AI VibeFlow overview

Core concepts

• DataPipeline: the core management object in AI VibeFlow. A DataPipeline corresponds to a set of data processing rules that cover both data preprocessing and vectorization, and runs automatically after it is created.

• Source data: defines the source of the raw data to be processed, including the source bucket, object prefix, file types (document, image, and video), and processing mode (existing data, incremental data, or both).

• Processing settings: define the workflow to process, including data preprocessing and vectorization. Processing settings support the following two modes:

  • Processing with AI VibeStudio (Recommended): provides data preprocessing and vectorization. Different processing templates are provided for different types of raw files such as images, videos, and documents, so that raw files can be parsed, preprocessed, and vectorized for various scenarios. This mode does not require a Model Studio API key.

  • Vectorization with Alibaba Cloud Model Studio: provides vectorization only. You can call the vector models of Alibaba Cloud Model Studio directly to vectorize data, with no data preprocessing capabilities. This mode requires a Model Studio API key, a vector model, a video frame capture frequency (available only for multimodal models), and the setting that specifies whether to cascade the deletion of the corresponding vectors when the source object is deleted.

• Destination data: defines the destination vector bucket, vector index, and vector key naming rule to which the vector data is written, and whether to write ObjectTag and UserMeta as scalar metadata. In AI VibeStudio mode, you can specify destination indexes separately for image embedding, image text embedding, video frame embedding, video text embedding, and document chunk embedding, and a single DataPipeline writes to up to five vector indexes.

• Exception handling: defines the processing policy when a single processing task fails (skip and continue, skip and log, or terminate immediately), and lets you specify the bucket and prefix that receive error messages.

Benefits

• End-to-end workflow: massive volumes of raw unstructured data, the data preprocessing and vectorization process, and vector data storage are all handled within OSS in a closed loop, which simplifies the production pipeline.

• Automatic processing: supports both batch vectorization of existing data and automatic vectorization of newly uploaded data, without manual intervention.

• Synchronized data updates: when raw data in the object bucket is modified, overwritten, or deleted, you can choose to synchronously update and delete the corresponding vector data in the vector bucket, which reduces manual maintenance costs.

• Flexible data preprocessing: different processing templates are provided for different types of raw files such as images, videos, and documents, so that raw files can be parsed and preprocessed for various scenarios without building your own parsing and chunking pipeline.

• Flexible vector models: in Model Studio mode, you can directly call the text and multimodal vector models provided by Alibaba Cloud Model Studio (such as text-embedding-v4 and tongyi-embedding-vision-plus) and select the appropriate model and dimensions based on your business needs. In AI VibeStudio mode, the service automatically matches the vision model and vector model to the selected search mode and file types, so you do not need to select a model yourself.

• Transparent costs: in Model Studio mode, vectorization uses your own Model Studio API key and embedding fees are charged directly by Model Studio, while OSS charges only for API requests and vector storage. In AI VibeStudio mode, model fees are measured in AI Credits and charged by OSS. Neither mode incurs additional task management fees.

• Traceable failures: for each failed request, information such as the ErrorCode, ErrorMessage, and RequestId can be saved as an object to a specified bucket, which makes it easier to locate issues and retry.

Scenarios

AI VibeFlow is suitable for all scenarios in which you need to preprocess and vectorize massive volumes of unstructured data in OSS. Typical use cases include:

• Semantic retrieval and RAG: vectorize documents, images, and videos in an OSS object bucket and write the vectors to a vector bucket, and then combine them with large language models to build Retrieval-Augmented Generation (RAG) applications.

• E-commerce image and text search: perform multimodal vectorization of product images and descriptions and write the vectors to a vector bucket to support image-to-image and text-to-image search.

• Advertising and marketing asset management: automatically vectorize marketing images and videos to enable semantic recall of assets and recommendation of similar content.

• Enterprise knowledge base construction: vectorize internal knowledge documents and question bank images and write the vectors to a vector bucket to enable unified semantic retrieval across business lines.

• Multimedia content governance: combine vector similarity with scalar metadata filtering to enable fine-grained management and distribution of image and video assets.

Billing

The fees incurred by AI VibeFlow depend on the mode selected in the processing settings, and the two modes are billed separately. AI VibeFlow does not charge additional task management fees.

Vectorization with Alibaba Cloud Model Studio

This mode incurs the following fees, which are charged separately by OSS and Alibaba Cloud Model Studio:

Processing with AI VibeStudio

This mode incurs the following fees, all of which are charged by OSS:

Processing with AI VibeStudio does not consume a Model Studio API key and does not generate Embedding model call fees in your Model Studio bill.

Limits

• A single account can create up to 1,000 DataPipelines in a single region.

• In Model Studio mode, a DataPipeline corresponds to one vector model and one vector index. In AI VibeStudio mode, a DataPipeline can write to 1 to 5 vector indexes, and the number depends on the selected search mode and file types.

• The source object bucket, the destination vector bucket, and the bucket that receives error messages must be located in the same region.

• The maximum capacity of a single vector index is 2 billion rows. After the index is full, the pipeline automatically enters the Paused state.

• When extracted as vector metadata, you can specify up to 10 ObjectTag keys and up to 8 UserMeta keys.

Permissions

When you create a DataPipeline, you must specify a RAM role that OSS uses to run vectorization tasks in the background. AI VibeFlow uses a dedicated trust policy (datapipeline-oss.aliyuncs.com) and cannot reuse other service roles such as AliyunOSSRole.

{
  "Statement": [
    {
      "Action": "sts:AssumeRole",
      "Effect": "Allow",
      "Resource": "acs:ram::<account-id>:role/<RoleName>"
    }
  ],
  "Version": "1"
}

{
  "Version": "1",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "ram:CreateRole",
        "ram:CreatePolicy",
        "ram:AttachPolicyToRole"
      ],
      "Resource": "*"
    }
  ]
}

{
  "Version": "1",
  "Statement": [
    {
      "Action": [
        "oss:GetObject",
        "oss:GetObjectTagging",
        "oss:GetBucketStat",
        "oss:ListObjects"
      ],
      "Resource": [
        "acs:oss:*:*:<source-bucket-name>",
        "acs:oss:*:*:<source-bucket-name>/*"
      ],
      "Effect": "Allow"
    },
    {
      "Action": [
        "oss:GetVectorIndex",
        "oss:GetVectors",
        "oss:PutVectors",
        "oss:DeleteVectors"
      ],
      "Resource": [
        "acs:ossvector:*:*:<vector-bucket-name>",
        "acs:ossvector:*:*:<vector-bucket-name>/*"
      ],
      "Effect": "Allow"
    },
    {
      "Action": [
        "oss:PutObject",
        "oss:GetBucketStat"
      ],
      "Resource": [
        "acs:oss:*:*:<error-output-bucket-name>",
        "acs:oss:*:*:<error-output-bucket-name>/*"
      ],
      "Effect": "Allow"
    }
  ]
}

{
  "Statement": [
    {
      "Action": "sts:AssumeRole",
      "Effect": "Allow",
      "Principal": {
        "Service": [
          "datapipeline-oss.aliyuncs.com"
        ]
      }
    }
  ],
  "Version": "1"
}

References

• AI VibeFlow getting started

• Use AI VibeFlow to produce vector data

• OSS Vectors overview

• Text and multimodal vectorization with Model Studio