Object Storage Service:OSS Vectors

Key concepts

• Vector bucket: A new bucket type that serves as a cloud resource for managing large-scale vector data.

• Vector index: You can create vector indexes in a vector bucket. A vector index is an index table for storing vector data. You can create multiple vector indexes in a single vector bucket to organize data for different business needs. When you initiate a retrieval query, the query returns results based on the similarity of the vector data within the specified vector index.

• Index mode: A vector index can use one of two modes: Standard Mode or Fusion Mode. The two modes differ in index structure, schema capabilities, and retrieval methods. You must specify the mode when you create a vector index. You can create vector indexes of different modes within the same vector bucket.

• Vector data: High-dimensional numerical arrays generated from unstructured data, such as images, videos, and documents, using vector embedding models. These arrays represent the content features of the data. Vector retrieval returns results based on the similarity between these vectors. You can use any vectorization service, such as ECS, PAI, or Model Studio, to generate vectors and then write them to a specified vector index by using the OSS API, SDKs, or the ossutil tool. When writing data, you can also attach metadata for subsequent scalar filtering.

• Scalar metadata: When you write vector data, you can also attach scalar metadata for subsequent scalar filtering.

Vector index modes

OSS Vectors offers two index modes to meet different requirements for cost, performance, and retrieval complexity. You can create indexes of both modes within the same vector bucket.

Benefits

• Low cost: Vector data has become essential infrastructure for various AI applications and is growing exponentially. OSS Vectors features a simple, transparent billing model that charges for only two items: vector data storage and the amount of data scanned during retrieval. This can reduce costs by more than 90% compared with traditional methods.

• Large scale: OSS Vectors is architected to store and manage massive volumes of vector data. It is built on a serverless framework that automatically handles scaling, eliminating the need for capacity planning.

• Easy to use: OSS Vectors provides a comprehensive set of APIs, SDKs, and the ossutil command-line tool. You can also manage vector data in the OSS console with operations such as retrieval, insertion, and bulk insertion.

• Unified management: You can manage vector buckets and standard buckets that store raw data in a consistent manner. For example, you can apply the same Bucket Policy for access control or configure a common log export path for auditing.

• Semantic retrieval: You can use the QueryVectors operation to query vector data in an index and receive results sorted by similarity. OSS Vectors also supports filtering based on scalar metadata. You can attach metadata when writing vector data to a vector bucket for post-filtering. When you create a vector index, you can also define non-filterable metadata. This type of metadata cannot be used for filtering but is returned with the retrieval results to provide descriptive information for the vector.

• Hybrid search (Fusion Mode): Fusion Mode lets you configure multiple vector fields and various scalar field types in a single index and retrieve them concurrently through multiple retrievers. It then fuses and ranks the results from all retrievers using RRF or Weighted Rank. This meets the needs of high-precision retrieval in multi-modal scenarios with complex conditions.

Use cases

Build low-cost RAG applications

As AI services expand, the exponential growth of vector data increases the pressure on storage and retrieval costs. For multi-modal retrieval scenarios like knowledge bases, AI assistants, and medical image retrieval, users are becoming more tolerant of retrieval latencies in the range of tens to hundreds of milliseconds. In these cases, using OSS Vectors Standard Mode as the storage foundation for your RAG applications can meet business requirements at an extremely low cost.

Build AI agents with tiered retrieval

Different AI agents have varying retrieval performance needs. You can centralize all your vector data in a cost-effective OSS vector bucket in Standard Mode. For business scenarios that demand high performance and low latency, you can store data in an OSS vector bucket in Fusion Mode. This allows you to build an AI agent application architecture with tiered retrieval.

Build a unified AI content platform

AI applications generate massive amounts of unstructured content (such as user-generated content, internal documents, and AI-generated content) and their corresponding vectors, which can lead to fragmented storage and retrieval systems. By storing raw data in standard OSS buckets and vector data in vector buckets, you can build an efficient AI data management platform for use cases like AIGC data management. A single set of APIs and SDKs allows you to manage both raw files and vector indexes, simplifying the creation of a unified AI content platform.

Semantic and multi-modal hybrid search

For online scenarios that demand higher recall and lower latency, such as e-commerce product search, video and image content understanding, and enterprise knowledge base Q&A, you can use Fusion Mode. This mode supports multi-column vectors, full-text search, and hybrid ranking across multiple retrievers. It flexibly combines multiple retrieval signals, including vector similarity, text keywords, scalar filtering, and geospatial data, to significantly improve retrieval results for complex queries.

Enterprise features

Endpoint access

OSS Vectors provides separate public and internal endpoints that are isolated from standard OSS buckets.

• Public endpoint: $bucketname-$uid.regionID.oss-vectors.aliyuncs.com

• Internal endpoint: $bucketname-$uid.regionID-internal.oss-vectors.aliyuncs.com

Note: You must use a third-level domain for all operations except ListVectorBuckets.

Secure transfer

HTTPS encrypts all data in transit.

Access control

• Bucket Policy: Supports resource-based authorization policies that allow you to control permissions at the vector bucket level or for one or more vector indexes.

• RAM Policy: Supports identity-based RAM authorization policies for fine-grained permission control over vector buckets, vector indexes, and data operations. These policies also support cross-account access authorization.

Logging

• Access log export: Supports exporting access logs to a specified bucket in real time or near-real time.

• Unified log format: The log format is fully compatible with standard OSS logs. It includes an additional BucketARN field to uniquely identify the vector bucket resource, which simplifies unified log analysis.

S3 Vectors protocol compatibility

Vector buckets are compatible with the Amazon S3 Vectors protocol. You can use the s3vectors client in the AWS SDK to access a vector bucket without rewriting your existing S3 Vectors application. You only need to point the endpoint to the access domain name of the vector bucket. For the list of compatible operations and the differences from S3 Vectors, see Compatibility with Amazon S3.

Quotas and limits

OSS Vectors has certain quotas and limits. When designing and implementing your vector storage and retrieval solution, plan your number of buckets, index scale, metadata structure, and API call strategy based on the following limits. To request an increase for any of the following quotas, contact Technical Support.

• A single Alibaba Cloud account can create a maximum of 100 vector buckets in a region.

• A single vector bucket can contain a maximum of 100 vector indexes.

Standard Mode limits

• A single vector index can store a maximum of 2 billion rows of vector data.

• Vector dimensions: 1 to 4,096.

• TopK range for vector retrieval requests: 1 to 500 by default.

• Maximum total size of metadata (filterable and non-filterable) for a single vector: 200 KB.

• Maximum size of a single filterable metadata field for a single vector: 2 KB.

• Maximum number of metadata fields (filterable and non-filterable) for a single vector: 100.

• Scalar metadata supports four data types: String, Number, Boolean, and List.

• When you use metadata to filter vectors:

  • The cumulative length of filterable metadata in a single filter instruction cannot exceed 20 KB.

  • The number of filterable metadata items in a single filter instruction cannot exceed 1,024.

  • Filter conditions support a maximum of 8 nested levels.

• When you use the QueryVectors or ListVectors operation for scalar filtering, the following operators are supported:

  • Equality: $eq, $ne

  • Range: $gt, $gte, $lt, $lte

  • Inclusion: $in, $nin

  • Existence: $exists

  • Logical: $and, $or, $nor

• For a single bucket, the PutVectorIndex operation is limited to 5 calls per second.

• For a single index, the PutVectors operation has a QPS (Queries Per Second) limit of 1,000. A single concurrent request can write up to 500 entries per batch. The total throughput across all concurrent requests cannot exceed 2,500 entries per second. For example, if each concurrent request writes 100 entries in a batch, the maximum supported write QPS is 25 (2,500 / 100).

• The maximum request body size for the PutVectors operation is 20 MB.

• For a single index, the QPS limit for the QueryVectors operation is 100.

  • Note: Retrieval QPS depends on factors such as the number of vectors in a single index, vector dimensions, the TopK value, and the scalar metadata conditions in the query. In a typical setup with 10 million rows of 1,024-dimensional vectors and a TopK value of 100, the query QPS can usually reach 100. As the index size or the TopK value increases, the actual retrieval QPS may decrease. OSS Vectors does not guarantee that the QPS will reach 100. Test with your workload to confirm actual performance.

• The maximum response size for the QueryVectors operation is 8 MB.

• The ListVectorIndexes operation returns a maximum of 500 indexes per page. You can use pagination to retrieve the next batch of indexes.

• The maximum concurrency for the ListVectors operation is 16.

Fusion Mode limits

• Field names must be unique within a single index and 1 to 63 characters long. They can contain uppercase and lowercase letters, digits, and underscores (_), and must start with a letter.

• Field count limits: no more than 100 fields in total, no more than 3 vector fields, and no more than 8 tokenized fields. Other scalar fields use the remaining quota. At least one field of the vector type is required.

• Scalar fields support the following types: string, long, double, bool, ip, and geoPoint.

• For numeric and IP fields (double, long, ip), the isArray flag is false by default. When array mode is enabled, a single array can contain a maximum of 128 elements.

• For string fields, when array mode is enabled, a single array can contain a maximum of 512 elements, each string element can be up to 4 KB, and all elements in the array can total up to 64 KB. A custom tokenizer can use the space, |, -, _, and , characters as delimiters. After tokenization, each token must be no larger than 4 KB; the server truncates any excess before indexing. For details about string field limits, see the schema documentation for creating an index.

• The Fusion Mode PutVectors operation has a maximum request body size of 20 MB and a QPS limit of 1,000. A single concurrent request can write up to 500 entries per batch, and the total throughput across all concurrent requests cannot exceed 2,500 entries per second. For example, if each concurrent request writes 100 entries in a batch, the maximum supported write QPS is 25 (2,500 / 100).

• Limits for the Fusion Mode-specific vector retrieval operation (QueryVectorsFusion):

  • This operation supports three query methods:

    • The knn field supports pure vector search and vector search combined with pre-filtering:

      • Supports single-path or multi-path vector queries, with a maximum of 3 KNN queries per request.

      • The topK parameter can be up to 500.

      • The numCandidates parameter can be up to 1,000.

    • The query field supports pure scalar search and full-text search:

      • Each $and, $or, or $nor logical operator supports a maximum of 512 sub-operators and up to 8 levels of nesting.

      • $in and $nin support a maximum of 1,024 elements.

      • The query text for $textMatch and $textMatchPhrase can be up to 512 bytes.

      • $geoPolygon supports a maximum of 16 points.

    • The retriever field supports multi-path hybrid queries with result fusion:

      • When retriever is present, only the following 5 top-level parameters are allowed: indexName, limit, partitionKeys, returnMetadata, and returnMetadataFields. All other parameters (including knn, query, sort, and nextToken) are prohibited.

      • The retrievers array must contain at least 1 and at most 3 elements.

      • For compound retrievers (RRF Retriever and Weight Retriever), windowSize can be up to 500.

  • Other limits for this operation:

    • limit can be up to 500.

    • A single request supports a maximum of 3 sort fields. Score-based sorting supports only descending (desc) order.

    • Pagination with nextToken is allowed only when the request contains only query (scalar search and full-text search). Vector search (knn) and multi-path retrieval (retriever) do not support pagination.

    • partitionKeys supports a maximum of 512 elements, with all elements totaling up to 4 KB.

    • returnMetadata is false by default, meaning no fields are returned. Set it to true to return all fields.

    • returnMetadataFields takes effect only when returnMetadata is true. It cannot return vector-type fields. Setting it to [] is equivalent to returnMetadata=false and returns no fields; setting it to null is equivalent to not specifying the parameter and returns all fields.

    • The maximum response size is 8 MB.

• Fusion Mode is currently in invitational preview and is available only in the Indonesia (Jakarta) region. Some quotas and API limits may differ from Standard Mode. For details, see the API documentation.

Billing

OSS Vectors officially started commercial billing on June 10, 2026 (UTC+8). Please monitor your bills. For more information about the billing method, see Vector billing items. Fusion Mode is currently free during the invitational preview.