Reranking with cross encoder and knn search

General Feedback
,
1

Hi all, I configured a reranking search pipeline to be used with knn search.

What I want to achieve is to first find 15 most relevant elements with knn search, and further rerank them to return the final 5 elements.

The query I used is:

response = client.search(
    index="test_index",
    search_pipeline="rerank_pipeline",  
    body={
        "query": {
            "knn": {
                "embedding": {
                    "vector": vector,  
                    "k": 15
                }
            }
        },
        "ext": {
            "rerank": {
                "query_context": {
                    "query_text": "architecture generation" 
                }
            }
        },
        "size": 5,
    }
)

The query_text “architecture generation” is a keyword that I identified from the last element if I only do knn search with k=15. Normally, with my query, it should rerank the last element to the first one.

But it turns out the elements that I got with my query are the same as I would get with

response1 = client.search(index= 'test_index',
    body= {
        "size": 5,
        "query": {
            "knn": {
                'embedding': {
                    "vector": vector,
                    "k": 5,
                }
            }
        }})

but reordered. And the element with “architecture generation” that I expect to see is not here.

What my query did is to retrieve 5 elements and rerank them, even though I specified k=15 for knn.

Am I missing something?

An alternative way is to use both k=15 and size=15 then post-process to keep the top 5. But I wonder if it is possible to achieve it without post-processing.

Thanks in advance.