A flexible RAG-based document question-answering system that supports both local processing (via Ollama) and cloud processing (via Google Gen AI). Choose the option that best fits your privacy, performance, and hardware needs.

• Multi-Mode Operation: Switch between local processing (private, offline), Google Gen AI
• Multiple File Support: PDFs, Word, CSV, and text files
• Lightweight Design: Optimized for low-power machines
• Privacy-First: Local mode keeps all data on your device
• Fast Cloud Option: Google Gen AI modes for superior performance when internet is available

pda/
├── __init__.py  
├── app.py # Main application with mode selection
├── cli.py # CLI display
├── config.py # app config magic numbers 
├── document_processor.py # Document loading and processing
├── document_store.py # Document retrieval
├── domain.py # QA service
├── error_handler.py # Make errors readable
├── hybrid_retriever.py # Retriever pipeline
├── llm_factory.py # Factory functions to setup LLMs
├── query_cache.py # cache implementation
├── rag_system.py # RAG implementation
├── requirements.txt # Project dependencies
├── requirements-dev.txt # dev dependencies
├── web_app.py # Streamlit interface
├── documents/ # Corpus
└── tests/
    ├── __init__.py
    ├── conftest.py
    ├── generate_test_data.py
    ├── pytest.ini
    ├── README.md
    ├── run_tests.sh
    ├── test_document_processor.py
    ├── test_document_store.py
    ├── test_hybrid_retriever.py
    ├── test_llm_factory.py
    ├── test_query_cache.py
    ├── test_rag_system.py
    └── test_retrieval_performance.py

mkdir pda
cd pda

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

pip install -r requirements.txt

python -c "import nltk; nltk.download('punkt_tab'); nltk.download('averaged_perceptron_tagger_eng')"

langchain>=0.1.0
langchain-community>=0.0.20
langchain-chroma>=0.1.0
langchain-core>=0.1.0
langchain-huggingface>=0.1.0
langchain-ollama>=0.1.0
sentence-transformers>=2.2.0
rank-bm25>=0.2.1
chromadb>=0.4.0
tqdm>=4.65.0
python-dotenv>=1.0.0
numpy>=1.21.0
huggingface-hub>=0.16.0
torch>=1.9.0
transformers>=4.21.0
ollama>=0.1.0
python-docx>=1.0.0
streamlit>=1.28.0

1. Add your documents:

   • Create a documents folder in your project
   • Add PDFs, Word, CSV, and text files you want to query

2. Run the application:

python app.py
||
streamlit run web_app.py

3. Choose your preferred mode when prompted:

   • Local: Private, offline processing using Ollama
   • Google Gen AI: Cloud-based processing using Gemini models

• Download from ollama.ai
• Pull a lightweight model:

# For low-power machines:
ollama pull phi3:mini

# For better quality (requires more RAM):
ollama pull llama3.1:8b-instruct-q4_K_M

1. Go to Google AI Studio
2. Create a new API key
3. Add it to your .env file:

GOOGLE_API_KEY=your_google_api_key_here

model_options = {
    "fast": "gemini-2.0-flash",      # Fastest, most efficient
    "balanced": "gemini-2.0-flash-lite",    # Better quality, still efficient
    "high_quality": "gemini-2.5-flash" # Best quality
}

• phi3:mini (3.8B) – Best for low-power systems
• llama3.1:8b-instruct-q4_K_M (8B) – Better quality (needs 8GB+ RAM)

• Adding document management (add/remove documents)
• Implementing conversation history
• JSON grammar constraint (Ollama feature)
• Strip images, headers, footers before chunking
• Add support for Excel

• Local mode issues: Check Ollama documentation and verify models are downloaded
• Google Gen AI issues: Check API key and internet connection
• Performance issues: Try smaller models or reduce chunk sizes

• Project Gutenberg
  • 60,000+ public domain eBooks (literature, history)
  • Format: Plain text (.txt)

• Internet Archive
  • Millions of books, texts, and documents
  • Format: PDF/TXT (filter for "Texts" and public domain)

• USA.gov
  • U.S. government publications (reports, laws, policies)
  • Format: PDF/TXT (public domain)
  • Examples: NASA, DOE, and federal agency archives