@jpountz as we discussed, I am opening a separate ticket for it

An infrastructure for boosting individual text fragments would allow handling of several very useful scenarios. For example:

1. Pre-analyze text to find any text fragments of importance, mark it up with boost instructions and submit to ES for indexing. Field mapping for such field could indicate that it supports embedded boosts and thus content of the field should be analyzed for boost instructions
   Quick Brown <hint boost="10">Fox</hint> Jumped...
   and boost payloads should be used at search time
2. Same infrastructure can be used for boosting multifield content based on boost specified for its contributing fields. Same applies to _all field
   please see Index time boost in multi_field ignored? #4108 for more details
3. Synonyms, stem etc. If you choose to put them together with the original tokens into the same field, you can have stemmer/synonym filter to give derived forms slightly lower boost. That may be an overkill performance wise since the same could be achieved by doing boolean MUST against combined stem/no-stem filed and SHOULD with OR against unstemmed field. I wonder which would be faster and better

BTW, or similar infrastructure could be used for other purposes than boost alone