Crimson AI NewsA CrimsonLingua Network service
EN ع
← Back to news
Research paper Hugging Face

RAGU: Open-Source GraphRAG Engine Uses Compact 7B Model to Outperform Larger LLMs

AI By Crimson AI Hugging Face Papers 20 July 2026 · 00:00 10 views
Share: X Telegram

Researchers introduce RAGU, a modular GraphRAG engine that separates extraction from consolidation, and Meno-Lite-0.1, a 7B model that outperforms Qwen2.5-32B on knowledge graph construction while reducing costs by up to 100x.

RAGU: Open-Source GraphRAG Engine Uses Compact 7B Model to Outperform Larger LLMs

Key points

A team from Novosibirsk State University and ITMO University has released RAGU, an open-source GraphRAG engine that challenges the assumption that large language models are necessary for high-quality knowledge graph construction. The system introduces a two-stage pipeline that separates entity extraction from consolidation, reducing noise and improving retrieval.

Central to RAGU is Meno-Lite-0.1, a 7B parameter model fine-tuned for language skills such as comprehension, extraction, and reasoning. The researchers found that these skills scale only weakly with model size (4x growth from 0.5B to 72B parameters), unlike world knowledge (21x growth). Meno-Lite-0.1 outperforms Qwen2.5-32B on knowledge graph construction by 12.5% in harmonic mean and matches it on end-to-end GraphRAG tasks.

RAGU's pipeline includes two-stage typed extraction, DBSCAN-based deduplication, LLM summarization, and Leiden community detection. On the GraphRAG-Bench (Medical) benchmark, RAGU achieves evidence recall up to 0.84, compared to ≤0.76 for competitors, and surpasses HippoRAG2 on synthesis tasks (Creative Generation AC: 59.0 vs. 56.9). The apparent advantage of HippoRAG2 on multi-hop QA is largely attributed to answer-format artifacts; after controlling for format, RAGU ties HippoRAG2 on BioASQ (72.9 vs. 72.4).

RAGU is designed for cost efficiency: it runs on a single consumer GPU and costs approximately $0.001 per document on rented GPUs, compared to ~$0.10 per document for API-based alternatives. At 100,000 documents, this translates to $100 vs. $10,000. The engine supports three storage backends (NetworkX, Neo4j, NanoVDB, Qdrant) and includes over 374 tests with a deterministic mock LLM server for CI.

The researchers acknowledge limitations, particularly in handling lexically distant synonyms and acronyms (e.g., "USA" vs. "United States"). Current retrieval-side patches use dense retrieval and community co-occurrence, but a dedicated entity-linking module is planned for future releases. RAGU is available under MIT license, and Meno-Lite-0.1 under Apache-2.0.

MetricRAGU (7B)HippoRAG2 (20B)Qwen2.5-32B
KG Construction (Harmonic Mean)+12.5% vs QwenBaseline
Evidence Recall (Medical)0.84≤0.76
Creative Generation AC59.056.9
BioASQ (format-controlled)72.972.4
Cost per document~$0.001~$0.10
Source
Hugging Face · Hugging Face Papers
Related news
Research paper
Hugging Face 31 Aug 2026

Hugging Face Unveils StepGuard: Step-Level Guardrails for Safer AI Agents

StepGuard, a new step-level guard model from Hugging Face, audits agent actions before execution, reducing attack success rates by...

1
Research paper
Hugging Face 31 Aug 2026

Hugging Face Researchers Unveil ABot-Recon for Stable Long-Horizon 3D Reconstruction

ABot-Recon, a new streaming 3D reconstruction model from Hugging Face, achieves stable long-horizon performance using only local t...

1
Research paper
Hugging Face 31 Aug 2026

ContextPilot: Teaching Agents Proactive Context Management via Fine-Grained RL

Hugging Face researchers introduce ContextPilot, a framework that enhances long-horizon agent reasoning by expanding context-editi...

1