A lightweight Jupyter notebook for exploring the Headnotes and Catchwords of EPO Board of Appeal G-Decisions - running entirely locally, delivering semantic search and RAG answers with transformer models.
Dataset (Zenodo): https://zenodo.org/records/14987955
- Data: G-Decisions, 120 headnotes/catchwords in English (with duplicates)
- Local-first & offline: All processing (embeddings + LLM) runs locally on your device.
- Focused corpus: Optimized for G-decision headnotes & catchwords (EN) to answer domain-specific queries.
- Semantic search + RAG answers: Retrieve the most relevant snippets and generate concise answers with cited sources.
- Notebook-native workflow: Run cell-by-cell, tweak parameters, inspect intermediate results, and visualize embeddings.
- Pluggable local models: Works with Ollama (e.g.,
nomic-embed-text,qwen2.5) or sentence-transformers—swap without code churn. - Persistent vectors: Uses Chroma for fast, on-disk indexing and quick restarts.
Disclaimer: Transformer models can produce incorrect or misleading statements (“hallucinations”). Always verify outputs against the original decisions. This project does not provide legal advice.