RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM
Mikhail Komarov, Ivan Bondarenko, Stanislav Shtuka, Oleg Sedukhin, Roman Shuvalov, Yana Dementyeva, Matvey Solovyov, Nikolay O. Nikitin
cs.CL, cs.AI
2026-07-13
Existing GraphRAG does a single noisy extraction pass; RAGU splits construction into multiple steps (two-stage entity and relation extraction, clustering-based consolidation) and pairs it with a 7B model trained to use context rather than recall, cutting indexing cost from about $0.10 per document on GPT-4o to about $0.001 locally.
Graph retrieval-augmented generation (GraphRAG) gives LLMs structured knowledge, but existing systems build the graph in a single extraction pass, producing noisy, duplicated entities and brittle retrieval. Practitioners default to GPT-4-class API models on a false premise: what an LLM needs inside a RAG pipeline (comprehension, extraction, in-context reasoning) is language skill, not factual recall, and factual recall scales near-linearly with parameters while language skill scales much more slowly. On the Qwen2.5 family, CheGeKa F1 rises 21.1x from 0.5B to 72B (recall-driven) while MultiQ rises only 4x (skill-driven). Open-source frameworks also suffer install failures and unsafe paths like eval() on raw LLM output.
Six-step construction: chunking (three chunkers); two-stage extraction (entities first, validated against the NEREL schema of 29 entity types and 49 relation types, then relations whose endpoints must be validated entities); consolidation (EntitySummarizer groups by name plus type, DBSCAN clusters, LLM summarizes duplicates), done before community detection, the denoising step single-pass systems like LightRAG lack; then Leiden community clustering, community summarization, and refinement.
The compact model Meno-Lite-0.1 is a 7B model from RuadaptQwen2.5-7B-Lite-Beta, with continued pretraining on 1.3B tokens (Russian and English educational and scientific text) and SFT on 50M tokens (NEREL extraction, multi-hop QA, query logs). SFT teaches the model to use context rather than recall facts, investing compute in language skill. It has a 128K context window (passkey retrieval 0.98 at 128K) and 47% better tokenizer efficiency on Russian.
The engine offers a three-tier storage abstraction (graph, KV, vector) with lifecycle callbacks, an async-first API, structured outputs validated through Pydantic v2, and five search engines (Local, Global, Naive, Mix, QueryPlan).
On GraphRAG-Bench (medical), indexing and answering both with Meno-Lite-0.1: LightRAG scores 26.2/20.2/22.6/14.4 (fact retrieval, complex reasoning, contextual summary, creative generation) with 3.9 coverage; HippoRAG 2 scores 72.4/68.4/65.0/56.9 with 34.7 coverage; RAGU scores 54.2/53.7/64.1/59.0 with 57.4 coverage. HippoRAG 2 wins fact retrieval and complex reasoning, but RAGU's coverage is far higher, and on broad-synthesis creative generation RAGU flips ahead (59.0 vs 56.9). RAGU has the highest evidence recall at every factoid level (84 vs at most 76 for rivals).
On information extraction, the 7B Meno-Lite-0.1 reaches harmonic mean 0.468, beating Qwen2.5-32B (0.416) by 12.5%. Cost per document: MS-GraphRAG on GPT-4o about 40k tokens and about $0.10; RAGU with Meno-Lite-0.1 locally about 8k tokens at a fixed GPU cost of about $0.001. For 100k documents that is about $10,000 versus about $100, two orders of magnitude apart.
This turns GraphRAG from "index with the most expensive large model" into "index locally with a 7B model," cutting cost by two orders of magnitude. For teams scaling up, multi-step denoising plus a small model trained only on language skill beats swapping in a bigger model. The engine is open-source and modular with swappable storage backends.
The core claim that language skill does not scale with size is validated on one model family (Qwen2.5) and selected tasks; the paper calls it a supported hypothesis, not a universal theorem. Meno-Lite-0.1 trades parametric recall for contextual grounding and should not be a standalone knowledge base; multi-hop reasoning degrades past 32K. SFT uses NEREL train and validation splits while the benchmark uses a held-out test split of the same source, so a residual overlap advantage cannot be ruled out. The default NetworkX backend does not scale to million-node corpora. Final graph quality is still sensitive to the extraction LLM; weak base models inject structural noise that consolidation cannot fix.