Added dot product everywhere were cosine similarity was used

[joancf]

Related Issue

Support for inner product similarity measures

• Support for inner product similarity measures (dot product) #265

Changes

almost everywhere (except some tests) where there was a cosine reference a parallel dot function/option is added

What changed?

most files where cosine was used

Testing and Validation

Not done , needs to be done, it is pending.
I'll try to generate and replace the plugin in my ES isntallation and check it.
but not sure which prodedures I must follow.

[alexklibisz]

Thanks for the contribution. I left some initial comments.

Here are some other items that don't quite fit as a comment:

• what happens when the vectors are not normalized? I don't want to return a negative score to Elasticsearch. I'm not sure how that would behave. I would rather detect this failure state and return an explicit exception. I'm not totally sure where we should detect it. I'm thinking maybe add an if-statement here and here. If the similarity is < 0, we throw an ElastiknnRuntimeException.
• We should add some regression tests to the RecallSuite, similar to these.

[alexklibisz]

I would rewrite this part slightly:

Cosine similarity is defined over $$[-1, 1]$$.
Dot similarity is defined over $$[-1, 1]$$, and if vectors have a magnitude of 1, then it's equivalent to cosine similarity.
Elasticsearch does not allow negative scores.
To work around this, Elastiknn applies simple transformations to produce L1, L2, Cosine, and Dot scores in accordance with the Elasticsearch requirements.

[alexklibisz]

We should make sure to catch this and return an error in the plugin.

[joancf]

i added max(0,distance) so, it will always be positive.

[alexklibisz]

General comment for this file: Let's also add a block of tests below similar to this:

elastiknn/elastiknn-api4s/src/test/scala/com/klibisz/elastiknn/api/XContentCodecSuite.scala

Lines 371 to 406 in 39fa610

	"CosineLsh" - {
	"roundtrip" in {
	for {
	_ <- 1 to 100
	(dims, l, k) = (rng.nextInt(), rng.nextInt(), rng.nextInt())
	mapping = Mapping.CosineLsh(dims, l, k)
	expected = Json.obj(
	"type" -> "elastiknn_dense_float_vector".asJson,
	"elastiknn" -> Json.obj(
	"model" -> "lsh".asJson,
	"dims" -> dims.asJson,
	"similarity" -> "cosine".asJson,
	"L" -> l.asJson,
	"k" -> k.asJson
	)
	)
	} {
	roundtrip[Mapping](expected, mapping)
	}
	}
	"errors" in {
	val ex1 = intercept[XContentParseException](decodeUnsafeFromString[Mapping]("""
	|{
	| "type": "elastiknn_dense_float_vector",
	| "elastiknn": {
	| "model": "lsh",
	| "dims": 33,
	| "similarity": "cosine",
	| "L": "33",
	| "k": 3
	| }
	|}
	|""".stripMargin))
	ex1.getMessage shouldBe "Expected [L] to be one of [VALUE_NUMBER] but found [VALUE_STRING]"
	}
	}

[joancf]

Hi @alexklibisz
First thanks for fast response, and the plug-in itself!!
let me apologize , for sending the pull request before deeply testing it. It's my first time doing things in Scala, and I'm a bit confused on how to do some things.
Finally i could compile and build the zip in my side, so I can try to run the plugin! (and check if it works and the performance)
But as I said my knowledge of scala is my knowledge of Java. ... and for some of the things you ask me, i'm not able to do them. (basically testing is where I think it will take a me a while to understand everything!)

About the changes you asked. I did all of them .
One I did in a different way was to ensure that the similarity was returning a positive value with max(0,1+dotProduct)
In this way we don't raise an exception and negative values will have a 0 similartiy

Thanks
Joan

[alexklibisz]

Thanks for the quick fixes! I can help with the tests once we're happy with the implementation.

I'll have to think about more about the strategy for accounting for non-normalized vectors. As a user I think I would rather get a failed response than a bunch of missing results. Also, we say that the similarity is in [0, 2], but AFAICT the current strategy would allow for scores > 2, wouldn't it?

[alexklibisz]

I think there might be something wrong with the table formatting:

Also, if we end up using it, we should describe the updated transformation here: max(0, 1 + dot product)

[alexklibisz]

Hi @joancf can you try adding the exceptions for scores outside [0, 2]? If you're having trouble I can try this, but probably not til the weekend or next week.

[joancf]

hi. @alexklibisz my company doesn't want to use it. So, I will finish it out-hours
I'll do my best with exceptions and testing.

[joancf]

@alexklibisz Now I'm back with this code, I wanted to extract some logs,
let me see If I can finish it

[joancf]

Hi, @alexklibisz ,
I added some tests (i normalized the vectors for dot product) and also I rebased the branch to main
I can't compile it, because there is an error in
[E008] Not Found Error: elastiknn-plugin/src/main/scala/com/klibisz/elastiknn/mapper/VectorMapper.scala:119:83
[error] 119 | new VectorMapper.FieldType(CONTENT_TYPE, context.buildFullName(thisBuilder.leafName), mapping),

maybe this a problem of the version of Elasticsearch I'm used ( I'm using 12.2 instead of 13) but I need to work on that version and I don't want to upgrade my environment.

[alexklibisz]

Hi, @alexklibisz , I added some tests (i normalized the vectors for dot product) and also I rebased the branch to main I can't compile it, because there is an error in [E008] Not Found Error: elastiknn-plugin/src/main/scala/com/klibisz/elastiknn/mapper/VectorMapper.scala:119:83 [error] 119 | new VectorMapper.FieldType(CONTENT_TYPE, context.buildFullName(thisBuilder.leafName), mapping),

maybe this a problem of the version of Elasticsearch I'm used ( I'm using 12.2 instead of 13) but I need to work on that version and I don't want to upgrade my environment.

In order to get this merged, it needs to be working with the latest version. If that doesn't work, you're welcome to fork and build your own artifacts for older versions. I'm not maintaining multiple/older versions of Elasticsearch.

[joancf]

ok,
Alex I synchronized my repo, Now I can compile and generate the plugin 8.15.0.0