Create a vector index

  • Article

In Azure AI Search, a vector store has an index schema that defines vector and nonvector fields, a vector configuration for algorithms that create and compress the embedding space, and settings on vector field definitions that are used in query requests.

The Create or Update Index API creates the vector store. Follow these steps to index vector data:

  • Define a schema with vector algorithms and optional compression
  • Add vector field definitions
  • Load prevectorized data as a separate step, or use integrated vectorization for data chunking and encoding during indexing

This article explains the workflow and uses REST for illustration. Once you understand the basic workflow, continue with the Azure SDK code samples in the azure-search-vector-samples repository for guidance on using these features in test and production code.

Tip

Use the Azure portal to create a vector index and try integrated data chunking and vectorization.

Prerequisites

  • Azure AI Search, in any region and on any tier. Most existing services support vector search. For services created before January 2019, there's a small subset that can't create a vector index. In this situation, a new service must be created. If you're using integrated vectorization (skillsets that call Azure AI), Azure AI Search must be in the same region as Azure OpenAI or Azure AI services.

  • Pre-existing vector embeddings or use integrated vectorization, where embedding models are called from the indexing pipeline.

  • You should know the dimensions limit of the model used to create the embeddings. Valid values are 2 through 3072 dimensions. In Azure OpenAI, for text-embedding-ada-002, the length of the numerical vector is 1536. For text-embedding-3-small or text-embedding-3-large, the vector length is 3072.

  • You should also know what the supported similarity metrics are. For Azure OpenAI, similarity is computed using cosine.

  • You should be familiar with creating an index. The schema must include a field for the document key, other fields you want to search or filter, and other configurations for behaviors needed during indexing and queries.

Prepare documents for indexing

Prior to indexing, assemble a document payload that includes fields of vector and nonvector data. The document structure must conform to the index schema.

Make sure your documents:

  1. Provide a field or a metadata property that uniquely identifies each document. All search indexes require a document key. To satisfy document key requirements, a source document must have one field or property that can uniquely identify it in the index. This source field must be mapped to an index field of type Edm.String and key=true in the search index.

  2. Provide vector data (an array of single-precision floating point numbers) in source fields.

    Vector fields contain an array generated by embedding models, one embedding per field, where the field is a top-level field (not part of a nested or complex type). For the simplest integration, we recommend the embedding models in Azure OpenAI, such as text-embedding-ada-002 for text documents or the Image Retrieval REST API for images.

    If you can take a dependency on indexers and skillsets, consider using integrated vectorization that encodes images and alphanumeric content during indexing. Your field definitions are for vector fields, but source data can be text or images, with vector arrays created during indexing.

  3. Provide other fields with human-readable alphanumeric content for the query response, and for hybrid query scenarios that include full text search or semantic ranking in the same request.

Your search index should include fields and content for all of the query scenarios you want to support. Suppose you want to search or filter over product names, versions, metadata, or addresses. In this case, similarity search isn't especially helpful. Keyword search, geo-search, or filters would be a better choice. A search index that includes a comprehensive field collection of vector and nonvector data provides maximum flexibility for query construction and response composition.

A short example of a documents payload that includes vector and nonvector fields is in the load vector data section of this article.

Add a vector search configuration

A vector configuration specifies the parameters used during indexing to create "nearest neighbor" information among the vector nodes:

  • Hierarchical Navigable Small World (HNSW)
  • Exhaustive KNN

If you choose HNSW on a field, you can opt in for exhaustive KNN at query time. But the other direction doesn't work: if you choose exhaustive, you can't later request HNSW search because the extra data structures that enable approximate search don't exist.

A vector configuration also specifies quantization methods for reducing vector size:

  • Scalar
  • Binary (available in 2024-07-01 only and in newer Azure SDK packages)

For instructions on how to migrate to the latest version, see Upgrade REST API.

2024-07-01 is generally available. It supports a vector configuration having:

  • vectorSearch.algorithms support HNSW and exhaustive KNN.
  • vectorSearch.compressions support scalar and binary quantization, oversampling, and reranking with original vectors.
  • vectorSearch.profiles provide for multiple combinations of algorithm and compression configurations.

Be sure to have a strategy for vectorizing your content. We recommend integrated vectorization and query-time vectorizers for built-in encoding.

  1. Use the Create or Update Index API to create the index.

  2. Add a vectorSearch section in the index that specifies the search algorithms used to create the embedding space.

     "vectorSearch": {
     "compressions": [
         {
             "name": "scalar-quantization",
             "kind": "scalarQuantization",
             "rerankWithOriginalVectors": true,
             "defaultOversampling": 10.0,
                 "scalarQuantizationParameters": {
                     "quantizedDataType": "int8"
                 }
         },
         {
             "name": "binary-quantization",
             "kind": "binaryQuantization",
             "rerankWithOriginalVectors": true,
             "defaultOversampling": 10.0,
         }
     ],
     "algorithms": [
         {
             "name": "hnsw-1",
             "kind": "hnsw",
             "hnswParameters": {
                 "m": 4,
                 "efConstruction": 400,
                 "efSearch": 500,
                 "metric": "cosine"
             }
         },
         {
             "name": "hnsw-2",
             "kind": "hnsw",
             "hnswParameters": {
                 "m": 8,
                 "efConstruction": 800,
                 "efSearch": 800,
                 "metric": "hamming"
             }
         },
         {
             "name": "eknn",
             "kind": "exhaustiveKnn",
             "exhaustiveKnnParameters": {
                 "metric": "euclidean"
             }
         }

     ],
     "profiles": [
       {
         "name": "vector-profile-hnsw-scalar",
         "compression": "scalar-quantization",
         "algorithm": "hnsw-1"
       }
     ]
 }

    Key points:

    • Names for each configuration of compression, algorithm, and profile must be unique for its type within the index.

    • vectorSearch.compressions.kind can be scalarQuantization or binaryQuantization.

    • vectorSearch.compressions.rerankWithOriginalVectors uses the original, uncompressed vectors to recalculate similarity and rerank the top results returned by the initial search query. The uncompressed vectors exist in the search index even if stored is false. This property is optional. Default is true.

    • vectorSearch.compressions.defaultOversampling considers a broader set of potential results to offset the reduction in information from quantization. The formula for potential results consists of the k in the query, with an oversampling multiplier. For example, if the query specifies a k of 5, and oversampling is 20, then the query effectively requests 100 documents for use in reranking, using the original uncompressed vector for that purpose. Only the top k reranked results are returned. This property is optional. Default is 4.

    • vectorSearch.compressions.scalarQuantizationParameters.quantizedDataType must be set to int8. This is the only primitive data type supported at this time. This property is optional. Default is int8.

    • vectorSearch.algorithms.kind are either "hnsw" or "exhaustiveKnn". These are the Approximate Nearest Neighbors (ANN) algorithms used to organize vector content during indexing.

    • vectorSearch.algorithms.m is the bi-directional link count. Default is 4. The range is 4 to 10. Lower values should return less noise in the results.

    • vectorSearch.algorithms.efConstruction is the number of nearest neighbors used during indexing. Default is 400. The range is 100 to 1,000.

    • "vectorSearch.algorithms.fSearch is the number of nearest neighbors used during search. Default is 500. The range is 100 to 1,000.

    • vectorSearch.algorithms.metric should be "cosine" if you're using Azure OpenAI, otherwise use the similarity metric associated with the embedding model you're using. Supported values are cosine, dotProduct, euclidean, hamming (used for indexing binary data).

    • vectorSearch.profiles add a layer of abstraction for accommodating richer definitions. A profile is defined in vectorSearch, and then referenced by name on each vector field. It's a combination of compression and algorithm configurations. This is the property that you assign to a vector field, and it determines the fields' algorithm and compression.

Add a vector field to the fields collection

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

Vector fields are characterized by their data type, a dimensions property based on the embedding model used to output the vectors, and a vector profile.

2024-07-01 is generally available.

  1. Use the Create or Update Index to create the index.

  2. Define a vector field with the following attributes. You can store one generated embedding per field. For each vector field:

    • type must be a vector data types. Collection(Edm.Single) is the most common for embedding models.
    • dimensions is the number of dimensions generated by the embedding model. For text-embedding-ada-002, it's 1536.
    • vectorSearchProfile is the name of a profile defined elsewhere in the index.
    • searchable must be true.
    • retrievable can be true or false. True returns the raw vectors (1536 of them) as plain text and consumes storage space. Set to true if you're passing a vector result to a downstream app.
    • stored can be true or false. It determines whether an extra copy of vectors is stored for retrieval. For more information, see Reduce vector size.
    • filterable, facetable, sortable must be false.

  3. Add filterable nonvector fields to the collection, such as "title" with filterable set to true, if you want to invoke prefiltering or postfiltering on the vector query.

  4. Add other fields that define the substance and structure of the textual content you're indexing. At a minimum, you need a document key.

    You should also add fields that are useful in the query or in its response. The following example shows vector fields for title and content ("titleVector", "contentVector") that are equivalent to vectors. It also provides fields for equivalent textual content ("title", "content") useful for sorting, filtering, and reading in a search result.

    The following example shows the fields collection:

    PUT https://my-search-service.search.azure.cn/indexes/my-index?api-version=2024-07-01&allowIndexDowntime=true
Content-Type: application/json
api-key: {{admin-api-key}}
{
    "name": "{{index-name}}",
    "fields": [
        {
            "name": "id",
            "type": "Edm.String",
            "key": true,
            "filterable": true
        },
        {
            "name": "title",
            "type": "Edm.String",
            "searchable": true,
            "filterable": true,
            "sortable": true,
            "retrievable": true
        },
        {
            "name": "titleVector",
            "type": "Collection(Edm.Single)",
            "searchable": true,
            "retrievable": true,
            "stored": true,
            "dimensions": 1536,
            "vectorSearchProfile": "vector-profile-1"
        },
        {
            "name": "content",
            "type": "Edm.String",
            "searchable": true,
            "retrievable": true
        },
        {
            "name": "contentVector",
            "type": "Collection(Edm.Single)",
            "searchable": true,
            "retrievable": false,
            "stored": false,
            "dimensions": 1536,
            "vectorSearchProfile": "vector-profile-1"
        }
    ],
    "vectorSearch": {
        "algorithms": [
            {
                "name": "hnsw-1",
                "kind": "hnsw",
                "hnswParameters": {
                    "m": 4,
                    "efConstruction": 400,
                    "efSearch": 500,
                    "metric": "cosine"
                }
            }
        ],
        "profiles": [
            {
                "name": "vector-profile-1",
                "algorithm": "hnsw-1"
            }
        ]
    }
}

Load vector data for indexing

Content that you provide for indexing must conform to the index schema and include a unique string value for the document key. Prevectorized data is loaded into one or more vector fields, which can coexist with other fields containing alphanumeric content.

You can use either push or pull methodologies for data ingestion.

Use Documents - Index to load vector and nonvector data into an index. The push APIs for indexing are identical across all stable and preview versions. Use any of the following APIs to load documents:

POST https://{{search-service-name}}.search.azure.cn/indexes/{{index-name}}/docs/index?api-version=2024-07-01

{
    "value": [
        {
            "id": "1",
            "title": "Azure App Service",
            "content": "Azure App Service is a fully managed platform for building, deploying, and scaling web apps. You can host web apps, mobile app backends, and RESTful APIs. It supports a variety of programming languages and frameworks, such as .NET, Java, Node.js, Python, and PHP. The service offers built-in auto-scaling and load balancing capabilities. It also provides integration with other Azure services, such as Azure DevOps, GitHub, and Bitbucket.",
            "category": "Web",
            "titleVector": [
                -0.02250031754374504,
                 . . . 
                        ],
            "contentVector": [
                -0.024740582332015038,
                 . . .
            ],
            "@search.action": "upload"
        },
        {
            "id": "2",
            "title": "Azure Functions",
            "content": "Azure Functions is a serverless compute service that enables you to run code on-demand without having to manage infrastructure. It allows you to build and deploy event-driven applications that automatically scale with your workload. Functions support various languages, including C#, F#, Node.js, Python, and Java. It offers a variety of triggers and bindings to integrate with other Azure services and external services. You only pay for the compute time you consume.",
            "category": "Compute",
            "titleVector": [
                -0.020159931853413582,
                . . .
            ],
            "contentVector": [
                -0.02780858241021633,
                 . . .
            ],
            "@search.action": "upload"
        }
        . . .
    ]
}

Check your index for vector content

For validation purposes, you can query the index using Search Explorer in the Azure portal or a REST API call. Because Azure AI Search can't convert a vector to human-readable text, try to return fields from the same document that provide evidence of the match. For example, if the vector query targets the "titleVector" field, you could select "title" for the search results.

Fields must be attributed as "retrievable" to be included in the results.

  • Review the indexes in Search management > Indexes to view index size all-up and vector index size. A positive vector index size indicates vectors are present.

  • Use Search Explorer to query an index. Search Explorer has two views: Query view (default) and JSON view.

    • Set Query options > Hide vector values in search results for more readable results.

    • Use the JSON view for vector queries. You can either paste in a JSON definition of the vector query you want to execute, or use the built-in text-to-vector or image-to-vector conversion if your index has a vectorizer assignment.

    • Use the default Query view for a quick confirmation that the index contains vectors. The query view is for full text search. Although you can't use it for vector queries, you can send an empty search (search=*) to check for content. The content of all fields, including vector fields, is returned as plain text.

    • See Create a vector query for more details.

Update a vector store

To update a vector store, modify the schema and if necessary, reload documents to populate new fields. APIs for schema updates include Create or Update Index (REST), CreateOrUpdateIndex in the Azure SDK for .NET, create_or_update_index in the Azure SDK for Python, and similar methods in other Azure SDKs.

The standard guidance for updating an index is covered in Update or rebuild an index.

Key points include:

  • Drop and rebuild is often required for updates to and deletion of existing fields.

  • However, you can update an existing schema with the following modifications, with no rebuild required:

    • Add new fields to a fields collection.
    • Add new vector configurations, assigned to new fields but not existing fields that have already been vectorized.
    • Change "retrievable" (values are true or false) on an existing field. Vector fields must be searchable and retrievable, but if you want to disable access to a vector field in situations where drop and rebuild isn't feasible, you can set retrievable to false.

Next steps

As a next step, we recommend Query vector data in a search index.

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

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