Semantic vs keyword search
Keyword search matches the words you typed and is precise about exact identifiers; semantic search matches meaning and is precise about intent. Keyword search fails when a listing uses different words for the same thing. Semantic search fails when you need an exact value, because it treats near-misses as matches. For listing search, where a query mixes a vague wish with hard constraints, neither is sufficient alone.
The short version
Use keyword search when you know the exact term and it will appear verbatim: a model number, a registration, a street name. Use semantic search when you can describe what you want but not name it. Real queries contain both — "a quiet flat near good schools under 500,000" is a vague wish and two hard constraints in one sentence — which is why choosing one is choosing to be wrong half the time.
Where keyword search wins
Identifiers and rare terms. If a query contains a string that appears in exactly the right documents and nowhere else, nothing beats matching it literally. Semantic search actively hurts here: it will find things that are merely similar to your exact identifier, which is not what "exact" means.
Where semantic search wins
Paraphrase and intent. "Somewhere to work from home" has no keyword. Neither does "a starter home that doesn't need work". Semantic search finds listings whose meaning matches even when no word overlaps, which is the majority of how people actually describe what they want.
Where both fail without a re-rank
Neither method checks anything. Keyword search does not know 500,000 is a maximum; semantic search does not know 2019 is not 2018. Both return plausible candidates and neither enforces a constraint. Enforcement is a separate step, applied to the shortlist after retrieval, and skipping it is the most common reason a search that looks sophisticated returns obviously wrong results.
What we do
Both retrievals run, their results merge, and the merged shortlist is re-scored against the constraints in your sentence, with image signals contributing to the ordering. See how it works for the full path from question to results.