HuggingFace Papers Jul 20, 2026 8/10 signal

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

contextresearchmodels

Summary

This paper introduces RAGU, an open-source, multi-step GraphRAG engine, and Meno-Lite-0.1, a compact 7B language model optimized for in-pipeline language skills. RAGU separates knowledge graph construction into distinct extraction and consolidation stages, including typed extraction, DBSCAN-based deduplication, and Leiden community detection, to produce cleaner graphs. The specialized Meno-Lite-0.1 model outperforms a 32B model on knowledge graph construction by a 12.5% relative harmonic mean. On the GraphRAG-Bench (Medical) benchmark, RAGU retrieves the most complete context (evidence recall up to 0.84 vs. ≤0.76) and surpasses the HippoRAG 2 baseline on synthesis tasks.

Problem

Existing GraphRAG systems suffer from three main issues. First, they typically use a single extraction pass to build knowledge graphs, which results in noisy and duplicated entities. Second, they rely on large, expensive LLMs based on the false premise that factual world knowledge is required for in-pipeline tasks, when in fact language skills like comprehension and extraction are sufficient and scale more weakly with model size. Third, many open-source frameworks lack engineering maturity, making them difficult to install, test, and deploy reliably.

Method

The authors developed RAGU, a modular GraphRAG engine with a multi-step pipeline: chunking, two-stage typed entity and relation extraction, DBSCAN-backed consolidation and summarization, and Leiden community detection. They also trained Meno-Lite-0.1, a 7B model fine-tuned for RAG-oriented language skills. The system was evaluated against LightRAG and HippoRAG 2 on the GraphRAG-Bench, BioASQ, MuSiQue, and 2WikiMultiHopQA benchmarks, using gpt-4o-mini as the common answer-generation LLM to isolate the impact of graph construction and retrieval quality.

Details

GraphRAG-Bench (Medical) Results:

RAGU demonstrates a cross-over in performance against HippoRAG 2. While HippoRAG 2 leads on factoid tasks, RAGU excels on synthesis tasks that require broader context, achieving superior coverage and evidence recall across all tasks.

System	Index LLM	Task	Answer Correctness	Coverage	Faithfulness
RAGU	Meno-Lite-0.1	Fact Retrieval	54.2	-	-
HippoRAG 2	Meno-Lite-0.1	Fact Retrieval	72.4	-	-
RAGU	Meno-Lite-0.1	Complex Reasoning	53.7	-	-
HippoRAG 2	Meno-Lite-0.1	Complex Reasoning	68.4	-	-
RAGU	Meno-Lite-0.1	Contextual Summarize	64.1	71.1	-
HippoRAG 2	Meno-Lite-0.1	Contextual Summarize	65.0	51.7	-
RAGU	Meno-Lite-0.1	Creative Generation	59.0	57.4	34.2
HippoRAG 2	Meno-Lite-0.1	Creative Generation	56.9	34.7	26.6

RAGU achieves the highest Evidence Recall at every factoid level (0.84 vs. ≤0.76 for competitors), confirming its consolidation pipeline retrieves more complete context.
LightRAG performed weakest on all tasks, with Answer Correctness scores ranging from 14.2 to 26.2.

Multi-Hop QA Results:

The performance gap between RAGU and HippoRAG 2 on multi-hop QA is shown to be largely an artifact of answer formatting. When prompts are standardized to produce terse answers, RAGU becomes competitive.

Benchmark	System (Terse Prompt)	Answer Correctness	HippoRAG 2 (Terse)
BioASQ	RAGU (Meno-Lite-0.1)	72.8	72.4
2WikiMultiHopQA	RAGU (Meno-Lite-0.1)	55.1	63.5
MuSiQue	RAGU (Meno-Lite-0.1)	40.7	54.4

Under verbose prompts, HippoRAG 2's AC on BioASQ was 74.1 vs. 54.5 for RAGU. Under terse prompts, RAGU's AC rose to 72.8, matching HippoRAG 2's 72.4.
HippoRAG 2 maintains a lead on MuSiQue, the most difficult multi-hop benchmark.

Meno-Lite-0.1 Model Evaluation:

The compact 7B model, Meno-Lite-0.1, outperforms significantly larger models on knowledge graph construction tasks, validating the hypothesis that language skills are more critical than world knowledge for this role.

Model	Size	NER (F1)	RE (F1)	Harmonic Mean
Meno-Lite-0.1	7B	0.504	0.347	0.468
Qwen2.5-32B	32B	0.536	0.239	0.416
gemma-3-27b	27B	0.544	0.224	0.396
Qwen2.5-7B	7B	0.477	0.192	0.356

Meno-Lite-0.1 outperforms Qwen2.5-32B by 12.5% relative on the harmonic mean score, driven by a large advantage in relation extraction (0.347 vs. 0.239 F1).
On end-to-end QA tasks, the performance difference between using Meno-Lite-0.1 and larger models for extraction is minimal (≤1.5 pp), indicating the RAGU pipeline is robust to the choice of extractor.

What's new

The paper introduces a multi-step graph consolidation pipeline for GraphRAG that explicitly separates extraction from deduplication, summarization, and community detection. It also proposes and validates the hypothesis that compact, skill-oriented LLMs are sufficient and cost-effective for in-pipeline RAG tasks, demonstrating this with Meno-Lite-0.1, a 7B model that outperforms a 32B model on knowledge graph construction. The work also emphasizes engineering maturity, delivering a production-ready, installable, and testable open-source package.

Limitations

The authors state four limitations. First, the evidence for their model scaling hypothesis is based on a single model family (Qwen2.5). Second, Meno-Lite-0.1 is optimized for contextual grounding at the expense of parametric factual recall and should not be used as a standalone knowledge base. Third, there is a potential for residual advantage on their IE benchmark due to distributional overlap in the training data's schema and domain. Finally, the default graph backend (NetworkX) does not scale to massive corpora without being replaced by a dedicated graph database adapter.

Conclusion

The authors conclude that LLMs inside a RAG pipeline primarily require language skills, not world knowledge, allowing smaller, more efficient models to be effective. Their system, RAGU, operationalizes this by using a multi-step consolidation pipeline that excels at synthesis and broad-context tasks, retrieving more complete evidence than competitors. They suggest that for practical applications, RAGU is preferable for tasks like summarization and creative generation on a single-GPU budget, while chain-traversal systems like HippoRAG 2 are better suited for precise, multi-hop factoid queries.

Read the source →

Don't read this site daily. Get it in your inbox.

The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.