RAG Pipeline Framework

RAG Pipeline Framework

A modular retrieval-augmented generation framework with pluggable loaders, chunking strategies, embedding backends, vector stores, hybrid search, reranking, and a full evaluation suite.

Build a RAG pipeline that actually works. Not a toy demo — a framework with four chunking strategies, BM25+dense hybrid search, reranking, and RAGAS-style evaluation metrics, all runnable locally with the included in-memory vector store.

What You Get

• Document loaders for plain text, Markdown, HTML, and PDF with metadata extraction
• Four chunking strategies — fixed-size, recursive text splitting, semantic (embedding-based topic detection), and sentence-window
• Embedding abstraction with OpenAI adapter and a deterministic demo embedder for offline testing
• Embedding cache that persists to disk so re-indexing unchanged documents costs nothing
• In-memory vector store with full BM25 keyword search — runs the complete pipeline with zero external services
• Hybrid search combining dense similarity and BM25 via Reciprocal Rank Fusion
• Reranking with a cross-encoder API adapter and a stdlib keyword-boost alternative
• RAG orchestrator that wires retrieval + generation into a single query() call
• Evaluation suite measuring context precision, context recall, faithfulness, and answer relevancy
• YAML configuration for every pipeline parameter
• Comprehensive test suite covering loaders, chunkers, embeddings, retrieval, and evaluation
• Strategy guides for chunking selection and retrieval tuning

File Tree

rag-pipeline-framework/
├── README.md
├── LICENSE
├── requirements.txt
├── src/
│   ├── __init__.py
│   ├── models.py                          # Document, Chunk, RetrievalResult dataclasses
│   ├── orchestrator.py                    # RAG query orchestrator
│   ├── embeddings.py                      # Embedding abstraction (OpenAI + demo + cache)
│   ├── vector_store.py                    # Vector store interface + in-memory BM25 impl
│   ├── loaders/
│   │   ├── __init__.py
│   │   ├── base_loader.py                # Abstract loader interface
│   │   ├── text_loader.py                # Plain text + Markdown loaders
│   │   ├── html_loader.py                # HTML loader (stdlib html.parser)
│   │   └── pdf_loader.py                 # PDF loader (pypdf/PyPDF2)
│   ├── chunking/
│   │   ├── __init__.py
│   │   └── strategies.py                 # Fixed, Recursive, Semantic, SentenceWindow
│   ├── retrieval/
│   │   ├── __init__.py
│   │   └── reranker.py                   # CrossEncoder + KeywordBoost rerankers
│   └── evaluation/
│       └── __init__.py                   # RAGAS-style evaluation metrics + reporting
├── configs/
│   └── pipeline_config.yaml              # Full pipeline configuration
├── tests/
│   ├── conftest.py                       # Path setup
│   └── test_pipeline.py                  # Comprehensive test suite
└── guides/
    ├── chunking-strategy-guide.md        # When to use each strategy + tuning
    └── retrieval-tuning-guide.md         # Dense vs hybrid, reranking, top-k selection

... continues with setup instructions, usage examples, and more.