rag

Retrieval-Augmented Generation (RAG) Pattern

Overview

The Retrieval-Augmented Generation (RAG) Pattern grounds agent responses in external, up-to-date knowledge by retrieving relevant documents or data before generating a response. Rather than relying solely on knowledge baked into model weights (which has a training cutoff and may hallucinate), RAG agents dynamically fetch the most relevant context and use it to produce accurate, grounded responses.

Core Principle: Don't hallucinate what you can retrieve — anchor every response in verifiable, retrieved knowledge.

When This Skill Applies

Activate this pattern when:

• The agent needs domain-specific knowledge not in the base LLM's training
• Responses must be grounded in authoritative documents (legal, medical, technical)
• Information changes frequently and training data is stale
• Users need citations and sources for claims made
• Private or proprietary knowledge must be accessed securely
• Reducing hallucination is a critical requirement

Rule of thumb: If the answer exists in a document and you need it to be accurate and verifiable — use RAG.

RAG Architecture

Standard RAG Pipeline

Query → [Embedding] → Vector Search → Retrieved Chunks
                                              ↓
                                   LLM + Retrieved Context
                                              ↓
                                     Grounded Response

Advanced RAG Variants

Variant	Description	Use Case
Naive RAG	Embed query → retrieve → generate	Simple Q&A
Advanced RAG	Query expansion, reranking, filtering	High-accuracy enterprise
Modular RAG	Pluggable retrieval strategies	Complex, multi-source
Agentic RAG	Agent decides when/what to retrieve	Dynamic reasoning
Graph RAG	Knowledge graph + vector retrieval	Complex entity relationships

DEFINE → PLAN → ACTION Workflow

DEFINE

Map the knowledge retrieval requirements:

1. What knowledge sources need to be indexed? (PDFs, databases, APIs, web)
2. What query types will users ask? (factual, comparative, analytical)
3. What is the required accuracy/hallucination tolerance?
4. How frequently does the knowledge change? (indexing strategy)
5. What metadata is available for filtering? (date, source, category)