Using two levels of nesting in knn vector search

Elastic Stack Elasticsearch
1

If we have an index with at least two nesting levels (i know it is not optimal in ES, but we need to have it) , and the dense vectors are in the second level. I can get from the results, the top level fields (they are unique by hit) or the low level ones from the inner hits. But I can't find a way to retrieve the fields of the middle level.
Let me set up an example,
This is the mapping:

{
  "mappings": {
    "properties": {
      "title": {
        "type": "text"
      },
      "chapters": {
        "type": "nested",
        "properties": {
          "number": {
            "type": "integer"
          },
          "c_vector": {
            "type": "dense_vector",
            "dims": 3,
            "similarity": "cosine"
          },
          "paragraphs": {
            "type": "nested",
            "properties": {
              "content": {
                "type": "text"
              },
              "p_vector": {
                "type": "dense_vector",
                "dims": 3,
                "similarity": "cosine"
              }
            }
          }
        }
      }
    }
  }
}

So, we have a book with chapters and paragraphs, chapters also have dense vectors, but they are not strictly needed for this topic, but will help to illustrate it.

So, searching for paragraphs, I can do this query:

{
  "knn": {
    "query_vector": [1,2,3],
    "field": "chapters.c_vector",
    "k": 3,
    "num_candidates": 3,
    "inner_hits": {
      "_source": false,
      "fields": [  "chapters.number"   ]
    }
  },
  "fields": ["title","chapters.number"]
  "_source": false
}

In the results,

  • from the HIT, i get the title and the the list of chapters,
  • From the inner hit , i get the top matching chapter
    ok, that's fine

But now, If i want to search paragraphs, the query would be

{
    "knn": {
        "query_vector": [0,2,2],
        "field": "chapters.paragraphs.p_vector",
        "k": 3,
        "num_candidates": 3,
        "inner_hits": {
            "_source": false,
            "fields": ["chapters.paragraphs.content","chapters.number"]
        }
    },
   "fields": ["title","chapters.number"],
    "_source": false
}

So, I can get the title, and the in inner hits I get the paragraph with highest score, but... How can I know to which chapter this paragraph belongs?
in the outer fields, i get all the chapters of the book,
Adding "chapters.number" in the knn -> inner_hits has no effect.

In regular searches , an nested search can be embedded into another nested search, and each may return it's own inner_hits, but I don't think this is possible in knn searches.

Summarizing: When there are multiple levels on nesting, and the knn search is done in the lower levels, how can the middle levels of a hit be obtained?

2

hey @jcodina !

This is an interesting one - inner_hits retrieve the inner hits for just one nested level. Having multiple nested levels involve a different approach, using the knn query instead of the top-level knn section:

{
    "query": {
        "nested": {
            "path": "chapters",
            "query": {
                "nested": {
                    "path": "chapters.paragraphs",
                    "query": {
                        "knn": {
                            "query_vector": [
                                0,
                                2,
                                2
                            ],
                            "field": "chapters.paragraphs.p_vector",
                            "k": 3,
                            "num_candidates": 3
                        }
                    },            
                    "inner_hits": {
                        "fields": ["chapters.paragraphs.content"],
                        "_source": false
                    }
                }
            },            
            "inner_hits": {
                "fields": ["chapters.number"],
                "_source": false
            }
        }
    },
    "_source": false,
    "fields": ["title"]
}

Check that we're using two nested queries, each one retrieving the corresponding inner_hits for each path section.

It is a bit complex to traverse the result, as you'll be dealing with nested inner hits. It might be a good idea to flatten this structure and include the chapter number into the paragraphs information.

Hope that helps!