How natural-language property search works, for Europe
Natural-language property search is a way of finding homes by typing a sentence the way you'd say it out loud, such as "two-bedroom apartment in Prague near a school, under €200,000", and getting a filtered map and list back. Instead of clicking through ten dropdowns, you describe what you want and the system pulls out the criteria.
This kind of search only became reliable in the last two years. Earlier filter-based portals could handle one country, one language, and a fixed schema. A buyer searching across Europe runs into a different problem: every country names its rooms, layouts and listing types differently, and most portals only speak one language at a time. Seeki was built to close that gap.
This post explains how the search works at a high level, why cross-border buying breaks traditional filters, and where the system still asks for help instead of guessing.
How it works
A sentence goes through a few steps before it becomes a result. First the system reads the sentence and pulls out the parts that map to a property database: a place, a buy-or-rent intent, a property type, a price range, a room count, must-have amenities, and any nearby points of interest. Then it resolves the place name against a hierarchical gazetteer of European locations, so "Prague", "Praga", "Praha" and "Prag" all land on the same city. The criteria become a database query against the marketplace, with prices converted to euros so currencies do not get in the way. The result is a map and a paginated list, and the URL updates to a shareable form you can bookmark or send.
The reason this matters more in Europe than in any single-country market is that the unit names do not agree across borders. A two-bedroom apartment in English is a "T3" in Portuguese real-estate terminology (the number counts rooms, not bedrooms). In Czech it is a "3+kk" or "2+1" depending on whether the kitchenette is part of the main room. In German listings it is "Drei-Zimmer-Wohnung". In Italian "trilocale". In Polish "M3". None of these map cleanly into the others. A dropdown that says "2 bedrooms" in English does not survive translation. It changes its meaning every country it crosses.
Natural-language search sidesteps this by treating the question as language and the result as data. You speak the vocabulary of your own market, and the system translates into one canonical schema before searching. The same mechanism powers the cross-border use case: a German buyer planning a holiday rental can type "house with a pool near Lake Balaton under €300,000" in German, and get listings from Hungary back, priced in euros, with the same filters they would use at home.
A worked example
| Step | What goes in | What comes out |
|---|---|---|
| Read the sentence | "2-bedroom apartment in Prague near a school under €200k" | place: Prague, buy or rent: buy, type: apartment, layout: 2-bedroom, max price: €200k, near: school |
| Resolve the place | "Prague" | the Prague area page in your language |
| Filter the marketplace | the criteria above | apartments in Prague matching the filters |
| Render | the matching listings | a map and a list, with a shareable URL |
A sentence in another language traces the same path. "Mieszkanie dwupokojowe w Warszawie do 400 000 zł" lands on the Warsaw apartments page with a price filter converted to euros internally.
Where the system asks for help
A sentence-based search is only as good as its handling of vague or impossible queries. Three cases come up regularly.
If the sentence is not about real estate, the chat returns a polite redirect and you keep typing. If the sentence is about real estate but the place name does not match anything in our gazetteer (an unknown village, an unsupported country, a misspelling we could not salvage), the chat asks for a nearby town or region instead of returning a city in the wrong country. If the sentence has no actionable criterion ("I want a nice flat"), the chat asks one targeted follow-up: which city, what budget, buy or rent.
These guards exist because silently picking a default is worse than asking one extra question. A search system that confidently returns the wrong city is the failure mode this kind of product fails at most often.
Frequently asked questions
Does it work in my language?
It works in English, Czech, Slovak, German, Dutch, Polish, Spanish, Italian, Portuguese, and French. You can write the query in any of them and get results back in the same language, with listings from across the EU markets Seeki covers, regardless of which language those listings were originally written in. The titles and descriptions of the listings themselves are translated on demand into your chosen language.
What if my query is too vague?
The system tells you it is too vague rather than guessing. The chat returns a short follow-up ("which city?", "what's your budget?", "buy or rent?") and waits. One round of clarification is normally enough to produce a useful search. If you would rather click filters, the result page still has all the dropdowns available, and the sentence-based search is the entry point, not the only path.
Can it find listings near a school, a metro stop, or a park?
Yes. The system picks up point-of-interest constraints if you mention them, and narrows results to listings within walking distance of a matching POI. Schools, kindergartens, metro and tram stops, parks, supermarkets, hospitals and a few other amenity types are recognised.
How does it handle different room-count systems?
The system knows the local conventions. A Portuguese "T3" and a German "Drei-Zimmer-Wohnung" land on the same internal layout value, and the database query handles both. You do not have to learn another country's terminology to search there.
What happens after I see the results?
The URL is shareable and indexable, so you can bookmark it or send it. Refining the search (changing the price, adding a filter, panning the map) keeps you on the same page, and no further model calls are needed once you have landed. Opening a listing shows the property detail with similar properties on the map, all in your language.
Is the language model deciding which property is best?
No. It only identifies what you asked for. Ranking and filtering happen inside the database against objective criteria (price, location, room count, the listing's own data), not against a model's preference. The system is a translator from sentences to filters, not a recommender.
Seeki is the marketplace, and the sentence-based search is one way in. For background on how prices and stock differ across the continent, the country price snapshot shows live €/m² for Czechia by region, and the Prague district breakdown walks through how a single city can sit in several different price tiers at once.