Smart Document Q&A Agent

An AI-powered document question-answering system using RAG (Retrieval Augmented Generation). Upload documents and ask questions about their content - get accurate answers with source references and confidence scores.

Features

• Multiple File Formats - Support for PDF, DOCX, TXT, and Markdown files
• Multi-Document Search - Load entire folders and search across all documents at once
• Confidence Scores - See how relevant each source is (0-100%)
• Source Highlighting - Know exactly which file and section the answer came from
• Conversation Memory - Follow-up questions work naturally
• Two Interfaces - Command-line (CLI) or Web UI (Streamlit)

Screenshots

Web Interface

┌─────────────────────────────────────────────────────────────┐
│  SIDEBAR                    │  MAIN CHAT AREA               │
│  ─────────                  │  ─────────────                 │
│  📄 Document Q&A            │  💬 Chat with Your Documents   │
│                             │                                │
│  [Upload Documents]         │  User: What is this about?     │
│   Drag & drop files         │                                │
│                             │  Assistant: Based on the       │
│  📚 Loaded Documents        │  document, it discusses...     │
│   • report.pdf (12 chunks)  │                                │
│   • notes.txt (8 chunks)    │  📚 Sources (4 chunks)         │
│                             │    report.pdf | 92% match      │
└─────────────────────────────────────────────────────────────┘

How It Works (RAG Pipeline)

Documents (PDF, DOCX, TXT, MD)
     |
     v
[1. Text Extraction] --> Extract text from each file
     |
     v
[2. Chunking] --> Split into ~1000 char chunks with overlap
     |
     v
[3. Embeddings] --> Convert chunks to vectors (384 dimensions)
     |
     v
[4. Vector Store] --> Store in ChromaDB for fast similarity search
     |
     v
[5. User Question] --> Convert question to vector
     |
     v
[6. Retrieval] --> Find most similar chunks + confidence scores
     |
     v
[7. Generation] --> Send chunks + question to Claude
     |
     v
Answer with Sources & Confidence Scores

Why RAG?

• You can't feed a 500-page document directly to an AI (too large, too expensive)
• RAG finds only the relevant parts and sends those to the AI
• Results in accurate, grounded answers based on YOUR documents

Quick Start

1. Clone the Repository

git clone https://github.com/AlonNaor22/Smart-Document-QA-Agent.git
cd Smart-Document-QA-Agent

2. Set Up Virtual Environment

# Create virtual environment
python -m venv venv

# Activate it
# Windows:
venv\Scripts\activate
# Mac/Linux:
source venv/bin/activate

3. Install Dependencies

pip install -r requirements.txt

4. Add Your API Key

Create a .env file in the project root:

ANTHROPIC_API_KEY=your_api_key_here

Get your API key from console.anthropic.com

5. Run the Application

Option A: Command Line Interface

python main.py

Option B: Web Interface (Recommended)

streamlit run app.py

Supported File Formats

Format	Extension	Description
PDF	.pdf	Adobe PDF documents
Word	.docx	Microsoft Word documents
Text	.txt	Plain text files
Markdown	.md	Markdown files

Usage

CLI Commands

Command	Description
[any question]	Ask about your document(s)
quit or exit	Leave the application
new	Load different document(s)
clear	Clear conversation history
help	Show available commands

Loading Documents

Single File:

Enter the path to your document OR folder
> C:\path\to\document.pdf

Multiple Files (Folder):

Enter the path to your document OR folder
> C:\path\to\folder

The system will automatically find and load all supported files in the folder.

Project Structure

Smart-Document-QA-Agent/
├── main.py                 # CLI entry point
├── app.py                  # Web UI (Streamlit)
├── requirements.txt        # Python dependencies
├── .env                    # API key (create this)
├── src/
│   ├── config.py           # All settings in one place
│   ├── document_loader.py  # Multi-format document loading
│   ├── vector_store.py     # Embeddings & ChromaDB
│   └── qa_chain.py         # Q&A logic with Claude
├── data/                   # Place your documents here
└── chroma_db/              # Vector database (auto-created)

Configuration

All settings are in src/config.py:

Setting	Default	Description
CHUNK_SIZE	1000	Characters per text chunk
CHUNK_OVERLAP	200	Overlap between chunks
TOP_K_RESULTS	4	Number of chunks to retrieve
MODEL_NAME	claude-sonnet-4-5	Claude model to use
TEMPERATURE	0.0	Response randomness (0 = deterministic)

Technologies Used

• LangChain - RAG pipeline orchestration
• Anthropic Claude - LLM for question answering
• ChromaDB - Vector database
• HuggingFace - Embeddings model (all-MiniLM-L6-v2)
• Streamlit - Web interface
• PyPDF - PDF text extraction
• docx2txt - Word document extraction

Understanding Confidence Scores

When you ask a question, each source chunk shows a confidence score:

Score	Meaning	Color (Web UI)
80-100%	Highly relevant - strong match	Green
60-79%	Good relevance - likely useful	Orange
Below 60%	Lower relevance - may be tangential	Red

Learning Resources

This project is documented for learning purposes. Each source file contains:

• Detailed docstrings explaining the concepts
• Inline comments explaining WHY, not just WHAT
• Educational notes section with tips for extending

Key concepts covered:

• RAG (Retrieval Augmented Generation)
• Text embeddings and vector similarity
• Prompt engineering
• Conversation memory
• Multi-document retrieval

License

MIT License - Feel free to use this project for learning or as a starting point for your own applications.

Author

Built as a portfolio project to demonstrate RAG implementation with LangChain and Claude.