Add a knn method to elasticsearch_dsl.search.Search

[miguelgrinberg]

Fixes #1679

This PR adds the knn option for k-nearest neighbor vector search.

[pquentin]

This looks good but I did not test locally yet. However, can you please add docs? Unfortunately, the docstring is not enough, you need to edit docs/search_dsl.rst too. Here's the current output: https://elasticsearch-dsl--1691.org.readthedocs.build/en/1691/search_dsl.html

[pquentin]

Looks great! I only have a small nit.

[pquentin]

Looks like boost is missing. It's specified here: https://github.com/elastic/elasticsearch-specification/blob/05c39a9dc46e107e4baaf082a6b56aae1ef5901c/specification/_types/Knn.ts#L26-L43. Documented here: https://www.elastic.co/guide/en/elasticsearch/reference/current/knn-search.html#_search_multiple_knn_fields. But weirdly missing fom https://www.elastic.co/guide/en/elasticsearch/reference/current/search-search.html#search-api-knn.

[miguelgrinberg]

@pquentin I wonder why is boost omitted from our documentation. But in any case, I have added it now.

[pquentin]

Thanks! LGTM.

[israellias]

We need an update for this change since this PR has arrived at ElasticSearch v8.12.0

I'm trying to use "knn" inside the "query" object. But I'm getting the following error:

elasticsearch_dsl.exceptions.UnknownDslObject: DSL class `knn` does not exist in query.

How could I add a new Knn class in the accepted ones? Something like

# elasticsearch-dsl/query.py

class Knn(Query):
    name = "knn"

[miguelgrinberg]

@israellias I have added #1770 to track this request. Thanks!

[israellias]

Thanks @miguelgrinberg
Do you have any workaround?
I'm trying

queryset = queryset.update_from_dict(
    {
        "query": {
            "knn": {
                "field": "embedding",
                "query_vector": res.data[0].embedding,
                "filter": filter_for_knn,
                "num_candidates": 30,
            }
        },
    }
)

But I'm noticing that update_from_dict still calls Q which results in doing the same as

queryset.query("knn", field="embedding", query_vector=res.data[0].embedding, filter=filter_for_knn, num_candidates=30)

and that is throwing the error since the class is simply not listed