Pinned query not good with KNN search

Elastic Stack Elasticsearch
,
1

Hi,

When using a normal knn query this works:

"knn": [
    {
    "inner_hits": {
      "_source": false,
      "fields": [
        "passages.text"
      ]
    },
    "field": "passages.vector.predicted_value",
    "k": 5,
    "num_candidates": 100,
    "query_vector_builder": {
      "text_embedding": {
        "model_id": "elastic__multilingual-e5-small-optimized",
        "model_text": "test text"
      }
    }
  },
  {
    "field": "title.predicted_value",
    "k": 5,
    "num_candidates": 100,
    "query_vector_builder": {
      "text_embedding": {
        "model_id": "elastic__multilingual-e5-small-optimized",
        "model_text": "test text"
      }
    }
  }]

Not only does an array not work, when just using 1 of the above knn queries:


{
 "size": 3,
 "query": {
   "pinned": {
     "ids": [ "1", "2" ],
     "organic": {
       "knn": {
    "inner_hits": {
      "_source": false,
      "fields": [
        "passages.text"
      ]
    },
    "field": "passages.vector.predicted_value",
    "k": 5,
    "num_candidates": 100,
    "query_vector_builder": {
      "text_embedding": {
        "model_id": "elastic__multilingual-e5-small-optimized",
        "model_text": "test query"
      }
    }
  }
     }
   }
 }
}

Returns the following error:

"caused_by": {
      "type": "x_content_parse_exception",
      "reason": "[7:15] [pinned] failed to parse field [organic]",
      "caused_by": {
        "type": "illegal_argument_exception",
        "reason": "[knn] queries cannot be provided directly, use the [knn] body parameter instead"
      }
    }

So currently I have the following issues;

  1. I am not able to do multiple knn queries ( Do knn queries on multiple fields.)
  2. I am not able to use knn queries with text, it seems I need to embed the text first.

Ps. I could not test with embeddings yet as I have problem with my inference pipeline.

Kind regards,
Chenko

2

@Chenko

Before 8.12, knn couldn't be provided as a query object directly. We only supported a top-level knn object.

Here are the docs for 8.12 on using knn in a query: Knn query | Elasticsearch Guide [8.12] | Elastic

In your case, for pinned query to work with knn, you must be at least on 8.12, and then utilize nested for your passage vectors.

3

Hi Ben,

Thanks for the swift response!

I See, this is indeed the case for me. I am currently using 8.11.

As for when I go to 8.12 I would be able to use text instead of first using the inference pipeline?

When you say I must utilize nested for my passage vectors, do you mean I need to do a nested query instead of a normal one with knn?
Could you give an example?

4

It seems like you are using nested vectors (chunked) right?

So, for pinned query and this, it would be something like:

{
    "size": 5,
    "query": {
        "pinned": {
            "ids": [
                "1",
                "2"
            ],
            "organic": {
				"nested": {
					"path": "passages",
					"inner_hits": {
						"_source": false,
						"fields": [
							"passages.text"
						]
					},
					"query": {
						"knn": {
							"field": "passages.vector.predicted_value",
							"query_vector": [...],
							"num_candidates": 100
						}
					}
				}
            }
        }
    }
}

Frustratingly, we didn't add the query vector builder stuff to knn query until later, sorry :(. That isn't until 8.14.

1 Like
5

Thanks for the example!

You are correct, I am indeed using nested vectors in the passages field. I am however also using normal vectors for the title.

I am now eagerly awaiting 8.14. any idea when it is set to release?

6

And the DSL for KNN does not allow passing a text string, only vectors. If you choose to use a trained model that's in the Elasticsearch tool, how can you obtain that vector on a local client?

Considering that Model A does not generate the same vector as Model B.

7

Hi there @Shell_Dias,

I encourage you to start new threads for new topics, so that it doesn't add too much clutter to already resolved topics.

With that said, I also encourage you to check out the query_vector_builder if you're interested in performing kNN search using a text string. You can also upload models using eland if you're not using a built in Elasticsearch model.

1 Like