{"slug": "splatrag-3-visualizer-scifact-0-8804-0-7624-with-5000-slang-rows-in-the-same", "title": "SplatRAG 3+ Visualizer: SciFact 0.8804 / 0.7624 with 5,000 slang rows in the same memory", "summary": "SplatRAG 3, a retrieval system using local Gaussian-splat memory, achieved a SciFact (BEIR test, 300 claims, k=10) recall@10 of 0.8804 and nDCG@10 of 0.7624 on a store that also contains 5,000 Urban Dictionary rows, surpassing its own floors of 0.88 and 0.75. The improvement over the previous public SplatRagBench hybrid (0.7822 nDCG@10 on the first 50 claims) comes from adding a third reciprocal rank fusion arm using the query's full 4096-d Qwen3-Embedding-8B cosine, while the 64-d HNSW and BM25 remain in the hybrid. The eval filters domain=scifact so slang cannot rank as a false positive, but the science documents still share the same memory field as the junk.", "body_md": "splatRAG 3 teaser — SciFact, poisoned store\n\nLocal Gaussian-splat memory. Append-only cold log. BM25 + 64-d HNSW. The 3-d field is PCA of those 64-d vectors — the picture, not the index.\n\nSciFact (BEIR test, 300 claims, k=10) on a store that also holds 5,000 Urban Dictionary rows, dreamed together:\n\n┌───────────────┬────────┬─────────┐\n\n│ │ R@10 │ nDCG@10 │\n\n├───────────────┼────────┼─────────┤\n\n│ Floors we set │ 0.88 │ 0.75 │\n\n├───────────────┼────────┼─────────┤\n\n│ This run │ 0.8804 │ 0.7624 │\n\n└───────────────┴────────┴─────────┘\n\nHandshake: splatrag handshake --k 10 --no-ingest. Eval filters domain=scifact so slang cannot rank as a false positive. The science documents still sat in the same field as the junk.\n\nThe old public SplatRagBench hybrid 0.7822 nDCG@10 is the first 50 claims (Nomic). Honest full-300 on that binary was 0.6664.\n\nWhat moved v3 over the floors: RRF (BM25×1.3, k=10) on the top 30, then a third RRF arm from the query’s full 4096-d Qwen3-Embedding-8B cosine. 64-d was already in hybrid; the leftover 4032 dimensions were sitting unused. Pure cosine rerank of that pool died (0.7174). SciFact is still lexical first.\n\nBasins pack in 64-d (k-NN, cosine 0.80). 3-d union-find had mashed 4,035 papers into one unlabeled well.\n\nBelow is example of AI’s memory. On below you see one big color cause basins still need to settle.", "url": "https://wpnews.pro/news/splatrag-3-visualizer-scifact-0-8804-0-7624-with-5000-slang-rows-in-the-same", "canonical_source": "https://discuss.huggingface.co/t/splatrag-3-visualizer-scifact-0-8804-0-7624-with-5-000-slang-rows-in-the-same-memory/179039#post_1", "published_at": "2026-08-21 03:57:20+00:00", "updated_at": "2026-08-21 04:18:09.938914+00:00", "lang": "en", "topics": ["machine-learning", "artificial-intelligence", "ai-research"], "entities": ["SplatRAG", "SciFact", "BEIR", "Urban Dictionary", "Qwen3-Embedding-8B", "SplatRagBench"], "alternates": {"html": "https://wpnews.pro/news/splatrag-3-visualizer-scifact-0-8804-0-7624-with-5000-slang-rows-in-the-same", "markdown": "https://wpnews.pro/news/splatrag-3-visualizer-scifact-0-8804-0-7624-with-5000-slang-rows-in-the-same.md", "text": "https://wpnews.pro/news/splatrag-3-visualizer-scifact-0-8804-0-7624-with-5000-slang-rows-in-the-same.txt", "jsonld": "https://wpnews.pro/news/splatrag-3-visualizer-scifact-0-8804-0-7624-with-5000-slang-rows-in-the-same.jsonld"}}