
Term expansion is a lexical trick. Feed it to a dense encoder and you drag the query off the answer.
The most consequential numbers in an ML system are the ones nobody tuned.
Engineer working on retrieval and evaluation for LLM systems. I write Expensive To Be Wrong, a blog about the numbers those systems report and the defaults nobody actually chose: why your hybrid search is closer to a voting machine than a ranker, why an accuracy with no error bar is an anecdote with a decimal point, and when a system should just say "I don't know." Mostly arithmetic you can redo with a calculator, and experiments you can run in an afternoon.
Markdown supported · ⌘↵ to post
No responses yet.