🦆 Duckling

What it does

Duckling extracts textual content and image descriptions from common document formats (PDF, images, CSV/XLSX) and returns them as standardized LangChain Document objects for downstream indexing or retrieval-augmented generation.

How it works

• Detects the input file format and dispatches to a dedicated converter.
• PDF converter uses Docling for parsing and OCR, extracting text and
• Drawing PDFs are handled by a drawing-focused pipeline that prompts
• Images are encoded and described via an LLM prompt; tables are

Technologies

• Docling / docling_core: PDF parsing, OCR and image artifact extraction.
• LangChain / langchain_core: Standard Document model used as output.
• LangGraph: Small state graph to route files by detected format.
• OpenAI-compatible LLM (via langchain_openai.ChatOpenAI): image and drawing description prompts and refinement.
• PyMuPDF (fitz) and OpenCV (cv2): page rendering and image handling.
• Transformers (HuggingFace tokenizer): token-aware chunking for text.

Minimal example

1. Create and activate a virtual environment and install dependencies:

python -m venv .venv
& .\.venv\Scripts\Activate.ps1
poetry install

2. Ensure LLM credentials and any environment settings are available (for example, place keys in a .env file read by the app).

3. Example usage (Python):

from duckling.graph import DucklingGraph

graph = DucklingGraph()
state = graph.run(r"C:\path\to\file.pdf", namespace="my-namespace")
documents = state.get("documents", [])