[BUG] Nested field missing sub embedding field will cause the IndexOutOfBoundsException

[wdongyu]

What is the bug?

Currently, if a nested field nestedField with two sub field textField and textFieldNotForEmbedding, and a doc like:

{
  "nestedField": [
    {
      "textFieldNotForEmbedding": "This is a text field"
    },
    {
      "textField": "This is another text field"
    }
  ]
}

applying pipeline configuration:

{
  "description": ...,
  "processors": [
    {
      "text_embedding": {
        "model_id": ...,
        "field_map": {
          "nestedField": {
              "textField": "vectorField"
          }
        }
      }
    }
  ]
}

Result:

java.lang.IndexOutOfBoundsException: Index 1 out of bounds for length 1

Expected result:

{
  "nestedField": [
    {
      "textFieldNotForEmbedding": "This is a text field"
    },
    {
      "textField": "This is another text field"
      "vectorField": [...]
    }
  ]
}

How can one reproduce the bug?

Change the mock doc in test code:

// textField --> textFieldNotForEmbedding

nestedObj1.put("textFieldNotForEmbedding", "This is a text field") 

In unit test, get the nlpResult from textEmbeddingProcessor.buildNLPResult, then set it to each field in original doc just like the embedding processor actually does:

Map<String, Object> nlpResult = textEmbeddingProcessor.buildNLPResult(
            knnMap,
            modelTensorList,
            ingestDocument.getSourceAndMetadata()
        );
nlpResult.forEach(ingestDocument::setFieldValue);

And run the unit test with:

./gradlew ':test' --tests "org.opensearch.neuralsearch.processor.TextEmbeddingProcessorTests.testBuildVectorOutput_withNestedList_successful"

What is the expected behavior?

Mentioned above.

What is your host/environment?

Operating system, version.

Do you have any screenshots?

If applicable, add screenshots to help explain your problem.

Do you have any additional context?

Add any other context about the problem.

[wdongyu]

When setting fields in IngestDocument, the processor should check if the textField is not empty :

// build nlp output for list of nested objects
for (Map<String, Object> nestedElement : (List<Map<String, Object>>) sourceAndMetadataMap.get(processorKey)) {
     // Adding non empty check for inputNestedMapEntry, Only fill in embedding when value is not null
     // if (inputNestedMapEntry.getValue().get(index) != null)
     nestedElement.put(inputNestedMapEntry.getKey(), results.get(indexWrapper.index++));
}

willing to open a pr if confirmed.

[vibrantvarun]

@martin-gaievski I think it is a bug and needs to be fixed. Can you once confirm?

[martin-gaievski]

yup, looks like it, we need a fix for it. @wdongyu is this on latest main and 2.x, or specific to a certain 2.x version?

[wdongyu]

@martin-gaievski I test on latest main, it should exist in all versions