Retrieval-Augmented Generation (RAG) Documentation Assistant

Most of what a company knows sits inside PDFs: manuals, contracts, specifications, reports. You know the answer is in there; you just can't find it. Ctrl+F only finds the word you typed - if the answer is phrased differently, it stays hidden.

Asking a general AI doesn't solve it either. The model never sees your document; it produces a plausible-sounding answer from general training data. With a contract or a manual, that is worse than no answer, because it looks confident.

What was needed was clear: a tool that reads your own documents, draws the answer out of them, and shows which document it came from.

I built an end-to-end RAG (retrieval-augmented generation) system that turns uploaded PDFs into a queryable knowledge base.

1. It splits documents into meaningful chunks. PDFs are read and divided into overlapping pieces that preserve context. Chunk size and overlap are adjustable from the interface - long contracts and short manuals don't want the same settings.
2. Each chunk becomes a meaning vector. Chunks are passed through an embedding model and stored in ChromaDB, so search works by meaning rather than keyword: it finds the termination clause even if you never typed "termination".
3. The vector database persists on disk. A document set is processed once and never re-embedded in later sessions. Several PDFs can be uploaded at once and queried as a single shared knowledge base.
4. Answers are generated only from retrieved text. When a question arrives, the system pulls the most relevant passages and constrains the model to answer from those alone. It never speaks from its own memory.
5. A usable interface. A Streamlit chat UI with PDF upload, conversation history, and live control over chunk size, overlap, number of retrieved passages and temperature. The whole thing is packaged with Docker.

The project was tested on the "A Short Guide to the EU" document and built entirely on free API tiers, at zero cost.

Where it fits: Internal procedures and quality manuals - a new hire asks "what do we do in this case?" and gets an answer with its source. Contract archives - "what was the penalty clause with this supplier?". Technical specifications and equipment manuals, so field teams find the right clause in seconds. Regulatory and compliance document sets. Past proposals and project documents, for finding precedent when writing a new bid.

Stack: Python, LangChain, ChromaDB, PyPDFLoader, Streamlit, Docker; cloud LLM APIs for embedding and generation.