Elasticsearch provides a way to return “inner hits” in search queries:

Essentially, that means, for any nested predicate, we’ll assign a name to that predicate, and the SearchResult will contain a separate result set with that name, whose hits are objects matched by the corresponding nested predicate.

See the Elasticsearch documentation for more details.

There are two main challenges:

1. How to represent the inner hits? Most likely, we’ll need either to load entities corresponding to the nested objects, or a way for developers to define a projection on the nested objects.

2. How to implement this efficiently with the Lucene backend? Let’s not look at the Elasticsearch 7.11+ implementation, the license doesn’t allow us to take inspiration from their code. We can look at Elasticsearch 7.10 and below as well as OpenSearch, though.