Most GraphRAG systems extract entities and relations in a single pass, which yields noisy entities and brittle retrieval. The paper's core insight is that the in-pipeline LLM mainly needs language skills (comprehension, extraction, reasoning over context) that scale weakly with model size, so a compact, domain‑adapted 7B extractor can match or beat much larger models for knowledge-graph construction.
Key Findings
- Pipeline separation: RAGU splits extraction and consolidation into typed two-stage extraction followed by DBSCAN-based deduplication, LLM summarization, and Leiden community detection — this reduces noisy, duplicated entities and improves retrieval precision.
- Compact extractor works: Meno‑Lite‑0.1 (7B, optimized for language skills) outperforms Qwen2.5-32B on knowledge-graph construction by +12.5% relative harmonic mean, and matches larger models on English GraphRAG tasks.
- Better context recall and synthesis: On GraphRAG-Bench (Medical), RAGU retrieves more complete contexts (evidence recall up to 0.84 vs ≤0.76) and surpasses HippoRAG2 on synthesis tasks; apparent multi-hop QA advantages for rivals were largely due to answer-format artifacts.
- Practical constraints: RAGU is modular, runs on a single GPU, installable via pip, and released under an MIT license, making it accessible for applied workflows without massive infra.
Who it's for and trade-offs
Great fit if you need higher-quality retrieval context for downstream LLM generation and want a reproducible, modular pipeline that runs on modest hardware. It favors domains where extraction and consolidation logic can be engineered (typed schemas, dedup rules) and where a compact, task-tuned LLM suffices. Look elsewhere if you require end-to-end factual knowledge beyond the provided context (global world knowledge not encoded in the pipeline) or if your use case depends on models pretrained with specific factual corpora rather than language-skill optimization.
Where it fits
RAGU sits between lightweight RAG setups (simple retrieval + prompt) and heavyweight, closed-source multibillion models: it prioritizes structured context quality and pipeline cleanliness over raw model scale, making it appealing for domain-specific GraphRAG applications and medical/technical benchmarks.