How does NeuralSearch find listings?
It runs keyword, semantic and image retrieval at the same time, merges the three result sets, then re-scores the merged shortlist against the specific constraints in your question. Retrieval decides what is considered; the re-scoring step decides what you see first.
Running the retrievals in parallel matters because each fails differently: keyword search misses paraphrase, semantic search drifts on exact values, and neither reads photographs. Merging keeps evidence from all three rather than committing to one. The re-ranking pass is where most of the final ordering comes from — it is the only stage that checks whether a listing actually satisfies what you asked for.