RAGU: A 7B Domain-Adapted Model Beats Qwen2.5-32B by 12.5% on Knowledge-Graph Construction

arXiv / HuggingFace Daily Papers (72 upvotes)·medium signal

RAGU (arXiv 2607.11683, July 13, Novosibirsk State University + collaborators) splits GraphRAG into separate entity-extraction and consolidation stages, using DBSCAN-backed deduplication, LLM summarization and Leiden community detection instead of a single pass. Their 7B Meno-Lite-0.1 exceeds Qwen2.5-32B by +12.5% relative harmonic mean on KG construction while matching it on English GraphRAG, and hits 0.84 evidence recall on GraphRAG-Bench Medical versus HippoRAG2's ≤0.76. It runs on one GPU and installs via pip — a rare directly-usable GraphRAG result.