🧠 Documind — AI-Powered Document Q&A

Ask questions. Get answers grounded in your document. No hallucinations.

Live Demo → docu-mind-ai.netlify.app

---

What is Documind?

Documind is a full-stack Retrieval-Augmented Generation (RAG) application that lets you upload any PDF and have a real conversation with it. It retrieves only the most relevant chunks from your document, passes them to an LLM, and returns answers that are strictly grounded in the source — refusing to answer when the evidence isn't there. Documind is the productionized evolution of RAG-built-from-scratch-v3 — taking the core retrieval pipeline and wrapping it into a fully deployed web application.

Built entirely from scratch — no LangChain, no LlamaIndex, no high-level RAG frameworks.

---

How It Works

PDF Upload → Text Extraction → Chunking → Embedding → Vector Storage
                                                              ↓
User Query → Embed Query → Cosine Similarity Search → Top Chunks
                                                              ↓
                                              LLM → Grounded Answer

Each component is a standalone Python module:

Module	Role
loader.py	Extracts raw text from uploaded PDFs
chunker.py	Splits text into overlapping semantic chunks
embedder.py	Converts chunks to vectors using sentence-transformers
vector.py	Stores and manages the in-memory vector index
retriever.py	Finds top-k most relevant chunks via cosine similarity and keyword search through BM25
app.py	FastAPI backend — handles requests, orchestrates the pipeline
index.html	Frontend UI with streaming response support

---

Features

• 📄 Upload any PDF and query it instantly
• 🔍 Semantic retrieval using sentence-transformers/all-MiniLM-L6-v2 and keyword search though "BM25"
• 🚫 Hallucination control — refuses to answer when similarity score falls below threshold
• ⚡ Streaming responses for a real-time chat feel
• 🌐 Fully deployed — backend on Render, frontend on Netlify

---

Tech Stack

Backend: Python, FastAPI, Sentence Transformers, scikit-learn, PyPDF2, NumPy, BM25 Frontend: HTML, CSS, JavaScript (vanilla)
Deployment: Render (backend) · Netlify (frontend)

---

Hallucination Control

If the similarity score of the best-matching chunk is below 0.5, the system refuses to answer and reports insufficient evidence. This ensures every response is traceable back to the document.

---

Local Setup

git clone https://github.com/akop-cyber/Documind.git
cd Documind
pip install -r requirements.txt
uvicorn app:app --reload

Then open index.html in your browser and point the API URL to localhost:8000.

---

Deployment Architecture

Frontend (Netlify)  ──→  Backend API (Render)
     index.html              app.py (FastAPI)
                               ↓
                         RAG Pipeline
                     (loader → chunker → embedder
                      → vector → retriever → LLM)

---

Author

Built by Aarav Kumar Ranjan Portfolio · GitHub · Kaggle

---

This project does not use LangChain, LlamaIndex, or any external RAG framework. Every component — chunking, embedding, retrieval, similarity scoring — is implemented manually.