Add vector functions to Sort Script Context

[mayya-sharipova]

See user filed problem here

[elasticmachine]

Pinging @elastic/es-search

[jimczi]

The code looks good to me but I wonder if we should allow this. The query can already be used to sort by these functions if it is set as the only scoring clause so I am concerned that we add multiple ways to achieve the same. Is there a compelling reason to add this functionality other than consistency ?

[csenol]

@jimczi Vectors are encoded and these are the only functions which makes them possible to use in sorting context.
I would not think vectors only for document/query embedding and using it for document scoring. Cosine similarity / dot products outputs can be used as an input to other sorting mechanism such as personalization, receny and so on with a combination of multi-staged sorting. So I think enabling them in script sorting is a reasonable Idea

[csenol]

@jimczi @mayya-sharipova Any updates on this issue ?

[mayya-sharipova]

@csenol Sorry for a late reply. If you are interested to sort by the output of the vector functions, would it be possible for your use case to usescript_score query and sort results by _score:

GET /my_index/_search
{
    "sort" : [
        "_score",
        "user"
    ],
    "query" : {
        "script_score" :  {...} // your vector function here
    }
}

We have internally redesigned vector functions which would make them difficult to be used in the sort context.

[csenol]

@mayya-sharipova Thanks for reply but I think this limits the usage of vectors almost into query/vector embeddings. There might be different sorting needs where relevancy score is not needed in the query context.

[mayya-sharipova]

@csenol

relevancy score is not needed in the query context.

You don't need to include relevancy scores. Here _score is the output of the script_score which could be just the result of computation from vector functions.

[csenol]

@mayya-sharipova Let's say it is OK for vector functions not to be in score context, but we still miss decay functions in sort context. What about adding them ?

[mayya-sharipova]

@csenol A similar question here: is there a reason why you can't use sort by _score, where _score can include a result from decay functions together with any other calculations you need?

GET /my_index/_search
{
    "sort" : [
        "_score",
        "user"
    ],
    "query" : {
        "script_score" :  {...} // your decay function here combined with any other calculations
    }
}

[csenol]

This is a bit patronizing how sorting should be used. In long term, you can even argue "sort" is useless and everything should be expressed with "script_score"

[csenol]

@mayya-sharipova I updated my cluster to ES 7.5.0 but I am still getting

Caused by: java.lang.IllegalArgumentException: Unknown call [dotProduct] with [2] arguments.
	at org.elasticsearch.painless.PainlessScript$Script.compile(dotProduct(params.query_embed, doc['my_vector']):1) ~[?:?]
	at org.elasticsearch.painless.node.ECallLocal.analyze(ECallLocal.java:125) ~[?:?]
	at org.elasticsearch.painless.node.SExpression.analyze(SExpression.java:62) ~[?:?]
	at org.elasticsearch.painless.node.SClass.analyze(SClass.java:205) ~[?:?]
	at org.elasticsearch.painless.node.SClass.analyze(SClass.java:162) ~[?:?]
	at org.elasticsearch.painless.Compiler.compile(Compiler.java:215) ~[?:?]
	at org.elasticsearch.painless.PainlessScriptEngine$2.run(PainlessScriptEngine.java:356) ~[?:?]
	at org.elasticsearch.painless.PainlessScriptEngine$2.run(PainlessScriptEngine.java:352) ~[?:?]
	at java.security.AccessController.doPrivileged(AccessController.java:389) ~[?:?]
	at org.elasticsearch.painless.PainlessScriptEngine.compile(PainlessScriptEngine.java:352) ~[?:?]