Unboxing RAG: one Python module at a time

What it does Local_Pdf_Chat_RAG is a learning-oriented document Q&A system built as a native Python RAG pipeline. It ingests PDFs, Word files, spreadsheets, Markdown, HTML, and plain text, splits them into chunks, and answers questions by retrieving relevant context from a local vector store before handing it to a language model. The whole thing is wrapped in a Gradio interface that can talk to either a local Ollama instance or the SiliconFlow cloud API.

The interesting bit The project treats pedagogy as architecture: the core/ directory is literally numbered by the RAG pipeline—document loader, text splitter, embeddings, FAISS vector store, BM25 index, hybrid retriever, reranker, generator—so you can read the code in the order the data flows. That transparency is rare; most tutorials give you a black-box chain, but here you can swap out the reranker or watch chunks render in the UI.

Key highlights

• Hybrid retrieval combining FAISS semantic search with BM25 keyword search, plus optional cross-encoder or LLM reranking
• Broad format support: PDF, TXT, DOCX, MD, HTML, CSV, XLS, and XLSX
• Automatic LLM backend detection: prefers SiliconFlow if an API key is present, falls back to local Ollama, or nudges you to configure one
• Optional SerpAPI web search to augment answers with live context
• Document chunk visualization inside the Gradio UI, so you can inspect how the splitter carved up your files

Verdict This is a teaching tool first and foremost: ideal if you want to learn RAG by reading clean, labeled pipeline code rather than framework abstractions. If you are already shipping production RAG systems, you will likely find the feature set familiar rather than novel.