Configure fused vectors
In scenarios such as e-commerce and multimodal search, vectorizing a single field (such as a product title or an image) often cannot satisfy both semantic intent matching and visual style matching at the same time. The fused vector feature lets you select multiple fields (such as a text field and an image field) and encode the content of these fields into a single vector value by using a multimodal model, thereby enabling cross-modal retrieval capabilities such as "search images by text" and "search text by image".
Usage notes
The fused vector feature is supported only by engine version 1.5.8 and later. If your current version is earlier, contact technical support to request an engine upgrade.
Supported models
The fused vector feature calls multimodal model services to encode field content into vectors. Different models impose different limits on the number and type of input fields:
|
Model source |
Model name |
Output dimensions |
Input limits |
|
Model Studio |
qwen3-vl-embedding |
Supported dimensions:
|
A single request supports up to 20 content inputs, of which no more than 5 can be images |
|
AI Search Open Platform |
ops-gme-qwen2-vl-2b-instruct |
1,536 dimensions |
A single request supports 1 image and 1 text as input content |
|
ops-mm-embedding-v1-2b |
1,536 dimensions |
A single request supports 1 image and 1 text as input content |
|
|
ops-mm-embedding-v1-7b |
3,584 dimensions |
A single request supports 1 image and 1 text as input content |
|
|
ops-mm-embedding-ecom-001 |
128 dimensions |
A single request supports 1 image and 1 text as input content |
Configure fused vectors
Fused vectors are configured in the Add a tableField Configuration step.
-
Go to the Field Configuration page and add the fields that participate in the fused vector to the field list. A fused vector requires at least 2 participating fields.
If a participating field involves image content, you must first configure an image data source for that field. In the Data Source column of the field list, select the corresponding data type (text or image) and configure the data source based on how the data is stored. The image type supports the following data sources: URL address, Base64 encoding, OSS object storage, and DLF-Object Table.
-
Click the Data Processing button at the bottom of the page, and select Fused vector from the Data processing template drop-down list.
-
In the Participating fields section, select the fields to be encoded into the fused vector (at least 2).
-
Configure the model service in Services:
-
Model: Select a multimodal model from the drop-down list. The default model is qwen3-vl-embedding on Model Studio.
-
Field Name: The system automatically generates the output field name of the fused vector (such as
_fused_vector). You do not need to modify it manually.
-
-
In the API Key field, enter the API key of the corresponding model service.
-
Read and select the risk notice and billing notice.
-
Click OK to complete the fused vector configuration.
After the configuration is complete, the system automatically fills the generated fused vector field back into the field list. The filled-back field has the following characteristics:
-
The Data Source column displays the names of the source fields that participate in the fused vector encoding.
-
The Field Type column carries a Fused tag, indicating that the field has been processed as a fused vector.
-
The Actions column supports the Edit operation, which reopens the data processing configuration panel and displays the configured information.
Other data processing templates
In addition to fused vectors, the data processing feature supports the following templates:
|
Template name |
Description |
|
Dense vectorization |
Converts text fields into dense vectors |
|
Dense + sparse vectorization |
Generates both dense vectors and sparse vectors for hybrid retrieval |
|
Image vectorization |
Converts image fields into vectors |
|
Image subject detection + image vectorization |
First performs subject detection on an image, and then vectorizes the detected subject regions. The subject detection service supports the following AI Search Open Platform models: general object detection (ops-object-detect-001), face detection (ops-object-detect-face-001), and apparel object detection (ops-object-detect-fashion-001). Supported since engine version 1.5.13. For details such as applicable scenarios and parameter requirements, see Best practices for multi-subject detection. |
|
Image content parsing |
Parses the text and content information in an image |
|
Image content parsing + image vectorization |
First parses the image content, and then performs vectorization |
|
Video processing |
Processes video field data |
Field data source configuration
On the field configuration page, Data Source is displayed as a separate column.
-
Default fields of a scenario template: The data type cannot be modified and retains the configuration preset by the template.
-
Fields added by users: You can select text or image as the data type.
Data sources supported by the Image type: URL address, Base64 encoding, OSS object storage, and DLF-Object Table.