Uredi

Deli putem

Index binary vectors for vector search

  • Članak

Azure AI Search supports a packed binary type of Collection(Edm.Byte) for further reducing the storage and memory footprint of vector data. You can use this data type for output from models such as Cohere's Embed v3 binary embedding models.

There are three steps to configuring an index for binary vectors:

  • Add a vector search algorithm that specifies Hamming distance for binary vector comparison
  • Add a vector profile that points to the algorithm
  • Add the vector profile to your binary field definition

This article assumes you're familiar with creating an index in Azure AI Search. It uses the REST APIs to illustrate each step, but you could also add a binary field to an index in the Azure portal.

Binary data types are generally available starting with API version 2024-07-01 and are assigned to fields using the Create Index or Create Or Update Index APIs.

Tip

If you're investigating binary vector support for its smaller footprint, you might also consider the vector quantization and storage reduction features in Azure AI Search. Inputs are float32 or float16 embeddings. Output is stored data in a much smaller format. For more information, see Reduce vector size through quantization, narrow data types, and storage options.

Prerequisites

  • Binary vectors, with 1 bit per dimension, packaged in uint8 values with 8 bits per value. These can be obtained by using models that directly generate "packaged binary" vectors, or by quantizing vectors into binary vectors client-side during indexing and searching.

Limitations

  • No Azure portal support in the Import and vectorize data wizard.
  • No support for binary fields in the AML skill that's used for integrated vectorization of models in the Azure AI Studio model catalog.

Add a vector search algorithm and vector profile

Vector search algorithms are used to create the query navigation structures during indexing. For binary vector fields, vector comparisons are performed using the Hamming distance metric.

  1. To add a binary field to an index, set up a Create or Update Index request using the REST API or the Azure portal.

  2. In the index schema, add a vectorSearch section that specifies profiles and algorithms.

  3. Add one or more vector search algorithms that have a similarity metric of hamming. It's common to use Hierarchical Navigable Small Worlds (HNSW), but you can also use Hamming distance with exhaustive K-nearest neighbors.

  4. Add one or more vector profiles that specify the algorithm.

The following example shows a basic vectorSearch configuration:

  "vectorSearch": { 
    "profiles": [ 
      { 
        "name": "myHnswProfile", 
        "algorithm": "myHnsw", 
        "compression": null, 
        "vectorizer": null 
      } 
    ], 
    "algorithms": [ 
      { 
        "name": "myHnsw", 
        "kind": "hnsw", 
        "hnswParameters": { 
          "metric": "hamming" 
        } 
      }, 
      { 
        "name": "myExhaustiveKnn", 
        "kind": "exhaustiveKnn", 
        "exhaustiveKnnParameters": { 
          "metric": "hamming" 
        } 
      } 
    ] 
  }

Add a binary field to an index

The fields collection of an index must include a field for the document key, vector fields, and any other fields that you need for hybrid search scenarios.

Binary fields are of type Collection(Edm.Byte) and contain embeddings in packed form. For example, if the original embedding dimension is 1024, the packed binary vector length is ceiling(1024 / 8) = 128. You get the packed form by setting the vectorEncoding property on the field.

  • Add a field to the fields collection and give it name.
  • Set data type to Collection(Edm.Byte).
  • Set vectorEncoding to packedBit for binary encoding.
  • Set dimensions to 1024. Specify the original (unpacked) vector dimension.
  • Set vectorSearchProfile to a profile you defined in the previous step.
  • Make the field searchable.

The following field definition is an example of the properties you should set:

  "fields": [ 
    . . . 
    { 
      "name": "my-binary-vector-field", 
      "type": "Collection(Edm.Byte)", 
      "vectorEncoding": "packedBit", 
      "dimensions": 1024, 
      "vectorSearchProfile": "myHnswProfile",
      "searchable": true
    },
   . . . 
  ]

See also

Code samples in the azure-search-vector-samples repository demonstrate end-to-end workflows that include schema definition, vectorization, indexing, and queries.

There's demo code for Python, C#, and JavaScript.