Advanced RAG Patterns

Simple RAG (Vector Search → LLM) is often called "Naive RAG". In production, it often fails because search results are messy or irrelevant. To build a world-class system, we need Agentic RAG — systems that can think, verify, and try again.

1. Query Transformation: Fixing the Question 🪄

Often, the user's question is poorly phrased for a search engine. We can use an LLM to rewrite it:

• Multi-Query Retrieval: Generate 5 different versions of the same question and search for all of them. This increases "recall".
• HyDE (Hypothetical Document Embeddings): The LLM writes a "fake" answer first. We then use that fake answer to search for a real document. Surprisingly, this works better than searching with a question!
• Sub-Question Decomposition: Break a complex question ("Compare Python vs Java in 2024") into two searches ("Python in 2024" and "Java in 2024").

2. Corrective RAG (CRAG) & Self-RAG 🛡️

What if the documents retrieved are completely useless? Naive RAG will just hallucinate an answer. Advanced RAG evaluates the quality.

1. Retrieve: Get top 5 documents.
2. Evaluate: A specialized "Grader" LLM checks if the documents are relevant.
3. Correct:
   • If Relevant: Proceed to generation.
   • If Irrelevant: Trigger a Web Search tool to find better info.
   • If Ambiguous: Use Reranking.

3. Reranking: Quality over Quantity 🏆

Vector search might return the top 100 documents quickly, but only top 3 fit in the prompt. Cross-Encoders (Rerankers) like Cohere or BGE are specialized models that look at the query and document together to give an absolute relevance score.

[!IMPORTANT] Reranking is the easiest way to improve RAG accuracy. It filters out the "noise" that semantic search often pulls in.

📚 Official Docs

• Cohere Rerank API — Production reranking service
• LangChain RAG Tutorial — End-to-end RAG pipeline
• LlamaIndex RAG Guide — Advanced query engines

4. Anthropic Contextual Retrieval 🔍

A major problem with naive RAG: chunks lose their context when split from the document.

"The defendant was found guilty." — What case? What crime? Out of context, this chunk is useless.

Anthropic's Contextual Retrieval prepends a short context summary to each chunk before embedding it:

📚 Official Resource

• Anthropic Contextual Retrieval Blog Post

5. Microsoft GraphRAG: Knowledge Graph + RAG 🕸️

Standard RAG retrieves relevant documents. GraphRAG builds a knowledge graph from your documents and enables queries that cross-reference entities and relationships:

When to use GraphRAG vs. Standard RAG:

• Standard RAG: "What does the contract say about payment terms?" → Local chunk retrieval
• GraphRAG: "What are the main themes across all 500 contracts?" → Global graph traversal

📚 Official Resource

• Microsoft GraphRAG Documentation

6. RAGAS: Scientific Evaluation Framework 📊

How do you know if your RAG is 80% or 90% accurate? We use the RAG Triad with RAGAS:

📚 Official Docs

• RAGAS Documentation — RAG evaluation framework
• RAGAS GitHub — Open-source implementation

7. RAG vs. Fine-Tuning: Decision Framework 🔀

Quiz

Key Takeaways

✅ Query Transformations (Multi-query, HyDE, Sub-questions) fix bad search results at the source.
✅ Corrective RAG (CRAG) adds a validation step to prevent hallucinations when retrieval fails.
✅ Reranking (Cohere, BGE) is the "silver bullet" for improving precision.
✅ Contextual Retrieval prepends context to chunks before embedding — reduces failures by 49%.
✅ GraphRAG enables global knowledge synthesis across large document collections.
✅ RAGAS provides a scientific, automated way to evaluate RAG quality.

Official Resources

• 📖 LangChain RAG Tutorial
• 📖 LlamaIndex RAG Guide
• 📖 RAGAS Documentation
• 📖 Microsoft GraphRAG
• 📖 Anthropic Contextual Retrieval

What's Next?

RAG relies on the Vector DB. But how do you handle billions of items without crashing?
Next Chapter: Vector Database Deep Dive: HNSW, Quantization, and Platform Selection.