AI-Powered Document Q&A 📚🔍

AI-Powered Document Q&A is an AI-powered document question-answering application built with Streamlit, LangChain, HuggingFace, and FAISS. It allows users to upload multiple PDF documents and interactively ask questions, receiving concise answers along with relevant source citations.

Check out the Streamlit App here https://gen-ai-document-rag.streamlit.app/

Streamlit shuts down the app after some time of inactivity. Please click on "Wake Up App" button to start the app again.

Demo Video

🚀 Features

• Multi-document Upload: Upload up to 5 PDF documents simultaneously.
• Conversational Q&A: Ask questions in natural language and receive context-aware answers.
• Source Citation: Answers include references to the original documents and pages.
• Relevance Ranking: Utilizes cross-encoder models to re-rank retrieved documents for improved accuracy.
• Memory Management: Maintains conversational context across multiple queries.

🛠️ Tech Stack

• Frontend: Streamlit
• Document Processing: pdfplumber
• Vector Store: FAISS
• Embeddings: Sentence Transformers
• Language Models: LangChain, TogetherAI, LlamaCpp
• Cross-Encoder: sentence-transformers/ms-marco-MiniLM-L-6-v2
• Environment Management: python-dotenv

🧩 Architecture Overview

The application follows a modular architecture:

Document Processing

• PDFs are uploaded and processed using pdfplumber.
• Text and tables are extracted, cleaned, and structured into LangChain Document objects.

Embedding and Vectorization

• Document chunks are created using RecursiveCharacterTextSplitter.
• Embeddings are generated using HuggingFace's Sentence Transformers (all-MiniLM-L6-v2).
• FAISS is used as the vector store for efficient retrieval.

Retrieval and Re-ranking

• Initial retrieval is performed using FAISS.
• A cross-encoder (ms-marco-MiniLM-L-6-v2) re-ranks the retrieved documents based on relevance to the query.

Conversational QA

• LangChain's ConversationalRetrievalChain integrates the retrieval system with a conversational memory buffer.
• Queries are answered using either a cloud-hosted LLM (TogetherAI's Llama-3.3-70B-Instruct-Turbo-Free) or a local LLM (Mistral-7B-Instruct).

Memory Management

• LangChain's ConversationBufferMemory is used to maintain context across multiple queries.

Frontend Interaction

• Streamlit provides an interactive UI for document upload, question input, and displaying answers with source citations.

🚦 Running the App Locally

1. Clone the Repository

git clone <your-repo-url>
cd DocumentQA

2. Configure Environment Variables

Create a .env file in the project root and add your API keys.
Get your API key from https://api.together.ai/ and add it to the .env file:

TOGETHER_AI_API_KEY=<your-api-key>

3. If Docker is Installed

docker compose up

4. If Docker is Not Installed

Set Up Environment

Create a virtual environment and install dependencies:

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Run the Application

streamlit run app.py

Open your browser and navigate to http://localhost:8501.

📌 Usage

• Upload PDF documents using the sidebar.
• Enter your questions in the chat input box.
• View answers along with cited sources and relevance scores.

⚙️ Customization

• Switching LLMs: Modify llm_handler.py to switch between cloud and local models.
• Chunk Size & Overlap: Adjust CHUNK_SIZE and CHUNK_OVERLAP in app.py to optimize retrieval performance.

📖 Dependencies

Key dependencies include:

• streamlit==1.45.0
• langchain==0.3.25
• langchain-community==0.3.23
• sentence-transformers==4.1.0
• faiss-cpu==1.11.0
• pdfplumber==0.11.6 (See requirements.txt for the full list.)

🚧 Future Improvements

• Enhanced UI/UX with additional Streamlit components.
• Support for more document formats (e.g., DOCX, TXT).
• Integration with additional LLM providers.