A small AI model beat a bigger rival at finding complete evidence — so maybe the model's size was never what mattered.

This explores why RAGU reports more complete retrieved evidence than HippoRAG 2 despite its small model, and whether model size is the relevant variable at all. In the RAGU paper's own account it isn't. The 7B model doesn't do the retrieving. It only builds the knowledge graph ahead of time, by pulling entities and relations out of text. The reported completeness (evidence recall up to 0.84, against at most 0.76 for HippoRAG 2) comes from what RAGU does with the graph after extraction. So "despite a smaller model" is closer to "the small model was never the bottleneck."

The paper's argument is that most GraphRAG systems treat graph construction as a single extraction pass. That pass produces noisy, duplicated entities and has no way to merge information scattered across chunks. RAGU splits the work into stages: typed two-stage extraction, deduplication, LLM summarization, and then Leiden community detection. Facts about the same entity end up in the same place, and communities come with summaries. A query therefore pulls back a consolidated picture instead of a scatter of fragments. The community-detection step is the same one that lets GraphRAG answer questions about a whole corpus, which plain chunk retrieval cannot do efficiently (Can community detection enable RAG systems to answer global corpus questions?). Completeness is what this design optimizes for.

It costs something. In the paper's comparison, HippoRAG 2 leads on retrieval precision and on chain-following multi-hop questions on MuSiQue, while RAGU leads on synthesis tasks that need the whole picture. The two systems are built for different jobs: one follows a thread of facts precisely, the other gathers everything relevant. The same trade-off appears elsewhere in RAG. LongRAG-style designs give up precise retrieval for coarse retrieval plus deep reading (Can long-context models resolve retriever-reader imbalance?). More broadly, retrieval failures tend to be architectural rather than a matter of tuning (Where do retrieval systems fail and why?), so a different retrieval structure fits a different kind of question. The paper also says part of RAGU's wider multi-hop gap under verbose prompts is an answer-format artifact.

The small model can work because the paper claims that extraction and reasoning over context are language skills, not factual recall. It says these skills grow only weakly with model size, while world knowledge scales steeply. Its evidence is Meno-Lite-0.1, a 7B model that beats Qwen2.5-32B at knowledge-graph construction (+12.5% relative harmonic mean) and matches it on English GraphRAG tasks. If that holds, a compact extractor costs little in quality, and a consolidating pipeline can spend its effort where completeness comes from.

The library can't say that the small model causes the completeness advantage. The scaling claim is asserted in the excerpt rather than demonstrated. The benchmarks are limited to GraphRAG-Bench (Medical) and MuSiQue. The library also doesn't say which model HippoRAG 2 used in the comparison, and the comparison isn't a controlled test that isolates model size. The supported reading is narrower. RAGU's completeness most plausibly comes from consolidating the graph and summarizing communities, and the compact model shows that step doesn't need a large extractor.