search-dataset-accuracy.yaml

# Example bfb search config that measures search accuracy (recall) against a
# reference dataset, using the same dataset format as vector-db-benchmark.
#
#   # 1. Upload the dataset corpus (assigns point id = dataset row index):
#   bfb upload --example upload-dataset-config
#
#   # 2. Search with the dataset's query set and measure recall:
#   bfb search --example search-dataset-accuracy -p 8 -t 4 \
#       --search-limit 10 --uri http://localhost:6334
#
# When a request's `source` is `type: dataset`, bfb draws query vectors from the
# dataset's *query set* (not the corpus) and, for every query, compares the
# returned point ids against the dataset's ground-truth nearest neighbors.
# Recall is reported under "--- Precision ---" as `|found ∩ expected[:k]| / k`.
#
# If the query set carries per-query `conditions` (the ann-filtering-benchmark
# datasets: laion-small-clip, arxiv-titles-*-filters, h-and-m-*-filters, the
# random-*-filters family), its ground truth answers the FILTERED query, so bfb
# applies each query's own conditions. A `filters:` block on the same request is
# then ignored — the dataset defines the filter. bfb prints how many queries
# carry conditions when it opens the query set.
#
# IMPORTANT: accuracy only lines up when the corpus was uploaded with the
# default integer id scheme (point id == dataset row index), which is what
# `bfb upload` does for `id: integer` collections.
#
# Supported query/ground-truth sources (auto-detected from the dataset files):
#   * h5 (ann-benchmarks): `test` (queries) + `neighbors` (ground truth)
#   * tar (ann-filtering-benchmark-datasets): `tests.jsonl` with
#     `query` / `closest_ids`
#   * sparse: `queries.csr` + `results.gt`

collection:
  # Must match the collection uploaded in step 1 (`--example upload-dataset-config`).
  name: glove-25-angular

requests:
  # Dense recall against the glove-25-angular query set (h5 `test`/`neighbors`).
  # `size` is optional for dataset sources (the dataset defines the dimension);
  # it is kept here only for documentation.
  - kind: dense
    size: 25
    source:
      type: dataset
      name: glove-25-angular
      format: h5
      path: glove-25-angular/glove-25-angular.hdf5
      link: http://ann-benchmarks.com/glove-25-angular.hdf5

  # Sparse recall example (uncomment and point at a sparse dataset that ships a
  # `queries.csr` + `results.gt` query set):
  #
  # - kind: sparse
  #   using: bm25
  #   source:
  #     type: dataset
  #     name: my-sparse-dataset
  #     format: sparse
  #     path: my-sparse-dataset/my-sparse-dataset
  #     link: https://example.com/my-sparse-dataset.tgz