RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM

RAGU is an open-source, modular GraphRAG engine that addresses three obstacles in existing systems: single-pass extraction, dependence on expensive LLMs, and engineering immaturity.

It separates extraction from consolidation via a six-stage pipeline: chunking, two-stage typed extraction (entities then relations) under the NEREL schema, DBSCAN-backed deduplication and LLM summarization, Leiden community detection, community summarization, and refinement.

The authors propose the Language/World Knowledge Hypothesis, stating that language skills scale weakly with model size, unlike world knowledge.

This motivates Meno-Lite-0.1, a 7B model fine-tuned for RAG-oriented language skills, which outperforms Qwen2.5-32B on knowledge-graph construction (+12.5% harmonic mean) and matches it on English GraphRAG tasks.

On GraphRAG-Bench (Medical), RAGU achieves the highest evidence recall at every factoid level (up to 0.84 vs.

≤0.76) and overtakes HippoRAG 2 on synthesis tasks (Creative Generation AC 59.0 vs.

56.9).

On multi-hop factoid QA, the apparent HippoRAG 2 advantage is largely an answer-format artifact; with terse prompts, RAGU ties on BioASQ (AC 72.9 vs.

72.4) and closes the 2WikiMultiHopQA gap.

RAGU is installable via pip install graph_ragu, runs on a single GPU, and is released under MIT license; Meno-Lite-0.1 is under Apache 2.0.

The system includes five search engines, a three-tier storage abstraction, Pydantic validation, and ~374 tests with a mock LLM server.

Cost analysis shows ~$0.001/doc vs.

~$0.10/doc for commercial APIs, making it ~100x cheaper at scale.

Limitations include reliance on a single model family for scaling evidence, degradation beyond 32K tokens for multi-hop reasoning, and sensitivity to extraction LLM quality.

The authors recommend RAGU for synthesis…