Set `track_total_hits: false` for Elasticsearch

[jpountz]

By default, Elasticsearch starts by evaluating hits exhaustively, until it hits track_total_hits matches (default: 10,000), and then kicks off dynamic pruning via block-max MAXSCORE/WAND.

The motivation behind this behavior is to enable applications to paginate through results easily, e.g. you shouldn't put a link to the 7th page if you display 20 hits per page and your query only has 100 hits (so 5 pages). Looking at docs, it doesn't look like Vespa does something similar, so the ecommerce_hybrid_search should probably disable hit count tracking with Elasticsearch by setting track_total_hits to false (which means the same as 0)? See https://www.elastic.co/docs/solutions/search/the-search-api#track-total-hits.

[radu-gheorghe]

Oh, nice catch! Vespa has targetHits which does a similar thing in this case, but the default is different (100 vs 10,000).

So we should either make track_total_hits=100 or targetHits=10000

We did that for vector search (num_candidates=100), but we missed this one, sorry. Not intentional, of course.

[jpountz]

In cases such as pure lexical search in the benchmark when hits are not re-ranked after retrieval, would Vespa still honor this targetHits parameter or would it automatically trim it to the number of returned hits because computing the top-100 BM25 matches and then keeping only the 10 best ones would be the same as retrieving the top-10 BM25 matches directly?

[radu-gheorghe]

I think so, because Vespa will return a totalCount still (much like ES does with track_total_hits). I'll double-check when we get to fixing this.