In Discourse-AI we started experimenting with semantic search. This is still early days and we are still exploring these systems.
Using LLMs to improve the search prompt is also a possible (albeit slow today) approach:
This technique is mentioned here: GitHub - texttron/hyde: HyDE: Precise Zero-Shot Dense Retrieval without Relevance Labels
Besides the 100% automated approaches
Our general strategy here is to iterate. We already have “watched words” in the product, I would not mind a feature that adds “Search Synonyms” where you specify common typos and common phrases you wish to “stuff”. It is not scheduled work but certainly something you could look at sponsoring.
There is precedence for this exact feature in Postgres per: https://www.postgresql.org/docs/current/textsearch-dictionaries.html#TEXTSEARCH-SYNONYM-DICTIONARY
The other area I am open to exploring (I am only lukewarm though on this) is allowing for a hidden “metadata” place on posts, where admins can stuff search terms. It is very very invisible and generally I recommend just “properly” stuffing the works so stuff is not hidden eg:
SEO
semantic, related, improving