VOIS – O(1) similarity search, 2.7x faster than FAISS HNSW with perfect recall

1 comments

• $PXQuantumLabs 7 hours ago

     I built a similarity search algorithm that beats FAISS HNSW on both speed     

  and recall. No graph construction, no learned indexes — phase-correlation routing with multi-probe search. O(1) per query, scales to any dimension.

  Tested on the same data, same machine (RTX 4060), same queries, verified against brute-force ground truth:

  VOIS Similarity Search — Patent Pending 100K vectors, 128-dim, top-10 retrieval, RTX 4060

                            QPS        Recall
    VOIS v3 (3 probes):   112,879      95.2%
    VOIS v3 (4 probes):   100,426      99.5%
    VOIS v3 (5 probes):    94,812     100.0%
    FAISS HNSW:            41,668      92.1%
  
    2.7x faster than HNSW with better recall
  
    Multi-dimension scaling (seperate method, same hardware)
  
    100K vectors, top-10, RTX 4060
  
    DIM    VOIS QPS    VOIS Recall   HNSW QPS   HNSW Recall
    128     145,163       100.0%       70,956       89.3%
    256     79,349        99.99%      31,875       80.9%
    512     29,750       100.0%       12,901       78.6%
    768     40,274        97.9%        8,991       74.3%
  
    500K vectors:
  
    DIM    VOIS QPS    VOIS Recall   HNSW QPS   HNSW Recall
    128    32,696       100.0%       31,139       67.3%
    256    14,384       100.0%       17,109       50.9%
    512    5,864       100.0%        8,700       41.8%
  
    HNSW recall collapses at higher dimensions. VOIS is immune.
    100% recall achieved at 2 probes for 512-dim and 768-dim.