RAG-Anything Read Its OWN Paper… and the Results Are WILD

I gave RAG-Anything (HKU's all-in-one multimodal RAG) its own research paper and asked it about the tables, figures and equations, then checked every answer against the real page. I also asked it two trick questions. Result in this run: 6 of 7 correct, including both trick questions. It declined one question whose answer was only in a chart. What's in the video: - How it works: MinerU parsing into "atomic units" → a cross-modal graph + a text graph merged by entity name → hybrid graph + vector retrieval → a vision-language model answers with the original image - What the paper claims (DocBench 63.4 overall vs 61.0 MMGraphRAG; 76.3 vs 66.0 on multimodal questions) and the caveats (author-run, graded by GPT-4o-mini, lowest "unanswerable" score at 46.0) - Running it on an RTX 4060 laptop (8 GB). Install fix: RAG-Anything 1.4.3 requires mineru[core] version 3.4.1 or newer with no upper limit, and MinerU 4 removed the -p flag it calls. Fix: install mineru[core] 3.4.x, below version 4 (I used 3.4.5). - Real cost: indexing the 18-page paper took 39.7 min and $0.31 on OpenRouter (GPT-5.6 Luna for text + vision, text-embedding-3-large). GPU parsing was ~2 min of that. Setup used: raganything 1.4.3 · mineru 3.4.5 · torch 2.6.0+cu124 · OpenRouter openai/gpt-5.6-luna (LLM + VLM) · openai/text-embedding-3-large (3072-d) · hybrid query mode. The 7 questions: 1. Table 2 multimodal accuracy → 76.3 ✓ 2. What does Equation 6 compute? ✓ 3. The Novo Nordisk case study (26,778M vs 32,928 / 11,503) ✓ 4. Graphs vs reranker (ablation 60.0 / 62.4 / 63.4) ✓ 5. Trick: "Legal domain of MMLongBench" (doesn't exist) → "not reported, 60.2 is DocBench" ✓ 6. False premise: "why does it beat MMGraphRAG on Academia?" (it doesn't: 61.4 vs 64.3) ✓ 7. Figure 2, 1–10 pages: which is better? (chart only) → declined ⚠ 7 questions on one paper is not a benchmark. Links - Repo: https://github.com/HKUDS/RAG-Anything - Paper: https://arxiv.org/abs/2510.12323 - RunPod: https://get.runpod.io/pe48 Chapters 0:00 Its own results table 0:41 What RAG-Anything is 1:31 How it works: parse, two graphs, hybrid search 3:27 What the paper claims (and the caveats) 4:34 Running it on an RTX 4060 + the MinerU 4 fix 6:48 7 tests against its own paper 9:44 Verdict Get GPUs Runpod: https://get.runpod.io/pe48 Get Hostinger: https://www.hostg.xyz/SHJiu Coupon Code "PROMPT" 10% discount 🔗 Connect with me: Email → [email protected] Ko-fi → https://ko-fi.com/promptengineer Patreon → https://www.patreon.com/PromptEngineer975 Book a Call → https://calendly.com/prompt-engineer48/call GitHub → https://github.com/PromptEngineer48 Twitter/X → https://twitter.com/prompt48 Website → https://promptengineer48.com/