Docs

Torch Object Storage

TOS Vectors

Overview

TOS Vectors is a new storage service designed for the AI-native era. With TOS Vectors, you can easily store, query, and manage vector data. Its semantic understanding and similarity search features transform traditional object storage, elevating it from merely "saving data" to actively "understanding data". Leveraging TOS's high reliability and elastic scalability, TOS Vectors offers a cost-effective vector storage for applications such as intelligent search, retrieval-augmented generation (RAG), and AI agents.

Core concepts

TOS Vectors consists of the following key components that work together:

Vector buckets: A new type of bucket designed for storing and querying vectors.

Vector indexes: You can organize your vector data within a vector bucket using vector indexes. You can also perform similarity queries on your vector data within vector indexes.

Vectors: The fundamental elements stored in a vector bucket. They are arrays of floating-point numbers that represent the semantic relationships between text, images, and audio. Similarity searches are performed based on semantic meaning during queries. You can also attach metadata to vectors to filter query results based on conditions such as timestamps and types.

Benefits

New architecture: TOS Vectors is built on a serverless architecture. This architecture addresses the high costs and poor scalability of traditional storage-compute architectures when handling massive amounts of infrequently accessed AI data.

Extremely low cost: TOS Vectors employs a simple, transparent billing model. It only charges for uploaded data, used storage capacity, and scanned data during retrieval. This significantly reduces the high computing and maintenance costs associated with traditional databases or standalone retrieval systems, cutting the total cost of storing vectors by over 90%.

Infinite elasticity: TOS Vectors uses a distributed system to store and manage large-scale vector data. Its serverless elastic scaling automatically adjusts the system based on load, ensuring high availability and performance.

Native support for semantic search: With TOS Vectors, you can query vector data within an index to get similarity-based results. You can filter based on scalar metadata to refine your results precisely. TOS Vectors also supports defining non-filterable metadata, which can be returned alongside search results to provide more detailed information.

Tiered storage: Use TOS Vectors for the long-term, low-cost storage of cold vector data. Use high-performance vector databases, such as VikingDB and Volcano Engine Cloud Search, for hot data. This combination enables tiered storage and on-demand scheduling of hot and cold data, effectively balancing speed, cost, and memory depth.

Easy to use: TOS Vectors provides developers with a dedicated set of APIs and SDKs, enabling tasks such as automation and visualization. Common management operations such as search queries, writing vector data, and batch insertion can be easily performed through the TOS console.

Use cases

Building low-cost RAG applications and knowledge bases

AI business development leads to explosive growth in vector data, which greatly increases storage and retrieval costs. TOS vector buckets provide an affordable foundation for RAG applications with higher tolerance for retrieval latency, including knowledge Q&A, intelligent customer service, and medical image retrieval.

Building AI agents with tiered retrieval

Different AI agents have different performance requirements for retrieval. All vector data can be stored centrally in ultra-low-cost TOS vector buckets for long-term or infrequently accessed memory. For business scenarios requiring high performance and low latency, hot data can be synchronized with a high-performance vector database. This creates an agent application architecture with tiered retrieval.

Building a unified AI content management platform

AI applications generate massive amounts of unstructured content (raw data) and vectorized results (vector data). This can easily lead to fragmented storage and retrieval systems. Storing raw data in TOS standard buckets and vector data in TOS vector buckets allows for the use of a single set of APIs/SDKs to manage and access both the original files and the vector indexes. This setup is ideal for efficient, unified AI content management.

Features

Vector management

Creating vector buckets and indexes.

Writing and deleting vectors in vector indexes.

Adding metadata to vectors.

Vector retrieval and recall (e.g., using cosine similarity or Euclidean distance), and metadata filtering.

Dedicated endpoints

Dedicated public and private network access domains, isolated from standard TOS buckets.

Secure transmission

Encrypted transmission via HTTPS to ensure data security during transfer.

Access control

Resource-based policies (e.g., bucket policies) to precisely control access at the vector bucket or single/multiple vector index level.

How it works

The following diagram illustrates how TOS Vectors works.

Constraints and limitations

Category	Description
Vector buckets	Each account can have up to 100 vector buckets per region.
Vector indexes	During the closed beta period, each vector bucket can hold up to 100 vector indexes.
Vectors	Each vector index can store up to 50 million vectors.
Retrieval performance	200 QPS.
Recall rate	90%.
Average retrieval latency	500 ms.

Billing notes

This feature is currently in a free closed beta and is only available in China (Beijing). Contact your account manager to request a trial.

Last updated: 2026.04.16 16:19:28