Support retrieving matched elements in an array of strings

[danitico]

Is your feature request related to a problem? Please describe.
When we want to search in a array of strings (imagine a list of paragraphs) and setting the summary to dynamic, we receive the whole list of paragraphs where the matched ones have the <hi> tags to highlight the set of words queried. If this list of paragraphs is small , it won't be a problem to receive everything and then "clean" the json response.

However, if we have a huge list of paragraphs, the previous "solution" won't scale.

Describe the solution you'd like
We would like to define a new type of summary in which we would like to retrieve the paragraphs that matched with our query. Furthermore, it will be useful if we can combine this feature with the dynamic summary feature, so we can have the same features as in a string field.

Describe alternatives you've considered
The key of this feature is what vespa thinks when we are talking about a match. As @jobergum says, vespa has several ways to match. The current implementation is for arrays of struct and maps, where vespa looks for exact matches of values using the sameElement operator.

I would recommend to take into account the match property defined on that field to "erase" the unimportant paragraphs of the summary

Additional context

• Related to https://blog.vespa.ai/semantic-search-with-multi-vector-indexing/
  Check the slack thread here.
• Registering a concrete type class breaks indexing of array<string> #23125

[jobergum]

Considering standard match:text for the array, is it enough that at least one of the query terms searching the array matches? E.g., query='what is the sun made of' would match many paragraphs as the query contains frequent words.

[danitico]

@jobergum Yeah! That's it

[Alexander-Mark]

I think an implementation of #29549 would solve this? We have the same issue (each doc has many paragraphs) and in streaming mode we sometimes experience latency upper bounded by disk IO, as well as additional overhead of transferring extra text over the network.

[andreer]

That would help, then the lower-scoring paragraphs could be removed from the result in a searcher to reduce the response size.