CourseMate AI: A Document Assistant That Refuses to Make Things Up

The Problem

Ask most chatbots a question about a 40-page report and you'll get a fluent, confident answer — that may or may not actually be in the document. For anything that matters — course notes, contracts, financial statements — "confident and wrong" is worse than no answer at all.

CourseMate AI is a Retrieval-Augmented Generation (RAG) assistant built to solve exactly that problem. Upload one or more PDFs, ask a question in plain English, and get an answer that is grounded strictly in the document's own text — with the exact page it came from attached. If the answer isn't in there, it says so, instead of guessing.

How It Works

The system is a five-stage pipeline that turns a static PDF into a searchable, citable knowledge base:
Upload – The user drops in one or more PDFs (course material, financial reports, anything text-heavy) through a Streamlit interface.

Chunk – Each document is recursively split into overlapping text chunks, small enough to embed precisely but large enough to preserve context.
Embed – Every chunk is converted into a vector using Mistral AI's embedding model, capturing its meaning numerically.
Store & Retrieve – Vectors are persisted in a ChromaDB store. When a question comes in, the system retrieves the most relevant chunks using Maximal Marginal Relevance (MMR) — a retrieval strategy that balances relevance against diversity, so the model sees a well-rounded set of evidence instead of five near-duplicate chunks.
Answer – LangChain assembles the retrieved chunks into a strict, context-only prompt. The LLM is instructed to answer only from what's in front of it, cite the page number, and explicitly say "not found in document" when the evidence isn't there.

Key Features

Zero-hallucination prompting — answers are grounded in retrieved text, not the model's general knowledge.
Page-level citations — every answer links back to where it came from, so users can verify it themselves.
Multi-PDF support — index and query across several documents at once, with live re-indexing when new files are added.
Tunable retrieval — an interactive sidebar exposes the retrieval knobs (k, fetch_k, lambda_mult) so users can trade off relevance versus diversity in real time.
Custom chat UI — a styled Streamlit front end with a source drawer under each answer, showing exactly which passages were used.

Tuning Retrieval: Why MMR Matters

A naive retriever just grabs the top-k most similar chunks to a question — which often means five chunks that all say almost the same thing. MMR instead asks: of the candidates I've found, which set together gives the most complete, least repetitive picture? Getting this right meant experimenting with three parameters:

k — how many chunks are ultimately handed to the LLM
fetch_k — the larger candidate pool MMR selects from before narrowing down
lambda_mult — the dial between pure relevance (1.0) and maximum diversity (0.0)

Small changes here had an outsized effect on answer quality — too little diversity and the model missed context spread across a document; too much and it lost focus on the actual question.

Tech Stack

LayerToolsIngestionPyPDF, LangChain text splittersEmbeddings & LLMMistral AIVector storeChromaDB (MMR search)OrchestrationLangChainInterfaceStreamlit

What's Next

The project currently runs as a local/Streamlit-hosted app. Natural next steps include persistent multi-user document libraries, swappable embedding/LLM providers, and evaluation metrics to track answer faithfulness over time.

Try It

The full source is available on GitHub: Course-mate-AI