[BUG] Filtered Search Recall Collapse on WIKI-1M under Single, Identical Query Labels

[sujingbo0217]

Describe the bug
While performing filtered ANNS on WIKI-1M dataset with uniform query labels (i.e., every query point has the same one label) to evaluate query performance under a specific level of specificity, the system returns identical neighbor lists for every query. This results in a total recall collapse toward 0. This behavior is observed on both IVF-Graph (high-specificity) and IVF-BFS (low-specificity) indexes.

Steps/Code to reproduce bug

1. Load WIKI-1M dataset.
2. Generate a query_labels.txt where each line is identical (e.g., lbls[0] = lbls[1] = ... = lbls[n-1] = [0]).
3. Run search using the following configuration (change spec_threshold to 2000 to evaluate IVF-BFS query for label 3079):

{
 "data_dir": "/data/ann/wiki_1M/",
 "data_fname": "base.fbin",
 "query_fname": "query.fbin",
 "data_label_fname": "base_labels.txt",
 "query_label_fname": "query_labels1.txt",
 "itopk_size": 32,
 "spec_threshold": 1500,
 "graph_degree": 32,
 "topk": 10,
 "num_runs": 1000,
 "warmup_runs": 10,
 "force_rebuild": true,
 "ivf_graph_fname": "ivf_graph.bin",
 "ivf_bfs_fname": "ivf_bfs.bin",
 "ground_truth_fname": "ground_truth_k10.ibin"
}

Environment details

• Dataset: WIKI-1M
• CPU: AMD Ryzen 7 5700G
• GPU: NVIDIA RTX 2080 Ti
• OS: Ubuntu 22.04 LTS
• NVIDIA Driver: 12.2
• CUDA Compiler: 12.9.86
• Host Compiler: g++ 11.4.0

Observed behavior
Host validation confirms (first two) query vectors are distinct ($L_2$ distance $\approx 0.62$), yet neighbors are duplicated (nbhs[0] = nbhs[1] = ... = nbhs[n-1]).
Cross-path confirmation:

• IVF-Graph: Label 0 and 1 (spec=15.33% and 13.05%), specificity > threshold — FAILED
• IVF-BFS: Label 3079 (0.20%), specificity < threshold — FAILED

Supporting evidence
IVF-Graph (WIKI-1M label 1, spec=13.05%) query results snippet:

IVF-Graph Index Stats:
 Total vectors:  980312
 Number of labels: 3814
 Graph size:     [21768457 × 32]
 Graph degree:   32

IVF-BFS Index Stats:
 Number of labels: 186
 Number of rows:  269772

QPS: 380420.79
Recall: 0.0000

=== Search Results Monitor (first 3 queries) ===
Query 0 (label=0):
 neighbors: 541004 316361 314017 331946 604415 448044 539479 291316 344698 103825
 gt:       370208 251743 484555 597579 386190 368896 860059 781968 802148 401846
 recall@10: 0/10
Query 1 (label=0):
 neighbors: 541004 316361 314017 331946 604415 448044 539479 291316 344698 103825
 gt:       341406 160363 370208 251743 484555 874429 712954 785765 457944 517840
 recall@10: 0/10
Query 2 (label=0):
 neighbors: 541004 316361 314017 331946 604415 448044 539479 291316 344698 103825
 gt:       549821 533776 915603 251743 484555 597579 573700 882094 276817 331536
 recall@10: 0/10

IVF-BFS (WIKI-1M label 3079 spec=0.20%) query results snippet:

IVF-Graph Index Stats:
  Total vectors:  980312
  Number of labels: 2976
  Graph size:     [20314630 × 32]
  Graph degree:   32

IVF-BFS Index Stats:
  Number of labels: 1024
  Number of rows:  1723599

QPS: 302732.20
Recall: 0.0023

=== Search Results Monitor (first 3 queries) ===
Query 0 (label=3079):
  neighbors: 430417 279428 926084 439948 62295 368557 266782 439498 559943 547068 
  gt:       430417 279428 926084 439948 62295 368557 266782 439498 559943 547068 
  recall@10: 10/10
Query 1 (label=3079):
  neighbors: 430417 279428 926084 439948 62295 368557 266782 439498 559943 547068 
  gt:       751135 773048 427405 742655 944674 345688 398253 943112 272594 565167 
  recall@10: 0/10
Query 2 (label=3079):
  neighbors: 430417 279428 926084 439948 62295 368557 266782 439498 559943 547068 
  gt:       226262 778697 417094 104588 281525 47437 862004 267955 913929 809383 
  recall@10: 0/10

Additional context
For comparison, I also generated synthetic labels in Zipfian distribution for the SIFT-1M dataset and ran the same uniform filtered query with label 1 (spec=75%). This yielded normal results.
The IVF-Graph results snippet:

IVF-Graph Index Stats:
 Total vectors:  1000000
 Number of labels: 51
 Graph size:     [3385661 × 32]
 Graph degree:   32

IVF-BFS Index Stats:
 Number of labels: 0
 Number of rows:  0

QPS: 1068508.05
Recall: 0.9292

=== Search Results Monitor (first 3 queries) ===
Query 0 (label=1):
  neighbors: 932085 934876 561813 695756 701258 455537 562594 908244 600499 893601 
  gt:       932085 934876 561813 708177 706771 695756 435345 701258 455537 562594 
  recall@10: 7/10
Query 1 (label=1):
  neighbors: 413071 880592 249062 400194 942339 880462 941776 586780 248426 849742 
  gt:       413071 706838 880592 249062 400194 942339 880462 941776 420802 586780 
  recall@10: 8/10
Query 2 (label=1):
  neighbors: 408764 408462 861882 406273 406324 551743 861530 402106 239766 823095 
  gt:       408764 408462 861882 406273 406324 551743 861530 402106 239766 823095 
  recall@10: 10/10

Interestingly, this does NOT occur on SIFT-1M (128D), even with uniform labels.

[moccch]

@sujingbo0217 Thanks for reaching out and the detailed report!

A couple of observations that might help narrow this down:

1. IVF-BFS (label 3079): I noticed Query 0 returns a perfect 10/10 recall, while the rest return identical results. It would be interesting to try running individual queries (e.g., just Query 1 by itself) to see if the issue persists and this would help determine whether it's a multi-query interaction problem or something else.
2. IVF-Graph (label 0): It might also be worth checking whether the returned neighbors actually belong to the queried label (i.e., do all returned neighbors have label 0?). If they do, the search is at least respecting the filter constraint and the issue is purely in query differentiation. If they don't, that points to a different cause.

[minjiazhang]

@sujingbo0217, thanks for raising the issue. It seems the other user is able to reproduce Wiki-1M. See #1. Could you take a look and see if there are any differences in setup or configuration?