This project implements a complete retrieval-augmented generation pipeline in native Python for educational purposes. It handles document loading and chunking from multiple formats (PDF, DOCX, TXT, etc.), stores embeddings in FAISS for vector search, and combines results with BM25 keyword retrieval. The system supports local models via Ollama or cloud inference via SiliconFlow API, includes cross-encoder reranking for result quality, and optionally integrates web search via SerpAPI for real-time information augmentation.