Open Energy Family – municipal-heat-planning-pdf-processing

Tools and scripts developed to support the MHPO development by automating data extraction and processing of municipal heat planning documents are stored here.

Key Dependencies and Models

Inference Runtime

Tool	Purpose
Ollama	Local model serving and inference runtime for all LLM and vision-language model interactions

Models

Model	Provider	Parameters	Usage in Pipeline
PP-DocLayoutV3	PaddlePaddle	—	Document layout detection (Stage 2): identifies tables, figures, titles, headers, footers, and other structural elements in PDF pages
gpt-oss:120b	OpenAI (open-weight)	120B	LLM-based section refinement (Stage 4): cleans extraction artefacts, normalizes titles and captions, removes directory pages, converts bibliographies to BibTeX
Qwen3-VL	Alibaba / Qwen	32B (Q8)	Vision-language model for image processing: converts table images to structured Markdown, generates detailed textual descriptions of figures, and produces captions where missing

Python Libraries

Library	Purpose
PyMuPDF (fitz)	PDF text extraction (rawdict mode) and page rendering to PNG
Pillow	Image handling, cropping detected layout regions
NumPy	Array operations for image crop processing
Transformers	Loading and running the PP-DocLayoutV3 layout detection model
ollama (Python)	Python client for Ollama API (chat, model management)
httpx	HTTP client with timeout control, used internally by the Ollama Python library
spaCy	NLP processing for text analysis and entity recognition

Collaboration

Everyone is invited to develop this repository with good intentions.