Handle empty input inference

[Samiul-TheSoccerFan]

This PR skips embedding generation when a sematic_text field is empty or with whitespace only.

PUT _inference/sparse_embedding/my-elser-model
{
  "service": "elser",
  "service_settings": {
    "num_allocations": 1,
    "num_threads": 1
  },
  "task_settings": {}
}

PUT index1
{
    "mappings": {
        "properties": {
            "semantic_text_field": {
                "type": "semantic_text",
                "inference_id": "my-elser-model"
            }
        }
    }
}

PUT index1/_doc/doc1
{
  "semantic_text_field": "test value"
}

PUT index1/_doc/doc2
{
  "semantic_text_field": ""
}

PUT index1/_doc/doc3
{
  "semantic_text_field": " "
}

GET index1/_search
{
  "query": {
    "match_all": {}
  },
  "fields": ["_inference_fields"]
}

doc2 and doc3 will return without _inference_fields

Multi field

PUT index3
{
  "mappings": {
    "properties": {
      "city": {
        "type": "semantic_text",
        "inference_id": "my-elser-model",
        "fields": {
          "sparse": {
            "type":  "semantic_text",
            "inference_id": "my-elser-model"
          }
        }
      }
    }
  }
}

PUT index3/_doc/doc1
{
  "city": ["new york", "new york fries"]
}

PUT index3/_doc/doc2
{
  "city": ["", " "]
}

GET index3/_search
{
  "query": {
    "match_all": {}
  },
  "fields": ["_inference_fields"]
}

doc1 and doc2 will return without _inference_fields

[Samiul-TheSoccerFan]

@elasticmachine update branch

[Mikep86]

Good start to this. Can you add a highlighter test where empty input is part of multi-chunk input (ex: ["some test data", " ", "now with chunks"])? The highlighter uses the offsets at query time to recreate the chunks, so this is a good test that the chunk offsets are valid end-to-end.

[Samiul-TheSoccerFan]

@elasticmachine update branch

[elasticsearchmachine]

Pinging @elastic/search-eng (Team:SearchOrg)

[elasticsearchmachine]

Pinging @elastic/search-relevance (Team:Search - Relevance)

[elasticsearchmachine]

Hi @Samiul-TheSoccerFan, I've created a changelog YAML for you.

[kderusso]

Nice work so far Samiul!

[kderusso]

Nitpick: We usually put the reason as when the fix was introduced e.g. 8.19.

[Samiul-TheSoccerFan]

Do we only mention 8.19 or we should mention 9.1.0 as well?

Skips embedding generation when semantic_text is empty or contains only whitespace, effective from 8.19 and 9.1.0.. How about this one?

[kderusso]

That's perfect!

[Mikep86]

Great iterations on this 🙌 ! This is coming along nicely. There's a bug when using the legacy format, which you can replicate with:

PUT test-index
{
  "settings": {
    "index.mapping.semantic_text.use_legacy_format": true
  },
  "mappings": {
    "properties": {
      "inference": {
        "type": "semantic_text"
      }
    }
  }
}

POST test-index/_doc/1
{
  "inference": "   "
}

GET test-index/_doc/1

You get a parsing error, when you should get:

{
    "_index": "test-index",
    "_id": "1",
    "_version": 1,
    "_seq_no": 0,
    "_primary_term": 1,
    "found": true,
    "_source": {
        "inference": {
            "text": "   ",
            "inference": {
                "inference_id": ".elser-2-elasticsearch",
                "model_settings": null,
                "chunks": []
            }
        }
    }
}

[Samiul-TheSoccerFan]

@elasticmachine update branch

[kderusso]

LGTM, nice work!

[Mikep86]

Great work, this is very close. Just a few things to clean up before this is good to merge :)

[Mikep86]

LGTM, thanks for the iterations!

[elasticsearchmachine]

💚 Backport successful

Status	Branch	Result
✅	8.x