A Streamlit app for asking questions about a PDF book (Docling for parsing, LangChain and FAISS for retrieval, a local LLM server for answers), plus a set of small Docling example scripts and notebooks.
| File | Purpose |
|---|---|
streamlit_app.py |
Streamlit UI: upload a PDF, chat, view source chunks and the Docling markdown |
docling_book_loader.py |
DoclingBookLoader and BookQASystem classes (parsing, chunking, FAISS index, QA chain) |
setup_script.py |
Installs the dependencies into the active environment |
AI-Engineer-Skool_booktutor.py |
A separate book-tutor script |
docling_examples_*.py |
Small examples: simple and custom conversion, tables and table export, scanned PDFs, LangChain, LlamaIndex RAG |
docling.ipynb, mayurji_docling-*.ipynb, sudarshan_koirala_docling.ipynb |
Tutorial notebooks |
docling-demo-project-bextuychiev/ |
A third-party demo project with its own README.md and requirements.txt |
data/ |
Sample PDFs (including scanned.pdf and table.pdf) |
scratch/ |
Exported table CSV and HTML output from the examples |
Create the conda environment from environment.yml and start the app:
conda env create -f environment.yml
conda activate docling
streamlit run streamlit_app.pyThe app expects a local OpenAI-compatible LLM server at http://localhost:1234/v1 (LM Studio, or any
server on that port; see Code/LOCAL_LLM_PLAN.md). Start it before asking questions.
Set in docling_book_loader.py:
- Docling pipeline: OCR on, table structure on, automatic accelerator with 8 threads
- Chunking: 1000 characters with 200 overlap (
RecursiveCharacterTextSplitter) - Retrieval: FAISS, MMR search, 5 chunks per query; the vector store is saved and reused on later loads
- LLM:
ChatOpenAIwith modellocal-model, base URLhttp://localhost:1234/v1, temperature 0
To use another provider, change the ChatOpenAI initialisation in BookQASystem.
- If the Bottleneck "ARRAY_API not found" error appears, rerun
python setup_script.pyin a fresh conda environment. - A connection error means the local LLM server is not running.
- The first load of a book is slow; later loads reuse the saved index.