🧠 Mindoc — Private Offline RAG Assistant

Mindoc is a privacy-focused, fully offline AI assistant that allows you to search and chat with your PDFs and PPTs locally. No cloud, no API keys — everything runs 100% on your device using efficient Small Language Models (SLMs).

---

🚀 Key Features

🔒 100% Offline

• No OpenAI
• No cloud dependencies
• No data leaves your device
• All models stored locally

📄 Multi-Document Ingestion

• Upload multiple PDFs and PPTs
• Fully private, local processing
• Fast and accurate extraction

🧠 Local LLM

Powered by LaMini-Flan-T5 (248M) optimized for CPU inference.

🔎 Hybrid Search Engine

Semantic Vector Search + Cross-Encoder Reranking

• Vector Model: all-MiniLM-L6-v2
• Reranker: ms-marco-MiniLM-L12-v2

⚡ Dual Search Modes

• Quick Mode: Fast, short answers (Top-2 docs)
• Deep Research: Multi-doc reasoning using Map-Reduce

🔗 Smart Citations

• Evidence-based answers
• Click on a citation → open PDF → auto-scroll to exact page

---

🛠️ Technical Architecture

1. Ingestion Pipeline

• Loader: PyMuPDFLoader
• Chunking: RecursiveCharacterTextSplitter
  • Chunk size: 1000 chars
  • Overlap: 200 chars
• Embeddings: SentenceTransformer (384-dim)
• Storage: ChromaDB (Local persistence)

2. Retrieval & Generation

1. Retrieve top-10 chunks with vector search
2. Re-rank with cross-encoder, keep best 3
3. Feed context → LaMini LLM → generate answer

---

🏗️ System Architecture

       ┌────────────┐
       │   Files    │
       │ PDF / PPTX │
       └──────┬─────┘
              │
              ▼
      ┌────────────────┐
      │ Document Loader│
      └───────┬────────┘
              │Chunks
              ▼
  ┌────────────────────────┐
  │Embeddings (Local Model)│
  └───────────┬────────────┘
              │Vectors
              ▼
    ┌──────────────────┐
    │ Vector Store     │
    │ FAISS / Chroma   │
    └─────────┬────────┘
              │
              ▼
    ┌───────────────────┐
    │ RAG Pipeline      │
    │ (Retrieve + LLM)  │
    └─────────┬─────────┘
              │
              ▼
       ┌─────────────┐
       │  FastAPI    │
       │  /query     │
       └─────────────┘

---

📦 Installation Guide

Prerequisites

• Python 3.10+ (3.12 recommended)
• Node.js & npm

---

🔧 Backend Setup (FastAPI)

cd backend

# Virtual environment
python -m venv venv
source venv/bin/activate    # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Download models (run once)
python download_model.py
python download_reranker.py

# Start backend server
uvicorn app.main:app --reload

🎨 Frontend Setup (React + Vite)

cd frontend

npm install
npm run dev

🖥️ Usage

Upload Documents

• Drag & drop PDFs
• Supports batch uploads
• Wait for “✅ Indexed” confirmation

Chat with Your Documents

⚡ Quick Mode

• Fast
• Lightweight
• Best for direct questions

🧠 Deep Research

• Reads many chunks
• Map-Reduce summarization
• Great for reports & summaries

Verify Sources

• Each answer includes clickable citations
• Opens full PDF and auto-scrolls to correct page

---

📂 Project Structure

mindoc/
├── backend/
│ ├── app/
│ │ ├── api/
│ │ ├── rag/
│ │ ├── services/
│ │ └── main.py
│ ├── data/
│ │ ├── chroma/
│ │ ├── models/
│ │ └── uploads/
│ ├── download_model.py
│ ├── download_reranker.py
│ └── requirements.txt
│
└── frontend/
├── src/
│ ├── App.jsx
│ ├── App.css
│ └── main.jsx
└── package.json

❗ Troubleshooting

sqlite3 errors (Python 3.12)

Install SQLite shim:

pip install pysqlite3-binary

Context Window Errors

Long chunks → crash. Fixed by enabling:

truncation=True

500 Search Errors

Usually caused by a missing reranker model. Run again:

python download_reranker.py

---

🔮 Future Roadmap

• OCR for scanned documents
• Model switching (LaMini ↔ Phi-2 ↔ Qwen 0.5B)
• Persistent conversation history
• Voice mode (offline ASR)

---