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

Abstract: Graph retrieval-augmented generation (GraphRAG) enhances LLMs with structured knowledge, yet existing systems construct knowledge graphs in a single extraction pass, producing noisy entities and brittle retrieval. RAGU, an open-source modular GraphRAG engine, addresses this by separating extraction from consolidation: entities and relations pass through two-stage typed extraction, DBSCAN-backed deduplication, LLM summarization, and Leiden community detection. A key insight motivates a compact extractor: the skills an in-pipeline LLM needs - comprehension, extraction, reasoning over context - are language skills that grow only weakly with model size, unlike factual world knowledge. Accordingly, we train Meno-Lite-0.1, a 7B model optimized for language skills, which outperforms Qwen2.5-32B on knowledge-graph construction (+12.5% relative harmonic mean) and matches it on English GraphRAG tasks. On GraphRAG-Bench (Medical), RAGU retrieves the most complete context at every factoid level (evidence recall up to 0.84 vs. $\leq$0.76) and overtakes HippoRAG2 on synthesis tasks; on multi-hop factoid QA, the apparent HippoRAG2 advantage is shown to be largely an answer-format artifact. RAGU is installable via $\texttt{pip install graph_ragu}$, runs on a single GPU, and is released under MIT. The source code is publicly available at https://github.com/RaguTeam/RAGU, and the Meno-Lite-0.1 model can be obtained from https://huggingface.co/bond005/meno-lite-0.1.

• The paper demonstrates a modular multi-step GraphRAG engine that decouples language skills from world knowledge scaling using a compact 7B model.
• It employs configurable chunking, two-stage extraction, deduplication, and community detection for robust, high-quality knowledge graph construction.
• Empirical evaluations show that RAGU outperforms larger models in context-rich tasks while reducing cost and infrastructure overhead.

RAGU: Modular Multi-Step GraphRAG with Compact Domain-Adapted LLM

Motivation and Hypotheses

RAGU is motivated by the recognition that contemporary GraphRAG systems suffer from excessive reliance on monolithic, single-pass information extraction and expensive large-scale LLMs for knowledge graph construction. The authors identify three core limitations: (1) single-pass entity/relation extraction yielding noisy and fragmented graphs, (2) dependency on large parametric models under the mistaken assumption that external world knowledge is essential, and (3) lack of engineering rigor and modularity in open-source implementations. To address these challenges, they introduce two artifacts: RAGU, a modular multi-step GraphRAG engine, and Meno-Lite-0.1, a 7B-parameter model domain-adapted for RAG tasks focused on language skills rather than factual recall.

The underlying hypothesis, validated empirically, is that language skills necessary for extraction, reasoning, and context summarization in RAG pipelines scale far more weakly with model size than world knowledge. This supports the replacement of large-scale LLMs with smaller models if the pipeline is architected for extraction-consolidation, not fact recall.

Figure 1: F1 scaling of world-knowledge (CheGeKa) vs. language-skill (MultiQ) in Qwen2.5-Instruct models reveals world-knowledge scales steeply (21.1×), while language skills increase only modestly (4×).

System Architecture

RAGU implements a fully modular, end-to-end pipeline that dissects the problem into explicit stages:

1. Chunking: Offers multiple configurable chunking strategies including semantic (embedding-based) and hybrid approaches.
2. Typed, Two-Stage Extraction: Extracts entities (NEREL schema) independently, validates types, then extracts relations constrained against this validated entity set, eliminating entity–relation mismatches.
3. Deduplication and Summarization: DBSCAN-based clustering and LLM-based summarization aggregate redundant entities and relations before downstream processing. This is a critical differentiator from single-pass pipelines.
4. Community Detection: Leiden clustering identifies subgraphs, followed by LLM-generated community summaries and optional refinement steps.

Figure 2: End-to-end indexing architecture in RAGU, delineating all artifact flows and processing stages with persistent storage at each swappable tier.

The search stack integrates five engines (ranging from vanilla vector RAG to hybrid dense-sparse retrieval and LLM-augmented global search), all exposed through an async-first, production-grade API, with a rigorous storage abstraction for graph, vector, and key-value backends.

Compact Domain-Adapted Model: Meno-Lite-0.1

Meno-Lite-0.1 is a compact, 7B-parameter LLM, fine-tuned for information extraction and downstream RAG-specific tasks based on contextual grounding rather than parametric knowledge. It demonstrates:

• 128K token context window with nearly perfect passkey retrieval (0.98 at 128K).
• 47% improved tokenizer efficiency relative to Qwen2.5 on Russian text.
• Extraction performance matching or exceeding models 4× its size in structured tasks.

The model’s training focus eschews literal world knowledge acquisition and instead emphasizes robust instruction following, context-aware extraction, multi-hop reasoning, and faithful summarization, yielding superior schema-constrained extraction while operating on a single consumer GPU.

Empirical Evaluation and Results

Evaluation is performed on multiple knowledge-intensive and multi-hop QA datasets, including the domain-intensive GraphRAG-Bench (Medical), BioASQ, MuSiQue, and 2WikiMultiHopQA. All systems use identical answer generation LLMs, isolating the impact of index LLMs and pipeline design.

On GraphRAG-Bench:

• Factoid QA: HippoRAG 2 maintains an edge in single-hop precision tasks (AC = 72.4 vs. 54.2 for RAGU), attributed to its personalized PageRank traversal for isolated fact lookup.
• Context-Breadth / Synthesis Tasks: RAGU overtakes HippoRAG 2 on Contextual Summarize and Creative Generation (Creative Generation AC = 59.0 vs. 56.9 for HippoRAG 2), and leads substantially in evidence Recall (up to 0.84 vs. ≤0.76).
• Coverage: RAGU consistently outperforms in retrieving all relevant supporting context.

Figure 3: Task complexity crossover on GraphRAG-Bench—RAGU surpasses HippoRAG 2 as synthesis/contextual complexity grows, supported by superior evidence recall.

In multi-hop QA:

• Under verbose answer protocols, HippoRAG 2 dominates (format bias).
• When controlling for output format, RAGU achieves parity or closes the gap on BioASQ and 2WikiMultiHopQA, with HippoRAG 2 retaining only a modest advantage on the hardest multi-hop chain-following benchmark (MuSiQue).

The model ablation demonstrates that increasing LLM parameter count from 3B to 14B has negligible impact (ΔAC ≤ 1.5pp), confirming the architectural hypothesis: robust graph construction is much more sensitive to the extraction-consolidation pipeline design than to model scale, given adequate language modeling.

Extraction and Reasoning Case Study

A case study exemplifies the multi-step pipeline, producing: (1) highly structured, type-validated entity/relation graphs from a short passage; (2) clean, clustered communities with interpretable summaries; and (3) accurate multi-hop retrieval over extracted knowledge.

Figure 4: Extracted knowledge graph on the “Ritchie passage” partitions semantic communities, exemplifying pipeline output in microcosm.

Engineering Maturity and Practical Considerations

RAGU distinguishes itself from contemporaries like HippoRAG 2 and LightRAG by virtue of robust engineering:

• Full async APIs, swappable storage tiers, and deterministic identifiers.
• Elimination of code execution from LLM output (no eval()), Pydantic-validated outputs, and ∼374 automated tests plus a deterministic mock LLM server.
• Deployment on a single GPU with cost at least two orders of magnitude below API-centric systems, substantially reducing infrastructure and environmental overhead.

Limitations include dataset/domain dependence (NEREL schema, Russian/English focus), pipeline scaling on massive graphs (current dependency on NetworkX), and domain adaptation requirements for other schema typologies.

Implications and Future Directions

The RAGU work makes a strong claim that, when extraction and consolidation are properly modularized, model size beyond moderate scales (7B) contributes minimal incremental gains for language-skill-focused RAG indices. This challenges prevailing “bigger is better” trends in RAG extraction, particularly in the cost-sensitive, open-source context.

For practitioners, the findings imply favoring smaller, optimized models and explicit consolidation modules for knowledge-intensive regulatory or creative tasks, while recognizing the ongoing advantage of chain traversal algorithms for precision factoid queries. Future work will need to explore generalization across more diverse languages, scaling consolidation mechanisms, and further systematization of schema adaptation for out-of-domain settings.

Conclusion

RAGU formalizes and substantiates the decoupling of language-skill scaling from world-knowledge scaling in knowledge graph construction for RAG pipelines. The modular extraction-consolidation design, combined with a compact, extraction-specialized LLM, achieves strong empirical results in context-rich, creative, and broad-coverage QA, matching or surpassing much larger models and systems in this regime. Engineering robustness and open-source accessibility are further distinguishing characteristics. The results challenge the necessity of relying on LLM-scale world knowledge for high-quality RAG and provide a clear blueprint for cost-effective, production-quality, domain-adapted GraphRAG deployment.