[ML] Implement JSONPath replacement for Inference API

[jonathan-buttner]

This PR adds a very minimal implementation of something similar to the JSONPath library. This is needed for the custom models PR here: #125679 MapPathExtractor recursively iterates through a provided map navigating to the specified field and extracting the values. It handles nested maps and lists within the map.

This code isn't currently used anywhere outside of the tests that reference it.

Difference from JSONPath

This implementation doesn't support many of the features that JSONPath does. It also deviates from JSONPath in its handling of arrays of maps. When extracting a field from a list of maps, JSONPath will flatten the result into a single array even if multiple arrays needed to be traversed to extract the field from the map. This implementation preserves each array that it encounters. This is important so that we can construct internal classes that represent the various result formats. For example when building the text embedding response we effectively need an array of an array of floats so it's helpful to preserve the outer and inner arrays when constructing the objects after we extract the data from the map.

The second example below depicts the difference.

Schema examples

I tried to keep the schema similar to JSONPath. There's no particular reason we need to do that though. Hopefully it's more familiar to users though.

• $. to start the path
• A . dot is used to traverse nested maps
• [*] to indicate that it's an array

$.field1.some_array[*].another_field
$.some_array[*].field1

Examples

Extracting arrays

{
    "request_id": "B4AB89C8-B135-xxxx-A6F8-2BAB801A2CE4",
    "latency": 38,
    "usage": {
        "token_count": 3072
    },
    "result": {
        "embeddings": [
            {
                "index": 0,
                "embedding": [
                    2,
                    4
                ]
            },
            {
                "index": 1,
                "embedding": [
                    1,
                    2
                ]
            }
        ]
    }
}

MapPathExtractor.extract(map, "$.result.embeddings[*].embedding") returns [[2, 4], [1, 2]]

Extracting multiple map fields

{
  "result": [
    {
      "key": [
        {
          "a": 1.1
        },
        {
          "a": 2.2
        }
      ]
    },
    {
      "key": [
        {
          "a": 3.3
        },
        {
          "a": 4.4
        }
      ]
    }
  ]
}

MapPathExtractor.extract(map, "$.result[*].key[*].a"); returns [[1.1, 2.2], [3.3, 4.4]]

NOTE: JSONPath will return: [1.1, 2.2, 3.3, 4.4]

[elasticsearchmachine]

Pinging @elastic/ml-core (Team:ML)

[prwhelan]

Suggested change

	* This implementation differs in out it handles lists in that JSONPath will flatten inner lists. This implementation
	* This implementation differs in how it handles lists in that JSONPath will flatten inner lists. This implementation

[prwhelan]

Do we have to assert or throw an exception if we don't start with $?

[jonathan-buttner]

Good point I'll add an exception.

[elasticsearchmachine]

💚 Backport successful

Status	Branch	Result
✅	8.x