Milvus can run hyperbolic search with plain inner product, but every benchmark so far comes from Zilliz

What we checked

• What concrete evidence would resolve the remaining editorial uncertainty for this item? partly confirmed. The primary post, the Milvus v3.0.1 release, the model card, the related papers and the companion Zilliz Cloud post all exist and agree with each other, so the item is real and specific. What is still missing is a third-party reproduction of the recall and latency numbers, or a public code repository. The post includes code snippets but links no standalone repo. [1][2][4][3]
• Does Milvus ship a native hyperbolic distance metric? (added) refuted. The post says outright that Milvus has no hyperbolic metric. The v3.0.1 release notes do not mention hyperbolic support. Their reranking additions are L1 reranking in function chains and weighted RRF. The method is a transformation applied to vectors that then use the ordinary IP metric. [1][2]
• Are the sign flip and the centering step exactly rank-preserving? (added) confirmed. The Lorentz inner product is −x0·y0 plus the sum of the spatial products. Negating the query's x0 therefore turns a standard dot product into the Lorentz score, and hyperbolic distance shrinks as that score grows. Subtracting a constant m from every document's x0 shifts each score for a given query by q0·m, which is the same for every document, so the ordering does not change. Part 1 of the series makes this argument and the algebra holds. [6][1]

What we found

• The method has two steps. At query time, negate the query's first coordinate (the 'height'), so that a standard metric_type="IP" search returns the Lorentz ranking. At ingest, subtract the corpus-mean height from each stored document ('centering'). The post describes two deployment strategies. Strategy A stores a single centered 513-dimensional vector and runs one search. Strategy B runs a cosine shortlist of 200 candidates over 512-dimensional directions, then reranks with the exact Lorentz score, either in the application or with a Milvus 3.0.1 function chain. [1]
• Centering changes how well the indexes approximate the search, not the exact ranking. Recall@10 over 1,000 text queries went from 0.974 to 0.989 on HNSW at ef=64, 0.989 to 0.996 at ef=128, 0.900 to 0.990 on IVF_FLAT at nprobe=32, and 0.884 to 0.969 on IVF_SQ8 at nprobe=32. Centered HNSW at ef=128 is reported as statistically equal to the shortlist-and-rerank strategy. [1]
• The post also reports limits. IVF_PQ performs poorly with or without centering (0.096 centered vs 0.115 raw recall@10), and the author advises against it. SCANN rejects the 513-dimensional vectors because it requires an even dimension count. The post makes no latency claims, saying laptop measurements were too noisy, and it notes that fp16 encoding means the embeddings are not exact. [1]
• Milvus v3.0.1 was released on September 9, 2026. Its notes list L1 reranking in search function chains and weighted RRF reranking, and say nothing about hyperbolic geometry. That fits the post's framing: the in-database rerank path uses general function-chain features, not a dedicated hyperbolic operator. The post's own check found the function chain matched application-side reranking on 100 queries over 10,000 images, with scores within 5×10⁻⁷. [2][1]
• Part 3 adds managed-service numbers on a Zilliz Cloud cluster in AWS eu-west-1. Centered single search at level 10 gave 0.977 recall@10 at 3.75 ms p50 and 1,698 peak QPS. A cosine shortlist plus rerank with 200 candidates gave 0.995 at 4.37 ms. The post also reports that hyperbolic vectors need level 10 to approach the recall CLIP ViT-B/16 reaches at level 1, with roughly 23% lower peak throughput. [3]
• The model, hyper3-clip-v1, is a roughly 0.1B-parameter hyperbolic vision-language model from hyper³labs. It is gated on Hugging Face under the OpenMDW-1.0 license, and Lorentz scoring is its default, with cosine and cone scoring as alternatives. The model card does not state the embedding dimension or curvature. The 513 dimensions and c=0.1 come from the Milvus posts. The accompanying paper is an ECCV 2026 Beyond Euclidean Workshop submission, dated August 29, 2026. [4][5]

What it means

If you are evaluating Lorentz-model embeddings, you probably don't need a specialised vector store. Negating the query's height and doing an IP search gives exact Lorentz top-k when the search is exhaustive. Subtracting the corpus-mean height at ingest is a one-line change that preserves the ranking and, per these tests, sharply improves IVF recall. It is also worth checking on whichever ANN engine you use, since the reason it helps, a smaller dynamic range in the large first coordinate, is not specific to Milvus. Avoid PQ-style compression for these vectors, and remember that the odd dimension count (n+1) can break indexes that need even dimensions. If you need recall above about 0.995, the cosine-shortlist-plus-exact-rerank design is the safer default.

Treat the numbers as a vendor's single-model, single-dataset result. Everything here comes from Zilliz-published posts using one checkpoint on ImageNet. Part 3 itself shows hyperbolic search needing more search effort than CLIP to reach the same recall. Rebuild the recall curves on your own corpus and model before committing. The geometry argument is solid, but the operating points will vary.

Still open

• Has anyone outside Zilliz reproduced the recall@10 figures, and is there a public benchmark repository?
• Does centering help as much for other hyperbolic models (for example HypRAG's text retrievers) and for text corpora rather than ImageNet images?
• How do the methods compare on retrieval quality, not just agreement with the exact Lorentz ranking, against strong Euclidean baselines on the same data?
• What does part 4 ('what hyperbolic geometry does to ranking') conclude, and does it change the deployment recommendation?