Pelorus Introduces Warped‑Product Graph Index for Hyperbolic Vector Search
A new warped‑product proximity‑graph index called Pelorus aims to keep hyperbolic search accurate while cutting memory use.
Source: Grafy Research LabA new indexing structure called Pelorus has been introduced for hyperbolic vector search, promising better fidelity and memory efficiency at million‑scale workloads. The method, described in a recent arXiv paper by Wenxuan Xiao, Xu Cao and colleagues, tackles long‑standing challenges of preserving geodesic order while keeping the index compact.
Why hyperbolic embeddings need a new ANN index
Hyperbolic embeddings are increasingly popular for representing hierarchical data such as taxonomies, knowledge graphs, and social networks because distances grow exponentially with depth, mirroring tree‑like structures. However, scaling approximate nearest neighbor (ANN) search to millions of points in hyperbolic space is difficult. Traditional flat inner‑product indices, which work well in Euclidean space, tend to scramble the ordering of vectors that lie on different radial shells, breaking the geodesic ordering that hyperbolic metrics rely on. Native hyperbolic graph indexes avoid this distortion but typically store every vector in full precision or quantize coordinates with a single conformal scale, inflating memory consumption and forcing repeated full‑precision reads during query time. According to the authors, reconciling geodesic fidelity with memory efficiency remains an open problem.
How Pelorus re‑thinks the index architecture
Pelorus proposes a warped‑product proximity‑graph that decouples three distinct components of a hyperbolic vector:
- Radial and conformal scalars – stored unquantized in to retain exact magnitude information.
- Spherical direction – encoded with a scale‑invariant 2‑bit code, capturing orientation while using only a fraction of the bits required for full floating‑point storage.
- Full‑precision vector – retained in but accessed only when a final, exact distance computation is needed.
By separating magnitude from direction, Pelorus builds a graph where edges are formed based on the warped product of these components. The 2‑bit directional codes act as a lightweight filter that quickly prunes irrelevant neighbors, while the scalars guide the search through the radial hierarchy. Only the most promising candidates trigger a deferred evaluation, preserving the exact hyperbolic geodesic distance for the final ranking. This design dramatically reduces the amount of data read per query without sacrificing the ordering guarantees that hyperbolic metrics demand.
What this means for developers and researchers
The primary benefit of Pelorus is a lower memory footprint combined with a query pipeline that scales to millions of vectors. For creators building hierarchical recommendation systems, knowledge‑graph retrieval services, or any application that relies on hyperbolic embeddings, the index offers:
- Reduced storage costs – directional information occupies only 2 bits per vector, cutting overall index size compared with full‑precision storage.
- Faster average‑case queries – the lightweight directional filter eliminates many distance calculations before the expensive step.
- Preserved geodesic accuracy – because radial magnitudes remain unquantized, the ordering of points across shells stays faithful to the true hyperbolic geometry.
The paper supplies code and artifacts, allowing practitioners to benchmark Pelorus against existing hyperbolic ANN methods on their own datasets. Early experiments in the supplementary material suggest competitive recall rates with up to an order‑of‑magnitude reduction in memory usage, though the authors note that the final evaluation still dominates the tail latency for very high‑recall settings.
Trying Pelorus with Grafy tools
Grafy provides several ways to explore and prototype the Pelorus index without leaving the platform:
- Reader – Load the arXiv PDF directly in /reader and ask the AI assistant to summarize sections, extract the pseudocode, or clarify the warped‑product construction.
- Code – Use /code to clone the public repository linked in the paper, run the provided Python notebooks, and experiment with custom datasets. The integrated terminal lets you install dependencies, execute the index build, and profile query latency.
- Graph – Visualize the proximity‑graph structure by feeding node and edge data into the /graph app. You can branch different parameter settings (e.g., varying the size of the directional codebook) and compare their impact on memory usage and recall.
- Chat – If you hit a roadblock, the multi‑model chat at /chat can help you troubleshoot implementation errors, suggest alternative hyperbolic libraries, or generate boiler‑plate code for integrating Pelorus into existing pipelines.
By combining these tools, creators can move from reading the research paper to a working prototype in a single workspace, keeping all experiments, code, and results linked together.
Remaining challenges and future directions
While Pelorus introduces an elegant separation of magnitude and direction, the approach still relies on a deferred computation for the final ranking, which can become a bottleneck for ultra‑low‑latency use cases. Additionally, the 2‑bit directional encoding assumes a uniform angular distribution; datasets with highly anisotropic directionality may suffer reduced pruning effectiveness. The authors acknowledge that extending the warped‑product graph to support dynamic updates (insertion or deletion of vectors) without rebuilding the entire index is an open research question. Future work could explore learned directional quantizers or hardware‑accelerated kernels to further shrink query times.
Frequently asked questions
How does Pelorus differ from traditional hyperbolic graph indexes?
Pelorus separates radial magnitude (stored in ) from spherical direction (encoded in 2‑bit codes) and only resorts to full distance calculations for the final shortlist. Traditional hyperbolic graphs either keep every vector in full precision or apply a single conformal scale, leading to higher memory use and more frequent expensive distance reads.
Can I use Pelorus with embeddings that are not strictly hyperbolic?
The warped‑product construction relies on the hyperbolic metric’s radial and angular decomposition. While the paper focuses on pure hyperbolic embeddings, the authors suggest that the same separation could be adapted to other constant‑curvature spaces, though performance guarantees would need separate evaluation.
Is the Pelorus code ready for production deployment?
The authors have released a reference implementation and benchmark scripts, but the repository is positioned as a research prototype. Production use would likely require additional engineering for scalability, fault tolerance, and integration with existing vector‑search services.
Source: Xiao, W., Cao, X., et al., “Pelorus: Warped‑Product Graph Indexing for Hyperbolic Vector Search,” arXiv:2611.22159, 2026.
Source: Grafy Research Lab — Pelorus: Warped-Product Graph Indexing for Hyperbolic Vector Search


