Vespa: Can we aggregate on nested fields? - vespa

Question

In search definition the fields inside struct can not have "attribute" indexing.
http://docs.vespa.ai/documentation/reference/search-definitions-reference.html#field_types
Also, struct and maps are not attribute by default.
Resulting search definition would look something like this:
struct nlp {
field token type string {
match: text //can't add indexing here
}
}
field n type nlp {
indexing: summary //can't add attribute here
}
How to add search definition such that we can group by "n.token"? Is it possible to add attribute or indexing for struct fields? Or group by on fields which is not attribute?

Answer 1

Struct field type cannot have attribute which is necessary pre-requisite if you want to run grouping with indexed search, see http://docs.vespa.ai/documentation/reference/search-definitions-reference.html#struct
The only thing you can really do with struct fields with mode=index is to have them as part of the summary (response). You could add a custom searcher which does the aggregation over the struct field analyzing the top K retrieved hits. See http://docs.vespa.ai/documentation/searcher-development.html
With mode=streaming you can run grouping over struct fields, more about streaming here http://docs.vespa.ai/documentation/streaming-search.html

Answer 2

can't comment yet hence posting as answer.
#jkb does this amount to us running a search on some field not in a struct, returning a lot of documents and then synthesising them in a searcher (chaining a searcher would essentially do similar stuff)?
is it also correct to say that nested fields in a document cannot be indexed and in-turn effectively searched (i don't want to use streaming search), and hence the structure must always be flat for index search to work?
i can achieve a flat structure in most cases but what about fields having array of objects where i'd want to search on some attribute of such an object?
"x" : [
object: { attributes},
object: {attributes}
]