docling

📚 Book QA System - Streamlit UI

A Streamlit-based web application for asking questions about PDF books using Docling for document processing and LangChain for question-answering.

Features

• PDF Upload: Easy file upload through the sidebar
• Document Processing: Uses Docling for advanced PDF processing with OCR and table structure recognition
• Vector Search: FAISS-based vector search for relevant document chunks
• Interactive Chat: Chat interface for asking questions about your book
• Source Documents: View the source chunks used to answer each question
• Document Preview: See the raw Docling markdown output in the right panel
• Persistent Storage: Vector stores are saved and reused for faster subsequent loads

Setup Instructions

1. Install Dependencies

Method A: Using Setup Script (Recommended)

python setup.py

Method B: Manual Installation

# First, install specific versions to avoid conflicts
pip install numpy==1.24.3 pandas==2.1.4 bottleneck==1.3.7

# Then install remaining dependencies
pip install -r requirements.txt

Method C: Alternative Requirements File

pip install -r requirements_alternative.txt

2. Set up Local LLM (Required)

This application uses a local LLM server. You need to set up a local model server at http://localhost:1234/v1.

Option A: Using LM Studio

1. Download and install LM Studio
2. Load your preferred model
3. Start the local server on port 1234

Option B: Using Ollama

1. Install Ollama
2. Pull a model: ollama pull llama2
3. Start the server: ollama serve
4. Update the API endpoint in doclingbookloader.py if needed

3. Run the Application

streamlit run streamlit_app.py

Usage

1. Upload PDF: Use the sidebar to upload a PDF book
2. Processing: Wait for the document to be processed by Docling
3. Ask Questions: Use the chat interface to ask questions about your book
4. View Sources: Expand the "Source Documents" section to see relevant chunks
5. Preview Content: Check the right panel to see the extracted markdown content

File Structure

├── streamlit_app.py          # Main Streamlit application
├── doclingbookloader.py      # Document loader and QA system classes
├── requirements.txt          # Python dependencies
└── README.md                # This file

Configuration

Docling Pipeline Options

The application uses the following Docling configuration:

• OCR: Enabled for text extraction from images
• Table Structure: Enabled with cell matching
• Accelerator: Auto-detection with 8 threads
• Chunk Size: 1000 characters with 200 character overlap

LLM Configuration

Current settings in doclingbookloader.py:

• Model: local-model
• API Base: http://localhost:1234/v1
• Temperature: 0 (deterministic responses)
• Search Type: MMR (Maximum Marginal Relevance)
• Retrieved Chunks: 5 per query

Troubleshooting

Common Issues

1. "AttributeError: ARRAY_API not found" (Bottleneck Error):

   # Solution 1: Use the setup script
   python setup.py
   
   # Solution 2: Manual fix
   pip uninstall bottleneck pandas numpy -y
   pip install numpy==1.24.3 pandas==2.1.4 bottleneck==1.3.7
   pip install -r requirements.txt
   
   # Solution 3: Alternative approach
   pip install -r requirements_alternative.txt

2. "Connection Error": Ensure your local LLM server is running

3. "FAISS Import Error": Install with pip install faiss-cpu

4. "Tokenizer Warnings": These are normal and don't affect functionality

5. "Memory Issues": Try reducing chunk size or using smaller models

Environment Issues

If you're having persistent dependency conflicts:

# Create a fresh conda environment
conda create -n bookqa python=3.10
conda activate bookqa

# Then run the setup
python setup.py

Performance Tips

• First Load: Initial processing takes time but creates a reusable index
• Subsequent Loads: Vector stores are cached for faster startup
• Large Files: Consider splitting very large PDFs for better performance

Customization

Change LLM Provider

To use a different LLM provider, modify the ChatOpenAI initialization in doclingbookloader.py:

# For OpenAI
llm = ChatOpenAI(
    model="gpt-3.5-turbo",
    openai_api_key="your-api-key",
    temperature=0,
)

# For Anthropic
from langchain_anthropic import ChatAnthropic
llm = ChatAnthropic(
    model="claude-3-sonnet-20240229",
    anthropic_api_key="your-api-key",
    temperature=0,
)

Adjust Chunk Settings

Modify the text splitter parameters in doclingbookloader.py:

text_splitter = RecursiveCharacterTextSplitter(
    chunk_size=1500,        # Increase for more context
    chunk_overlap=300,      # Increase for better continuity
    separators=["\n\n", "\n", " ", ""]
)

License

This project is open source and available under the MIT License.