Edit

Filters in vector queries

  • Article

You can set a vector filter modes on a vector query to specify whether you want filtering before or after query execution.

Filters determine the scope of a vector query. Filters are set on and iterate over nonvector string and numeric fields attributed as filterable in the index, but the purpose of a filter determines what the vector query executes over: the entire searchable space, or the contents of a search result.

This article describes each filter mode and provides guidance on when to use each one.

Prefilter mode

Prefiltering applies filters before query execution, reducing the search surface area over which the vector search algorithm looks for similar content. In a vector query, preFilter is the default.

Diagram of prefilters.

Postfilter mode

Post-filtering applies filters after query execution, narrowing the search results.

Diagram of post-filters.

Benchmark testing of vector filter modes

To understand the conditions under which one filter mode performs better than the other, we ran a series of tests to evaluate query outcomes over small, medium, and large indexes.

  • Small (100,000 documents, 2.5-GB index, 1536 dimensions)
  • Medium (1 million documents, 25-GB index, 1536 dimensions)
  • Large (1 billion documents, 1.9-TB index, 96 dimensions)

For the small and medium workloads, we used a Standard 2 (S2) service with one partition and one replica. For the large workload, we used a Standard 3 (S3) service with 12 partitions and one replica.

Indexes had an identical construction: one key field, one vector field, one text field, and one numeric filterable field. The following index is defined using the 2023-07-01-preview syntax.

def get_index_schema(self, index_name, dimensions):
    return {
        "name": index_name,
        "fields": [
            {"name": "id", "type": "Edm.String", "key": True, "searchable": True},
            {"name": "content_vector", "type": "Collection(Edm.Single)", "dimensions": dimensions,
              "searchable": True, "retrievable": True, "filterable": False, "facetable": False, "sortable": False,
              "vectorSearchConfiguration": "defaulthnsw"},
            {"name": "text", "type": "Edm.String", "searchable": True, "filterable": False, "retrievable": True,
              "sortable": False, "facetable": False},
            {"name": "score", "type": "Edm.Double", "searchable": False, "filterable": True,
              "retrievable": True, "sortable": True, "facetable": True}
        ],
        "vectorSearch":
        {
            "algorithmConfigurations": [
                {"name": "defaulthnsw", "kind": "hnsw", "hnswParameters": {"metric": "euclidean"}}
            ]
        }
    }

In queries, we used an identical filter for both prefilter and postfilter operations. We used a simple filter to ensure that variations in performance were due to filtering mode, and not filter complexity.

Outcomes were measured in Queries Per Second (QPS).

Takeaways

  • Prefiltering is almost always slower than postfiltering, except on small indexes where performance is approximately equal.

  • On larger datasets, prefiltering is orders of magnitude slower.

  • So why is prefilter the default if it's almost always slower? Prefiltering guarantees that k results are returned if they exist in the index, where the bias favors recall and precision over speed.

  • Postfiltering is for customers who:

    • value speed over selection (postfiltering can return fewer than k results)
    • use filters that are not overly selective
    • have indexes of sufficient size such that prefiltering performance is unacceptable

Details

  • Given a dataset with 100,000 vectors at 1536 dimensions:

    • When filtering more than 30% of the dataset, prefiltering and postfiltering were comparable.
    • When filtering less than 0.1% of the dataset, prefiltering was about 50% slower than postfiltering.

  • Given a dataset with 1 million vectors at 1536 dimensions:

    • When filtering more than 30% of the dataset, prefiltering was about 30% slower.
    • When filtering less than 2% of the dataset, prefiltering was about seven times slower.

  • Given a dataset with 1 billion vectors at 96 dimensions:

    • When filtering more than 5% of the dataset, prefiltering was about 50% slower.
    • When filtering less than 10% of the dataset, prefiltering was about seven times slower.

The following graph shows prefilter relative QPS, computed as prefilter QPS divided by postfilter QPS.

Chart showing QPS performance for small, medium, and large indexes for relative QPS.

The vertical axis is QPS of prefiltering over QPS of postfiltering. For example, a value of 0.0 means prefiltering is 100% slower, 0.5 on the vertical axis means prefiltering is 50% slower, 1.0 means prefiltering and post filtering are equivalent.

The horizontal axis represents the filtering rate, or the percentage of candidate documents after applying the filter. For example, 1.00% means that one percent of the search corpus was selected by the filter criteria.