A RAG pipeline that ingests PDFs, Markdown files, and source code into a hybrid vector store and answers questions via multi-query retrieval, cross-encoder reranking, and parent-child context expansion.
Built with Qdrant, MongoDB, LangChain, Unstructured / Docling, FastEmbed BM25, and a HuggingFace cross-encoder. Models (LLM + embeddings) connect via any OpenAI-compatible API endpoint.
- Setup - installation, config, all arguments, Docker
- Architecture - repo structure, what each part does
- Ingestion - how ingestion works, how to extend it
- RAG pipeline - the retrieval algorithm and design decisions
- Dev reference - day-to-day development commands
- Development Notebook - the notebook used to develop and test the pipeline