retain the scores of a single full text operation with multiple full text functions/operators

[tteofili]

when using ES|QL in scenarios where relevance ranking is important (e.g., full text search, vector search, hybrid search, etc.) scores are currently computed within the _score column, as a "side effect" of the different full text capabilities (match, qstr, kql functions, : operator).
the final values contained in the _score column are the result of (summing the scores of) possibly multiple full text functions/operators, so users might need to isolate some such "parts" of one ES|QL query to better debug/understand how the final scores come together.
e.g., this could be done with a score() function, together with an EVAL:

FROM test METADATA _score
  | WHERE match(content, "brown")
  | WHERE match(content, "fox")
  | EVAL first_score = score(match(content, "brown"))
  | KEEP id, _score, first_score
  | SORT id

In addition to that, isolating the scores of single full text operations in ES|QL might be nice in order to build custom scoring formulas.

[elasticsearchmachine]

Pinging @elastic/es-analytical-engine (Team:Analytics)

[pabloem]

hi! @tteofili - for my edification, what would be the input and output of a query like the one you suggest in your comment?

[tteofili]

for my edification, what would be the input and output of a query like the one you suggest in your comment?

if you're asking re the input/output of the score function, such function would accept an expression representing (a boolean combination of) full text expressions (e.g., match(title, "foo") AND qstr("bar")) whereas the output would be a list of double values (e.g., in a new column specified in the EVAL command), representing the score of each resulting doc. note that the score doesn't contribute in any way to filtering in/out rows, hence some rows might have first_score equals to 0.0.

given a query like

FROM test METADATA _score
  | WHERE match(content, "brown")
  | WHERE match(content, "fox")
  | EVAL first_score = score(match(content, "brown"))
  | KEEP id, _score, first_score
  | SORT id

the output will be a set of rows matching the query match(content, "brown") AND match(content, "fox"), sorted by id.
for each row we'll output:

• its id
• its score according to the match(content, "brown") AND match(content, "fox") query within the _score column
• its score for the match(content, "brown") query within the first_score column.