\n\n**GraphRAG doesn't have to be expensive, brittle, or messy.**\n\nWe're the team behind **RAGU** — an open-source GraphRAG engine built jointly by the Laboratory of Applied Digital Technologies at Novosibirsk State University and the Laboratory of Automated Multi-Agent AI at ITMO University. Its compact extractor, **Meno-Lite-0.1**, also powers our 1st-place system in SemEval-2026 Task 8 (MTRAG) ([paper](https://huggingface.co/papers/2605.04523)), so this isn't a toy model. And we're here to challenge a common assumption: *you don't need a massive LLM to build high-quality knowledge graphs.*\n\n**What we tested and what we built:**\n\n**The \"Language vs. World Knowledge\" hypothesis.** We ran experiments across 6 model sizes (0.5B to 72B). Language skills — comprehension, extraction, reasoning over context — scale slowly with model size (4× growth), while world knowledge scales steeply (21×). Yet existing GraphRAG systems throw 32B–72B models at extraction, wasting compute on facts they don't need to recall.\n\n**Meno-Lite-0.1** is a 7B model fine-tuned specifically for RAG-oriented language tasks. It outperforms Qwen2.5-32B on knowledge-graph construction (+12.5% harmonic mean). On end-to-end GraphRAG QA, its standalone extraction edge shrinks to ≤1pp — not because the fine-tuning failed, but because a good consolidation pipeline reduces sensitivity of QA quality to the extractor. Net result: 32B-class extraction at 7B cost.\n\n**RAGU itself** is a modular GraphRAG engine with a 6-stage pipeline that separates extraction from consolidation: two-stage typed extraction, DBSCAN deduplication, LLM summarization, Leiden community detection. Cleaner graphs, better retrieval.\n\n**Benchmarks:**\n\n- On GraphRAG-Bench (Medical): RAGU retrieves the most complete context at every factoid level — evidence recall up to 84% vs. ≤76% for competitors.\n- Outperforms HippoRAG 2 on synthesis tasks (Creative Generation AC: 59.0 vs. 56.9).\n- Multi-hop QA gap? Turns out it's mostly an answer-format artifact. Controlling for format, RAGU ties HippoRAG 2 on BioASQ (72.9 vs. 72.4).\n- RAGU does this with a locally served 7B model against HippoRAG 2's 20B extractor.\n- Cost: ~$0.001/doc on rented GPUs vs. ~$0.10/doc for API-based alternatives. At 100k docs, that's $100 vs. $10,000.\n\n**Practical details:**\n\n- `pip install graph_ragu` — runs on a single consumer GPU.\n- ~374 tests + deterministic mock LLM server, so CI works without API keys.\n- Three swappable storage backends (NetworkX or Neo4j, NanoVDB or Qdrant).\n- MIT-licensed code, Apache-2.0 model weights.\n\nWe built RAGU because open-source GraphRAG shouldn't mean choosing between academic prototypes and expensive APIs. Read the paper, try the code, let us know what breaks.\n\nCode: https://github.com/RaguTeam/RAGU\nModel: https://huggingface.co/bond005/meno-lite-0.1\nDemo videos:\n- [2 min](https://youtu.be/bicJDMJuQfg): basic RAGU setup and first run\n- [11 min](https://youtu.be/F6dFm9zyNck): exploring alternative pipeline configurations\n\n---\n\n*P.S. Appendix A documents the engineering gaps we found in HippoRAG 2 at a pinned commit: raw `eval()` on LLM outputs, missing retries, no storage abstraction. We show side-by-side how RAGU handles each of these — Pydantic-validated outputs, proper retry logic, swappable backends.*","html":"<img src=\"https://cdn-uploads.huggingface.co/production/uploads/62b1e0f76a5435fd9a60a8dc/8JdjDFgxqdytPELkYcu5I.png\" width=\"100%\" style=\"max-width:100%; height:auto;\">\n\n<p><strong>GraphRAG doesn't have to be expensive, brittle, or messy.</strong></p>\n<p>We're the team behind <strong>RAGU</strong> — an open-source GraphRAG engine built jointly by the Laboratory of Applied Digital Technologies at Novosibirsk State University and the Laboratory of Automated Multi-Agent AI at ITMO University. Its compact extractor, <strong>Meno-Lite-0.1</strong>, also powers our 1st-place system in SemEval-2026 Task 8 (MTRAG) (<a href=\"https://huggingface.co/papers/2605.04523\">paper</a>), so this isn't a toy model. And we're here to challenge a common assumption: <em>you don't need a massive LLM to build high-quality knowledge graphs.</em></p>\n<p><strong>What we tested and what we built:</strong></p>\n<p><strong>The \"Language vs. World Knowledge\" hypothesis.</strong> We ran experiments across 6 model sizes (0.5B to 72B). Language skills — comprehension, extraction, reasoning over context — scale slowly with model size (4× growth), while world knowledge scales steeply (21×). Yet existing GraphRAG systems throw 32B–72B models at extraction, wasting compute on facts they don't need to recall.</p>\n<p><strong>Meno-Lite-0.1</strong> is a 7B model fine-tuned specifically for RAG-oriented language tasks. It outperforms Qwen2.5-32B on knowledge-graph construction (+12.5% harmonic mean). On end-to-end GraphRAG QA, its standalone extraction edge shrinks to ≤1pp — not because the fine-tuning failed, but because a good consolidation pipeline reduces sensitivity of QA quality to the extractor. Net result: 32B-class extraction at 7B cost.</p>\n<p><strong>RAGU itself</strong> is a modular GraphRAG engine with a 6-stage pipeline that separates extraction from consolidation: two-stage typed extraction, DBSCAN deduplication, LLM summarization, Leiden community detection. Cleaner graphs, better retrieval.</p>\n<p><strong>Benchmarks:</strong></p>\n<ul>\n<li>On GraphRAG-Bench (Medical): RAGU retrieves the most complete context at every factoid level — evidence recall up to 84% vs. ≤76% for competitors.</li>\n<li>Outperforms HippoRAG 2 on synthesis tasks (Creative Generation AC: 59.0 vs. 56.9).</li>\n<li>Multi-hop QA gap? Turns out it's mostly an answer-format artifact. Controlling for format, RAGU ties HippoRAG 2 on BioASQ (72.9 vs. 72.4).</li>\n<li>RAGU does this with a locally served 7B model against HippoRAG 2's 20B extractor.</li>\n<li>Cost: ~$0.001/doc on rented GPUs vs. ~$0.10/doc for API-based alternatives. At 100k docs, that's $100 vs. $10,000.</li>\n</ul>\n<p><strong>Practical details:</strong></p>\n<ul>\n<li><code>pip install graph_ragu</code> — runs on a single consumer GPU.</li>\n<li>~374 tests + deterministic mock LLM server, so CI works without API keys.</li>\n<li>Three swappable storage backends (NetworkX or Neo4j, NanoVDB or Qdrant).</li>\n<li>MIT-licensed code, Apache-2.0 model weights.</li>\n</ul>\n<p>We built RAGU because open-source GraphRAG shouldn't mean choosing between academic prototypes and expensive APIs. Read the paper, try the code, let us know what breaks.</p>\n<p>Code: <a href=\"https://github.com/RaguTeam/RAGU\" rel=\"nofollow\">https://github.com/RaguTeam/RAGU</a><br>Model: <a href=\"https://huggingface.co/bond005/meno-lite-0.1\">https://huggingface.co/bond005/meno-lite-0.1</a><br>Demo videos:</p>\n<ul>\n<li><a href=\"https://youtu.be/bicJDMJuQfg\" rel=\"nofollow\">2 min</a>: basic RAGU setup and first run</li>\n<li><a href=\"https://youtu.be/F6dFm9zyNck\" rel=\"nofollow\">11 min</a>: exploring alternative pipeline configurations</li>\n</ul>\n<hr>\n<p><em>P.S. Appendix A documents the engineering gaps we found in HippoRAG 2 at a pinned commit: raw <code>eval()</code> on LLM outputs, missing retries, no storage abstraction. We show side-by-side how RAGU handles each of these — Pydantic-validated outputs, proper retry logic, swappable backends.</em></p>\n","updatedAt":"2026-07-20T05:13:33.415Z","author":{"_id":"62b1e0f76a5435fd9a60a8dc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655824626110-noauth.png","fullname":"Ivan Bondarenko","name":"bond005","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":85,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8276038765907288},"editors":["bond005"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1655824626110-noauth.png"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.11683","authors":[{"_id":"6a55e685a9d74d6e65bbd8f3","name":"Mikhail Komarov","hidden":false},{"_id":"6a55e685a9d74d6e65bbd8f4","name":"Ivan Bondarenko","hidden":false},{"_id":"6a55e685a9d74d6e65bbd8f5","name":"Stanislav Shtuka","hidden":false},{"_id":"6a55e685a9d74d6e65bbd8f6","name":"Oleg Sedukhin","hidden":false},{"_id":"6a55e685a9d74d6e65bbd8f7","name":"Roman Shuvalov","hidden":false},{"_id":"6a55e685a9d74d6e65bbd8f8","name":"Yana Dementyeva","hidden":false},{"_id":"6a55e685a9d74d6e65bbd8f9","name":"Matvey Solovyov","hidden":false},{"_id":"6a55e685a9d74d6e65bbd8fa","name":"Nikolay O. Nikitin","hidden":false}],"publishedAt":"2026-07-13T00:00:00.000Z","submittedOnDailyAt":"2026-07-20T00:00:00.000Z","title":"RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM","submittedOnDailyBy":{"_id":"62b1e0f76a5435fd9a60a8dc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655824626110-noauth.png","isPro":true,"fullname":"Ivan Bondarenko","user":"bond005","type":"user","name":"bond005"},"summary":"Graph retrieval-augmented generation (GraphRAG) enhances large language models 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. leq0.76) 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 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.","upvotes":36,"discussionId":"6a55e686a9d74d6e65bbd8fb","githubRepo":"https://github.com/RaguTeam/RAGU","githubRepoAddedBy":"user","githubStars":48,"organization":{"_id":"62b1e262f4a72794188b6757","name":"NSU","fullname":"Novosibirsk State University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655825086683-62b1e0f76a5435fd9a60a8dc.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"62b1e0f76a5435fd9a60a8dc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655824626110-noauth.png","isPro":true,"fullname":"Ivan Bondarenko","user":"bond005","type":"user"},{"_id":"6667c253ec712f50d63753f5","avatarUrl":"/avatars/29077977c5ebb92c6122484705b2c867.svg","isPro":false,"fullname":"cossmo","user":"cossmo","type":"user"},{"_id":"668bdbf4db71d9daf74909ec","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/668bdbf4db71d9daf74909ec/HspmqM3ITYR4vn59z1hxt.jpeg","isPro":false,"fullname":"Mikhail Komarov","user":"AsphodelRem","type":"user"},{"_id":"660da14e620d024d7062fc7f","avatarUrl":"/avatars/e37a0a6a44469c3bfe94b537e49b6192.svg","isPro":false,"fullname":"Vika_kondrashuk","user":"Kondrashuk","type":"user"},{"_id":"661d4fdfbca1a6038b362e8a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/AcSzMXgu7FaXn6FHvFAkA.jpeg","isPro":false,"fullname":"Nikolay Nikitin","user":"nnikitin","type":"user"},{"_id":"5f91b1208a61a359f44e1851","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/5f91b1208a61a359f44e1851/4hQd2Bt8AeSnQsluli5f1.jpeg","isPro":true,"fullname":"ai-forever","user":"ai-forever","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"},{"_id":"65de080d7b8cd1aab189a284","avatarUrl":"/avatars/b9e84281959004689c5be6401801c408.svg","isPro":false,"fullname":"Yana Dementyeva","user":"YanaDementyeva","type":"user"},{"_id":"643984dceb7c5616ef3f5d54","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/643984dceb7c5616ef3f5d54/10JRkblrRIEVci6UJwvPz.jpeg","isPro":false,"fullname":"Andrey Kuznetsov","user":"kuznetsoffandrey","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"66a74fd68686cd5495598198","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66a74fd68686cd5495598198/RrVQlvZxAlyCzN9ZRRgf3.jpeg","isPro":false,"fullname":"Ivan Komarov","user":"datakomarov","type":"user"},{"_id":"6669d8dc1537c1bb0f8f0913","avatarUrl":"/avatars/57dccbfb0825e7108ecdff2c244219bd.svg","isPro":false,"fullname":"Ildar","user":"ParaSilv","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":3,"organization":{"_id":"62b1e262f4a72794188b6757","name":"NSU","fullname":"Novosibirsk State University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655825086683-62b1e0f76a5435fd9a60a8dc.png"},"query":{}}">

Cite arxiv.org/abs/2607.11683 in a model README.md to link it from this page.

Cite arxiv.org/abs/2607.11683 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2607.11683 in a Space README.md to link it from this page.