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

The Core Problem and Solution

Most graph retrieval-augmented generation (GraphRAG) systems extract knowledge in a single pass—entities, relations, and summaries all created once from the source text. This approach is fast but brittle: you get noisy entities that don't consolidate well, duplicate entries that don't get caught, and retrieval that fails because the graph is messy. RAGU fixes this by separating extraction from consolidation into a multi-step pipeline. Entities and relations go through typed extraction twice, then get deduplicated using DBSCAN clustering, summarized by an LLM, and organized into communities using Leiden algorithm. The result: cleaner graphs that retrieve more relevant context.

Why a Smaller Model Works Better

The authors' key insight reshapes how to build these systems. They observe that extraction—understanding text, pulling out entities, reasoning over context—is fundamentally a language skill, not a factual knowledge task. Language understanding improves only weakly as models grow larger; what scales with size is memorized world knowledge (facts, figures, names). For a task like "extract all medical entities from this passage," you don't need a 32-billion-parameter model full of facts. You need something that reads carefully.

This motivated training Meno-Lite-0.1, a 7-billion-parameter model optimized for extraction tasks. Despite being 4.6× smaller than Qwen2.5-32B, it outperforms the larger model by 12.5% on knowledge-graph construction metrics (measured as harmonic mean of precision and recall). It matches the 32B model on downstream GraphRAG tasks—meaning the cleaner extraction pipeline compensates for the smaller model's limited factual knowledge.

What the Benchmarks Show

On GraphRAG-Bench (Medical), RAGU demonstrates clear wins:

• Evidence recall reaches 0.84 across factoid levels, versus ≤0.76 for competing systems. This means when you ask the system for evidence about a medical claim, it retrieves more of the relevant passages.
• On synthesis tasks (asking the system to reason across retrieved facts), RAGU overtakes HippoRAG2, a strong baseline.
• On multi-hop factoid QA (questions requiring chaining multiple facts), HippoRAG2 appeared better—but the authors show this was largely due to answer formatting differences, not actual reasoning quality.

The whole system runs on a single GPU, making it practical for organizations without massive compute budgets. The code is open-source (MIT license) and installable via pip, with the trained model available on Hugging Face.

Why It Matters

This challenges a common assumption in LLM development: that bigger is always better. For structured extraction tasks, a purpose-built smaller model with a smarter pipeline outperforms a generic large one. This has immediate practical implications—faster inference, lower memory cost, and better results. The modular design also means researchers can swap components (different deduplication methods, different community detection algorithms) without rebuilding from scratch. For domains like medicine, law, or finance where structured knowledge matters, RAGU shows a more efficient path forward.