RAG_Techniques

Implements a standard RAG pipeline architecture with document ingestion, embedding generation, vector storage, semantic retrieval, and LLM-based generation. Uses a modular pattern where each stage (chunking, embedding, retrieval, generation) is independently configurable, allowing developers to swap components (e.g., different embedding models, vector databases, LLM providers) without rewriting the pipeline. The architecture follows a consistent interface across 40+ technique implementations, enabling pedagogical progression from simple RAG to advanced variants.

Implements intelligent document chunking strategies that go beyond fixed-size splitting by using semantic boundaries (sentence/paragraph breaks, code blocks) and configurable chunk size optimization. The technique analyzes document structure to preserve semantic coherence while optimizing for embedding model context windows and retrieval performance. Includes methods to test different chunk sizes against a query workload to empirically determine optimal chunk dimensions, with metrics tracking retrieval quality vs. computational cost tradeoffs.

Implements Self-RAG and Corrective RAG (CRAG) techniques where the system generates answers, then validates them against retrieved context and self-corrects if validation fails. The system uses learned or rule-based validators to assess whether generated answers are supported by retrieved context, and if validation fails, triggers retrieval refinement (new queries, different retrieval strategies) and regeneration. This approach creates a feedback loop within the generation process, enabling the system to detect and correct hallucinations or unsupported claims without requiring external feedback.

Extends RAG to handle multi-modal documents containing both text and images by using multi-modal embedding models that encode images and text into a shared embedding space, enabling retrieval across modalities. The system processes images (extracting text via OCR, generating captions, or using vision models) and text separately, embeds them into a unified space, and retrieves relevant content regardless of modality. This approach enables queries to find relevant images when asking text questions and vice versa, supporting richer document understanding.

Provides a comprehensive evaluation framework (DeepEval) for assessing RAG system quality across multiple dimensions: retrieval quality (precision, recall, NDCG), answer quality (faithfulness, relevance, coherence), and end-to-end performance. The framework includes pre-built metrics, dataset management, and evaluation pipelines that can be integrated into development workflows. Developers can define evaluation criteria, run automated evaluations against test datasets, and track metrics over time to monitor RAG system quality and detect regressions.

Provides standardized benchmark datasets and evaluation protocols for comparing RAG techniques and implementations. The repository includes curated test datasets with queries, expected answers, and ground-truth retrieved documents, enabling developers to benchmark their RAG systems against known baselines. Benchmarks cover different domains (general knowledge, technical documentation, research papers) and query types (factual, conceptual, reasoning), allowing developers to assess RAG performance across diverse scenarios and compare their implementations against published baselines.

Provides parallel implementations of all RAG techniques using both LangChain and LlamaIndex frameworks, showing how the same logical RAG concepts map to different framework abstractions. Each technique has implementations in both frameworks, allowing developers to understand RAG architecture independent of framework choice and to compare framework approaches. This dual-implementation strategy helps developers make informed framework choices and understand how to port RAG implementations between frameworks.

Provides standalone, executable Python scripts for each RAG technique that can be run immediately without modification (with API keys configured). Scripts include all necessary imports, configuration, and error handling, demonstrating production-ready patterns. Each script is self-contained and can serve as a template for implementing the technique in production systems. Scripts include examples with real data, showing end-to-end execution from document loading through answer generation.

Implements multiple query transformation techniques (query rewriting, expansion, decomposition) that improve retrieval by reformulating user queries into forms more likely to match relevant documents. Techniques include HyDE (Hypothetical Document Embeddings) which generates synthetic relevant documents from queries, HyPE which generates hypothetical passages, and multi-query expansion that creates semantically similar query variants. Each transformation is applied before retrieval to increase the likelihood of finding relevant chunks, with optional fusion of results from multiple query variants.

Enhances retrieved chunks with contextual metadata by automatically generating or extracting chunk headers, parent document context, and hierarchical position information. When a chunk is retrieved, the system includes its semantic context (what section of the document it belongs to, what the surrounding chunks discuss) alongside the chunk content itself. This enrichment happens during indexing (headers are computed and stored with chunks) and retrieval (context is appended to retrieved chunks before passing to the LLM), improving the LLM's ability to understand chunk meaning without requiring larger context windows.

Combines results from multiple retrieval strategies (dense semantic search, sparse BM25 keyword search, hypothetical document embeddings) using fusion algorithms (Reciprocal Rank Fusion, weighted scoring) to produce a unified ranked result set. Each retrieval strategy is executed independently, then results are merged using configurable fusion methods that balance semantic relevance (from dense retrieval) with keyword matching (from sparse retrieval). This approach captures both semantic and lexical relevance without requiring a single unified index.

Implements a two-stage retrieval pipeline where an initial retriever (fast, approximate) returns candidate chunks, then a cross-encoder reranker (slower, more accurate) scores and reorders results based on query-document relevance. The reranker uses transformer models that jointly encode the query and document to compute relevance scores, providing more accurate ranking than embedding-based similarity. This approach maintains retrieval speed (initial retrieval is still fast) while improving result quality through expensive but accurate reranking on a smaller candidate set.

Builds multi-level document indices where documents are recursively summarized into hierarchies (leaf chunks → summaries → higher-level summaries) and retrieval traverses this hierarchy top-down. The system first retrieves relevant high-level summaries, then recursively retrieves more detailed chunks from relevant branches, reducing the number of embeddings needed and improving retrieval efficiency. This approach is particularly effective for large document collections where flat indices become inefficient, enabling both faster retrieval and better handling of documents with varying levels of detail.

Implements dynamic retrieval strategies that adapt based on query characteristics, routing different query types to different retrieval methods. The system analyzes incoming queries to determine optimal retrieval strategy (e.g., simple keyword search for factual lookups, semantic search for conceptual questions, graph-based retrieval for relationship queries) and applies the appropriate method. This routing can be rule-based (query classification) or learned (trained classifier), enabling the system to use the most efficient and effective retrieval method for each query type without requiring all queries to use the same strategy.

Implements iterative retrieval where initial retrieval results are evaluated, and based on evaluation (relevance feedback, answer quality assessment), the system refines queries or retrieval parameters and retrieves again. The feedback loop can be explicit (user indicates whether results are relevant) or implicit (system evaluates answer quality and decides whether to retrieve more context). This approach enables the system to improve results through iteration without requiring perfect initial retrieval, particularly useful for complex queries that may need multiple retrieval rounds to gather sufficient context.

Implements RAG using knowledge graphs (GraphRAG, RAPTOR) where documents are converted into structured knowledge graphs with entities and relationships, and retrieval operates on graph structure rather than flat chunks. The system extracts entities and relationships from documents, builds a graph index, and retrieves relevant subgraphs based on query entities and relationship patterns. This approach enables relationship-aware retrieval (finding documents about related entities) and supports complex queries that depend on understanding connections between concepts, not just individual chunks.