semisol on Nostr: “Reasoning” in this case is being able to process the document to make ...

“Reasoning” in this case is being able to process the document to make extrapolations like linking 2026 oil prices <=> Iran war. An LLM with token-by-token generation + tools + CoT can do this pretty well, but embedding models are intentionally trained to classify similarity and not make extrapolations like this.

So it is up to either the document writer, or the querier, to stuff the relevant keywords in the embedded text