[BUG] Invalid JSON in payload error despite sending a valid JSON with all required parameters

[NeuralFlux]

What is the bug?
Sending a predict request to a model that uses SageMaker connector like so

POST /_plugins/_ml/models/_6gdD40BZqSAbrEiV6DT/_predict
{
  "parameters": {
    "inputs": "test sentence"
  }
}

produces an error

{
  "error": {
    "root_cause": [
      {
        "type": "illegal_argument_exception",
        "reason": "Invalid JSON in payload"
      }
    ],
    "type": "illegal_argument_exception",
    "reason": "Invalid JSON in payload"
  },
  "status": 400
}

despite having a complete and valid JSON.

How can one reproduce the bug?
Steps to reproduce the behavior:

1. Setup a connector to a SageMaker endpoint (any endpoint is okay since the error is in the connector itself)
2. Deploy the model and note the model ID
3. Predict curl -XPOST "http://localhost:9200/_plugins/_ml/models/<Model ID>/_predict" -H 'Content-Type: application/json' -d' { "parameters": { "inputs": "test sentence" } }'
4. See error

What is the expected behavior?
The request should trigger processing of request_body in the connector without any errors.

What is your host/environment?

• AWS OpenSearch Service 2.11
• ML Commons

Do you have any additional context?
Passing "inputs": ["test sentence"] works. However, I need the embedding on just the sentence without the extra square brackets. Moreover, ${parameters.input}[0] works without any errors but gave different embedding for a test sentence than what was expected.

[ramda1234786]

Hi @NeuralFlux works well for me. May be some problem in request_body while creating connector

It should be

"request_body": "{ \"inputs\": \"${parameters.inputs}\" }"

[NeuralFlux]

I realized a JSON has different standards for being "valid". May I know which standard is used to check the payload? Plain strings are valid JSON documents according to RFC 7159 and RFC 8259.

[b4sjoo]

Hi @NeuralFlux, could you please share your connector configuration?

[NeuralFlux]

Sure thing, it's

"actions": [
      {
         "action_type": "predict",
         "method": "POST",
         "headers": {
            "content-type": "application/x-text"
         },
         "url": "<INFERENCE_ENDPOINT>",
         "request_body": "${parameters.inputs}"
      }
   ]

[mingshl]

this might be resolved using model interface #2354