Post #3287710
2026-06-03 17:12 UTC
@reykjalin@social.treehouse.systems @b0rk@social.jvns.ca Serious research question: suppose a search engine that uses an LLM to *interpret queries* but not to produce text; the output is the traditional set of documents that seem to be statistically relevant to the query.
The search engine's operators will need to train the model on the documents it can retrieve. How much *additional* natural-language text will they need to feed in to get something that does acceptably well at query interpretation?
possibly relevant to @marginalia@mastodon.social
Replies (2)
-
@drwho@masto.hackers.town 2026-06-03 17:19
@zwol@masto.hackers.town @reykjalin@social.treehouse.systems @b0rk@social.jvns.ca @marginalia@mastodon.social That's a search in a vector database. You don't really need an LLM for that, just a decent UI.
-
@marginalia@mastodon.social 2026-06-03 18:09
@zwol@masto.hackers.town @reykjalin@social.treehouse.systems @b0rk@social.jvns.ca Seems fairly similar to BERT, which is basically a predecessor to modern LLMs. Unsure of how much data you'd need to train something like that, transformer models in general are very resource hungry for training so they're kinda beyond what you can play with without datacenter GPU access.