Custom models - OpenSearch - Alibaba Cloud Documentation Center

OpenSearch Vector Search Edition allows you to call custom models. You can create custom models on the Custom Models tab and perform management operations on the models, such as view, edit, debug, disable, or restore a model.

Create a model

In the left-side navigation pane of the details page of your OpenSearch instance, choose Model Management > Models. On the Custom Models tab of the page that appears, click Create Model. In the Create Custom Model panel, configure the parameters based on the custom model format.

Custom model format

{
  "modelName": "custom_model",
  "modelType": "text_embedding",
  "dimension": 1536,
  "url": "http://xxx/api/v2/services/embeddings",
  "method": "POST"
  "request": {
    "header": {
      "Authorization": "fj42091fjlgnlsvu023nva",
      "Content-Type": "application/json"
    },
    "urlParams": {
      "build": {},
      "search": {}
    },
    "requestBody": "{\"input\": [\"%{input}\"],\"input_type\": \"${input_type}\"}",
    "parameters": {
      "build": {
        "input_type": "document"
      },
      "search": {
        "input_type": "query"
      }
    }
  },
  "response": {
    "embeddings": "$.result.embeddings[*].embedding"
  }
}

Parameters

Parameter	Description
modelName	The name of the model.
modelType	The type of the model. Valid values: text_embedding: dense text model text_sparse_embedding: sparse text model image_embedding: image vector model image_analyze: image content analysis model
dimension	The number of dimensions of the model. You need to configure this parameter if you set the modelType parameter to text_embedding or image_embedding.
url	The endpoint of the model service.
method	The HTTP method used to access the model service. Valid values: PUT and POST.
request	The request method.
request.header	The HTTP header used to access the model service, which is passed to the model service without any processing.
request.urlParams	The parameters in the URL. Some model services need to pass parameters through a URL, such as ?abc=123.
request.urlParams.build	The parameters that are passed when an index is created.
request.urlParams.search	The parameters that are passed during a query.
request.requestBody	The template string of the request body. Template variable format: %{Parameter name}. Example: { "input": "%{input}" }. Built-in variable: %{input}: the input text. %{modal}: the modal. If a multimodal model is used, you can specify the modal to use in a query.
request.parameters	If you want to distinguish between parameters in the build phase and parameters in the search phase in the request body, you can define them in this parameter. build: the parameters in the index creation phase. search: the parameters in the query phase.
response	The format used to parse results. A JSON path is used.

JSON path description

Operators

Operator	Description
$	The root element. A path must start with $.
*	The wildcard that indicates any value. You can use a wildcard to replace a name or number.
.<name>	Used to access the attributes of objects.
[<number>]	Used to access the elements of array subscripts.

Examples

{
  "store": {
    "books": [
      {
        "title": "java",
        "price": 56
      },
      {
        "title": "c++",
        "price": 99
      }
    ]
  }
}

Extract the titles of all books
- json path: $.store.books[*].title
- Result: ["java", "c++"]
Extract the title of the first book
- json path: $.store.books[0].title
- Result: ["java"]

Manage a model

Debug a model
On the Custom Models tab of the Models page, find the model that you want to manage and click Debug in the Actions column to test the effect of the model.
Disable a model
After a model is disabled, the table that invokes the model no longer processes the related data and queries. Proceed with caution.
Delete a model
Only disabled models can be deleted.
Restore a model
You can restore disabled models.

Call a model

When you configure a table, click Configure for the field that you want to preprocess. You can select the created custom model from the displayed models.