An open-source collection of tutorials and runnable scripts for over 30 advanced Retrieval-Augmented Generation (RAG) techniques, including Graph RAG, Agentic RAG, and various retrieval optimization strategies, implemented primarily with LangChain.

RAG_Techniques Project Overview#

Project Positioning#

RAG_Techniques is an open-source project initiated and maintained by Nir Diamant, dedicated to becoming the "technical encyclopedia" in the RAG field. The project provides not only theoretical explanations but also executable code based on Jupyter Notebooks and standalone Python scripts. Content ranges from simple vector retrieval to complex Graph RAG and Agentic RAG architectures.

Core Problems Solved#

Insufficient Accuracy: Addresses inaccurate retrieval in traditional simple RAG when handling complex queries
Context Loss: Resolves context fragmentation after chunking long documents
Evaluation Difficulty: Provides standardized evaluation methods (DeepEval, GroUSE) to measure RAG system performance
High Implementation Barrier: Reduces difficulty of integrating advanced RAG techniques into production through modular Notebooks and utility functions

Technology Categories (34 Techniques)#

Basic RAG Techniques#

Simple RAG - Entry-level retrieval-based QA system for beginners
Reliable RAG - Adds verification and refinement mechanisms to ensure accuracy
Optimizing Chunk Sizes - Text chunking optimization balancing context retention and retrieval efficiency
Proposition Chunking - Decomposes text into concise, complete, meaningful sentences

Query Enhancement Techniques#

Query Transformations - Query rewriting, step-back prompting, sub-query decomposition
HyDE (Hypothetical Document Embedding) - Generates hypothetical questions to improve query-data matching
HyPE (Hypothetical Prompt Embeddings) - Pre-computes hypothetical prompts, improving retrieval accuracy by up to 42%

Context Enhancement Techniques#

Contextual Chunk Headers - Creates document-level and paragraph-level context headers
Relevant Segment Extraction - Dynamically constructs multi-chunk text segments
Semantic Chunking - Splits documents based on semantic coherence rather than fixed size
Contextual Compression - Compresses retrieved information while preserving query-relevant content

Advanced Retrieval Methods#

Fusion Retrieval - Combines keyword search and vector search
Reranking - LLM scoring, cross-encoders, metadata-enhanced sorting
Hierarchical Indices - Dual-layer system of document summaries and detailed chunks
Multi-modal RAG - Multi-modal retrieval (caption generation and Colpali methods)

Advanced Architectures#

Graph RAG with LangChain - Knowledge graph integration
Microsoft GraphRAG - Microsoft's open-source GraphRAG system
RAPTOR - Tree-organized recursive abstraction processing for retrieval
Self-RAG - Adaptively decides whether to use retrieved information
Corrective RAG (CRAG) - Dynamically evaluates and corrects retrieval process
Agentic RAG with Contextual AI - Building production-grade agentic RAG pipelines

Application Scenarios#

Financial Document Analysis: Agentic RAG for complex query resolution
Knowledge Base QA: Enterprise-level document retrieval and QA systems
Multi-modal Content Processing: PDF, PPT and other multimedia data processing
Complex Reasoning Tasks: Multi-hop questions and complex query processing

Project Structure#

RAG_Techniques/
├── all_rag_techniques/              # Jupyter notebook implementations
├── all_rag_techniques_runnable_scripts/  # Runnable Python scripts
├── data/                            # Sample data
├── evaluation/                      # Evaluation methods
├── images/                          # Technical diagrams
├── helper_functions.py             # Core helper functions
└── CONTRIBUTING.md                 # Contribution guidelines

Core Helper Functions#

encode_pdf(): Encodes PDF documents to vector store
retrieve_context_per_question(): Retrieves context for questions
bm25_retrieval(): BM25 retrieval
exponential_backoff(): Exponential backoff retry mechanism

Quick Start#

# Clone repository
git clone https://github.com/NirDiamant/RAG_Techniques.git

# Navigate to technique of interest
cd all_rag_techniques/[technique-name]

# Configure environment variables
export OPENAI_API_KEY="your_key_here"

Core Dependencies#

langchain-community
langchain-text-splitters
langchain-openai
langchain-core
pydantic
rank_bm25
PyMuPDF (fitz)
numpy
faiss-cpu

Configuration Capabilities#

Embedding Providers#

class EmbeddingProvider(Enum):
    OPENAI = "openai"
    COHERE = "cohere"
    AMAZON_BEDROCK = "bedrock"

Model Providers#

class ModelProvider(Enum):
    OPENAI = "openai"
    GROQ = "groq"
    ANTHROPIC = "anthropic"
    AMAZON_BEDROCK = "bedrock"

Notebook Structure Standards#

Each technique Notebook includes: Title and overview, detailed explanation, visual representation (Mermaid diagrams), implementation steps, usage examples, comparative analysis, and reference resources.

Agents Towards Production: Production-grade GenAI agent building
GenAI Agents Repository: AI agent implementation collection
Prompt Engineering Techniques: Prompt engineering strategies

RAG Techniques