Darvin - A RAG Study Buddy I Built For My Friend Drowning in PDFs

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

Darvin is a terminal + FastAPI RAG chat where you give it the path to a PDF (or upload via API), and it answers questions only from that PDF, with citations.

Who I built it for: My friend is preparing for semester exams and has 10+ bulky PDFs (notes, previous papers, manuals). Ctrl+F wasn't cutting it, and uploading everything to ChatGPT felt wrong for his personal notes. He needed a patient practice partner that would quiz him and answer strictly from his material without hallucinating.

So I built Darvin for him: drop in his PDFs, ask in plain English, get grounded answers + follow-ups with chat memory.

What he said after trying it: "Bro this actually cites the page number, now I don't have to scroll 200 pages."

Demo

• Test video: (youtube)[https://www.youtube.com/watch?v=r6MwwP4BkHU]
• GitHub: https://github.com/Rishavj90/darvin

Demo flow:

1. python test.py -> prompt >
2. Ingest PDF via POST /upload (checks .pdf, parses with PyMuPDF, chunks, stores)
3. Ask via POST /ask or terminal loop -> answer + sources + contexts + updated history

Code

https://github.com/Rishavj90/darvin

Stack: Python, FastAPI, LangChain, langchain-groq, ChromaDB, sentence-transformers, PyMuPDF

Key files:

• backend/api.py - /upload and /ask routes
• backend/ingestion/parse.py, chunk.py, store.py - PDF -> pages -> chunks -> Chroma
• backend/query/result.py - top-5 retrieval + grounded prompt + ChatGroq
• backend/query/memory.py - rewrites follow-up into standalone question using last 5 turns
• test.py - terminal chat loop

How I Built It

Open-source AI at the core - 2 layers:

1. Open-weight model: llama-3.1-8b-instant (Meta Llama 3.1 8B weights). I run inference through Groq for speed, but the brain itself is open weights - I can swap GROQ_MODEL in .env to any other Llama / Mistral / Gemma endpoint without changing code.
2. Open-source harness + retrieval: LangChain + langchain-text-splitters + ChromaDB + sentence-transformers. No black-box vector API. Embeddings are local (EMBEDDING_MODEL), stored locally (CHROMA_PATH, COLLECTION_NAME).

Pipeline:

PDF -> PyMuPDF parse -> LangChain splitter chunk_pages -> sentence-transformers encode -> Chroma store
Query -> encode -> Chroma top-5 -> build context with [Source, Page] -> Llama-3.1 prompt (strict: "ONLY use context, say 'I cannot find...' otherwise, cite after each statement") -> answer
Follow-up -> memory.get_standalone_question() rewrites using history -> ask()

.example.env:

EMBEDDING_MODEL=
CHROMA_PATH=
COLLECTION_NAME=
GROQ_API_KEY=
GROQ_MODEL=llama-3.1-8b-instant

Why terminal first? My friend lives in terminal, and I had ~1 weekend. FastAPI is already there so I can add a Streamlit / MERN UI later (I already ship MERN on Render).

Why Does Open Innovation Matter?

This project only works because the pieces are open:

1. Swappable brain: Closed API would lock the prompt + behavior. With Llama open weights, if Groq pricing/limits change, I move the same prompt to Ollama locally, Together, or self-hosted vLLM. My friend can even run a smaller quantized Llama fully offline in hostel with no internet.
2. Inspectable retrieval: LangChain splitters + Chroma + sentence-transformers are all open. When answers were bad, I could actually debug chunk size, embedding model, and top-k=5. Closed RAG-in-a-box hides that.
3. Zero cost to gift: No per-seat SaaS. He clones the repo, sets 5 env vars, runs pip install -r requirements.txt. Data stays in his ./chromaDB and ./tmp_files, not on someone else's server.
4. No hallucination by design: The system prompt forces grounding + citations. Because I control the harness, I can enforce "do not use outside knowledge" - something you can't guarantee with a generic chatbot.

Closed would have been faster for a demo, but open made it giftable, debuggable, and portable.

My Agent Session

Built mostly by hand + terminal testing. No DevRelay session to embed.

Prize Categories

None - not using a partner tech this weekend. Entering for overall + completion badge.