How it works
NeuralSearch answers a plain-language question by running three retrievals at once — keyword, semantic and image — merging them, then re-ranking the merged set against the specifics you asked for. You describe what you want in a sentence; you get results ordered the way you would have ordered them yourself, and you can refine by continuing the conversation rather than adjusting filters.
What you can search
Two catalogues today: vehicles and property. Both accept ordinary sentences rather than filter forms. "A low-mileage estate car under 20,000 with a tow bar" and "a bright two-bedroom flat with a balcony, near a park" are both valid queries — the constraints are read out of the sentence, not typed into boxes.
Three retrievals, not one
A keyword search finds listings that literally contain your words. A semantic search finds listings that mean the same thing in different words. An image search finds listings that look like what you described. Each catches things the others miss, so all three run and their results are merged before anything is ranked.
Then it re-reads the shortlist
Merging gives a candidate pool, not an answer. The shortlist is re-scored against the specifics of your request — the numbers, the must-haves, the look — which is where most of the ordering actually comes from. Retrieval decides what is considered; re-ranking decides what you see first.
Refining is a conversation
"Same but cheaper" or "something with more light" continues the search rather than starting a new one. The context of what you already asked is carried forward, so you narrow by talking instead of by re-entering filters.
Where the listings come from
Listings are gathered from public sources and normalised into one shape, so a search spans catalogues that otherwise use different words for the same thing. Every listing keeps a link to its original source, and the original is always the authority on price and availability.