Load Your Own ONNX Model

Use the load model operation to load an embedding or reranking model into the database.

The load model operation imports a model from object storage, in ONNX format, for use in embedding generation and reranking operations. Once loaded, the model can be used for integrated table embeddings or for standalone inference.

• Python
• Curl
• PL/SQL

Caution:

After loading an ONNX model using DBMS_VECTOR_DATABASE, drop the model only by using DBMS_VECTOR_DATABASE.DROP_MODEL.

Do not drop a model loaded through DBMS_VECTOR_DATABASE using DBMS_DATA_MINING.DROP_MODEL, DBMS_VECTOR, or another database API. Those APIs can remove the underlying database model without removing the DBMS_VECTOR_DATABASE metadata that references it.

This can leave an invalid model entry. For example, DBMS_VECTOR_DATABASE.LIST_MODELS might return a model whose details are NULL, and subsequent DBMS_VECTOR_DATABASE operations, such as UPSERT_VECTORS, can fail when they reference that model. To remove a model loaded through DBMS_VECTOR_DATABASE, use the following syntax:

BEGIN
  DBMS_VECTOR_DATABASE.DROP_MODEL('<model_name>');
END;
/

See the following for an example of using load_model to load a models from Oracle Object Storage:

• Load an embedding model:

  from oracle_vecdb import OracleVecDB, Configuration
  
  client = OracleVecDB(Configuration(
      rest_url="https://<host>/ords/<schema>/_/db-api/stable/vecdb/",
      access_token="<bearer-token>", # or username="<user>", password="<pass>"
  ))
  
  response = client.load_model(
      model_name='all-MiniLM-L6-v2',
      url='https://objectstorage.us-phoenix-1.oraclecloud.com/n/namespace/b/bucket/o/model.onnx'
  )
  print(response)
  
  #verify that the model was loaded using list_models
  models = client.list_models()
  print([item.model_name for item in models.items or []])

• Load a reranking model:

  from oracle_vecdb import OracleVecDB, Configuration
  
  client = OracleVecDB(Configuration(
      rest_url="https://<host>/ords/<schema>/_/db-api/stable/vecdb/",
      access_token="<bearer-token>", # or username="<user>", password="<pass>"
  ))
  
  response = client.load_model(
      model_name='reranker_model',
      url='https://objectstorage.example.com/models/reranker_quantized.onnx',
      model_params={'metadata': {'function': 'regression'}}
  )
  print(response)
  
  #verify that the model was loaded using list_models
  models = client.list_models()
  print([item.model_name for item in models.items or []])

A JSON response is returned confirming that the model was loaded successfully. An error is raised if the model already exists or if the provided URL is inaccessible.

Example response:

{
    "model_name": "SAMPLE_MODEL",
    "algorithm": "ONNX",
    "mining_function": "EMBEDDING",
    "creation_date": "2026-03-12T11:27:21Z",
    "attributes": [
        {
            "name": "DATA",
            "value": "TEXT",
            "data_type": "VARCHAR2",
            "data_length": 32767
        },
        {
            "name": "ORA$ONNXTARGET",
            "value": "VECTOR",
            "data_type": "VECTOR",
            "data_length": 1593,
            "vector_info": "VECTOR(384,FLOAT32)"
        }
    ]
}

For more information about the load_model operation, see Python API Reference.

See how POST /vecdb/models/ can be used in the following examples:

• Load an embedding model:

  curl -X POST \
    "https://<host>:<port>/ords/<schema>/_/db-api/stable/vecdb/models/" \
    -H "Content-Type: application/json" \
    -H "Accept: application/json" \
    # Choose ONE authentication method:
  
    # Option 1: Basic authentication
    -u "<user>:<password>" \
  
    # Option 2: OAuth Bearer token
    # -H "Authorization: Bearer <access_token>" \
  
    -d '{
      "modelName": "DOC_EMBED_MODEL",
      "url": "https://objectstorage.example.com/models/doc_embed.onnx"
    }'

• Load a reranking model:

  curl -X POST \
    "https://<host>:<port>/ords/<schema>/_/db-api/stable/vecdb/models/" \
    -H "Content-Type: application/json" \
    -H "Accept: application/json" \
    # Choose ONE authentication method:
  
    # Option 1: Basic authentication
    -u "<user>:<password>" \
  
    # Option 2: OAuth Bearer token
    # -H "Authorization: Bearer <access_token>" \
  
    -d '{
      "modelName": "RERANK_MODEL",
      "url": "https://objectstorage.example.com/models/reranker_quantized.onnx",
      "modelParams": {
        "metadata": {
          "function": "regression"
        }
      }
    }'

Responses:

• 201 Created – returns the registered model record with status information.

  {{
    "model_name": "DOC_EMBED_MODEL",
    "algorithm": "ONNX",
    "mining_function": "EMBEDDING",
    "creation_date": "2026-05-01T10:00:00.000000+00:00",
    "attributes": [
      {
        "name": "DATA",
        "value": "TEXT",
        "data_type": "VARCHAR2",
        "data_length": 4000,
        "vector_info": null
      }
    ]
  }

• 400 - the request body included invalid parameters.
• 404 - model not found.

For more information about POST /vecdb/models/, see REST API Reference.

See how DBMS_VECTOR_DATABASE.LOAD_MODEL can be used in the following examples.

• Load an embedding model:

  dbms_vector_database.load_model(
      model_name => 'DOC_EMBED_MODEL',
      url => 'https://objectstorage.example.com/models/doc_embed.onnx'
  );

• Load a reranking model:

  dbms_vector_database.load_model(
      model_name => 'RERANK_MODEL',
      url => 'https://objectstorage.example.com/models/reranker_quantized.onnx',
      model_params => JSON('{"metadata": {"function": "regression"}}')
  );

Example response:

{
  "model_name": "DOC_EMBED_MODEL",
  "algorithm": "ONNX",
  "mining_function": "EMBEDDING",
  "creation_date": "2026-05-01T10:00:00.000000+00:00",
  "attributes": [
    {
      "name": "DATA",
      "value": "TEXT",
      "data_type": "VARCHAR2",
      "data_length": 4000,
      "vector_info": null
    }
  ]
}

For more information about the PL/SQL implementation, including parameters, see LOAD_MODEL.