

When should you use Multi-stage vs. Agentic RAG?

• Multi-stage RAG: Best for queries that can be clearly broken into distinct sub-questions, where the primary challenge is retrieving comprehensive information.
• Agentic RAG: Ideal for highly dynamic, multi-step problems that might require various types of reasoning, external interactions, and tool usage beyond just retrieval.

Both enhance RAG, but for different kinds of complexity.

Let's check your understanding of these advanced RAG patterns.

Great job! In this lesson, we explored advanced RAG patterns for handling complex scenarios:

• Multi-stage RAG: Decomposes complex queries into sub-queries, performing iterative retrieval to gather comprehensive context.
• Agentic RAG: Uses an LLM as an intelligent agent that can reason, plan, and use various tools (including RAG) to achieve a goal.

These techniques are crucial for building robust and intelligent LLM applications that go beyond simple question-answering.