Support one_to_one in ML Inference Search Response Processor

[mingshl]

Description

Many document to one prediction is the default config in ML Inference Search Response Processor. This PR adds one document to one prediction support following up with #2688

in 2.16, ml inference search response processor support collect documents into a list of model input and make one prediction call.

for example,

the search response returns as following, and we would like to use the text field in the document to send to model input.

{
  "took": 100,
  "timed_out": false,
  "_shards": {
    "total": 22,
    "successful": 22,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 2,
      "relation": "gte"
    },
    "max_score": 1,
    "hits": [
      {
        "_index": "test_index",
        "_id": "1",
        "_score": 1,
        "_source": {
          "text": " this is document 1"
      },
     {
            "_index": "test_index",
            "_id": "2",
            "_score": 1,
            "_source": {
              "text": " this is document 2"
          } 
      ] 
}

one_to_one inference is false setting will combine the two document containing the field text into ["this is document 1", "this is document 2"] and this is one round of prediction.

in 2.17, we support one_to_one inference is true setting, that will make two rounds of prediction, the first prediction, we send "this is document 1" and the second prediction we send "this is document 2".

when users want to use a model that accept a list of model input, for example, using openai embedding model, we can set one_to_one is false,

when users want to use a model that accept a single string model input, for example, using bedrock embedding model, we can set one_to_one is true. The most common use case for one-to-one inference is rerank use case using xgboost, which usually take a single document and compare with the search string and return a single score.

Related Issues

#2173
#2444
#2839

#2879

Check List

• New functionality includes testing.
• New functionality has been documented.
• API changes companion pull request created.
• Commits are signed per the DCO using --signoff.
• Public documentation issue/PR created.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.
For more information on following Developer Certificate of Origin and signing off your commits, please check here.

[mingshl]

We need to bump the version in the build.gradle file to let the CI runs.

> Could not resolve all dependencies for configuration ':opensearch-ml-plugin:runtimeClasspath'.
   > Conflict found for the following module:
       - org.apache.httpcomponents.core5:httpcore5 between versions 5.2.5 and 5.2.2
       - 

[zane-neo]

Why encapsulate the original exception to RuntimeException?

[mingshl]

to avoid 500x error

[dhrubo-os]

Why are we getting 5x error in the first place?

[Zhangxunmt]

Suggest use OpenSearchStatusException to replace RuntimeException.

[mingshl]

I don't see 5xx exception, it's for precautions. changed to OpenSearchStatusException similar to

ml-commons/plugin/src/main/java/org/opensearch/ml/utils/IndexUtils.java

Line 134 in 9663053

throw new OpenSearchStatusException("Failed to search index " + indexName, RestStatus.BAD_REQUEST);

added in commit use OpenSearchStatusException in error handling

[dhrubo-os]

Thanks for the explanation. IMO, one_to_one doesn't not seem intuitive. May be single_document_inference? Or document_wise_inference?

in 2.16, we support one_to_one inference is true setting,

So we already released this one_to_one inference in 2.16?

Where do we save this setting?

[mingshl]

Thanks for the explanation. IMO, one_to_one doesn't not seem intuitive. May be single_document_inference? Or document_wise_inference?

in 2.16, we support one_to_one inference is true setting,

So we already released this one_to_one inference in 2.16?

Where do we save this setting?

aw ha, it should be 2.17, yeah this change is going to add to 2.17. just modified the description.

[mingshl]

Thanks for the explanation. IMO, one_to_one doesn't not seem intuitive. May be single_document_inference? Or document_wise_inference?

in 2.16, we support one_to_one inference is true setting,

So we already released this one_to_one inference in 2.16?

Where do we save this setting?

in the ml search response processor, we already had it in 2.16 but it's not supported to turn on to true yet.

[dhrubo-os]

in the ml search response processor, we already had it in 2.16 but it's not supported to turn on to true yet.

can we change the settings name? or we need to stick with one_to_one?

[mingshl]

in the ml search response processor, we already had it in 2.16 but it's not supported to turn on to true yet.

can we change the settings name? or we need to stick with one_to_one?

the name is added since 2.16, if we change the name in 2.17, it might case confusion.

[mingshl]

All tests passed