Docs

DocuMind

Project Overview

Students and professionals waste hours reading documents. DocuMind turns any PDF into a conversational AI assistant. Upload → Ask → Learn in seconds.

Project Structure

Directory Overview
nextjs/— Next.js UI placeholder (quick demo)
rag/— Python RAG (Retrieval-Augmented Generation) building blocks
utils/— Helper utilities (zipping assets, ensure directories, etc.)
config/— Configuration files for modularity and DRY principles

RAG Modules

Core Components

loader.py

PDF loading utilities (uses `pypdf` by default). Extracts raw text from PDFs.

chunker.py

Naive text chunking function to split large documents into smaller chunks suitable for embeddings.

embedder.py

Wrapper around `sentence-transformers` to create vector embeddings for chunks and queries.

vectordb.py

FAISS-based in-memory vector store wrapper with simple add/search APIs.

qa.py

High-level flow that ties the components together: load PDF → chunk → embed → build/search index → return top contexts.

Runtime Flow

How the System Works
  1. 1.User uploads a PDF and submits a question via the Next.js UI.
  2. 2.Frontend sends the file and question to the backend (e.g., a FastAPI endpoint at `POST /ask`).
  3. 3.Backend uses `rag/loader.py` to extract text and `rag/chunker.py` to create chunks.
  4. 4.`rag/embedder.py` converts chunks to embeddings; `rag/vectordb.py` builds a FAISS index and performs similarity search using the question embedding.
  5. 5.`rag/qa.py` returns the top matching chunks and a `combined_context` string. Optionally the backend then calls an LLM (OpenAI / other) with the question + contexts to generate a polished answer.
  6. 6.The frontend displays the answer and optionally shows the supporting contexts.

Backend Setup

API Configuration
Location: backend/app.py (not scaffolded here to keep the skeleton minimal)
Endpoint: POST /ask
Expects: file, question (multipart form)

Next Steps

Development Roadmap
  1. 1.Add `backend/app.py` with a `POST /ask` endpoint that saves uploaded file to a temp folder, calls `rag.answer_question_from_pdf()`, and optionally calls an LLM (OpenAI) to synthesize an answer from the `combined_context`
  2. 2.Add a minimal `Dockerfile` or `Procfile` for easy deployment.
  3. 3.Add sample PDF(s) in `demo/` and a short demo GIF for presentation.
  4. 4.Add unit tests in `rag/tests/` for `chunker`, `embedder` (smoke), and `loader`.

Hackathon Tips

Best Practices
  • •Keep prompts in a single file for quick iteration (e.g., `rag/prompts.py`).
  • •Use `sentence-transformers` + `faiss-cpu` for fast local retrieval without needing a cloud setup.
  • •If LLM quota or latency is a concern, prepare a fallback: pre-generate answers for a few sample PDFs and keep them in `demo/` for a smooth live demo.

Styling Configuration

Glassmorphism Theme

Theme: Glassmorphism

Framework: Tailwind CSS

Global Utilities:

.glass— Glass effect background
.glass-border— Glass effect border
.badge— Badge styling utility

GlassCard Props:

title— string - Card title
subtitle— string - Card subtitle
avatar— string - Avatar URL/icon
children— ReactNode - Card content

Performance Tips

From Configuration
  • •Keep heavy images as optimized webp and serve with next/image for automatic optimization.
  • •For animations, prefer CSS transitions over JS where possible.

Table of Contents