Too long; read later - here's a demo : SPARQL Sliders Test
"I'd like a search tool for furniture that works like Google's Flight Search does for flights. That is, with sliders so I can say what type of furniture (table), what range of widths (1-2m), lengths (2-5m), and heights (1-2m), what material (wood), what thickness, what price range, etc, I'd like, with the list of available products updating in real time."
For convenience/lack of available data the demo runs against dpPedia via the SNORQL SPARQL Explorer. As furniture and it's dimensions wasn't available it uses cities and their populations and elevations.
So how would you get real data?
First of all, furniture vendors could either provide dumps of their data or, more Webby, mark up their sites with RDFa and/or HTML5 microdata using e.g. the GoodRelations e-commerce vocabulary.
Ultimately, for a front end like these sliders to work, the data would need to go in a store with a SPARQL endpoint. But, triplestores shouldn't be thought of as just a wacky alternative to a SQL database. A triplestore is just a cache of a little chunk of the Linked Data Web. The question of where the store resides and how the data is collected is entirely open. Following the more traditional DB model, a service might aggregate the data published by known furniture suppliers and provide the endpoint online.
But alternately, a local user agent (I think Chris Bizer had a little Java example, can't find the link...there are others) could crawl the Web to answer the query just-in-time. The advantage of this approach is that it's more thorough and the only real option for totally arbitrary queries, the downside being that it's answer will probably take longer than milliseconds. But remember triplestores are caches, not every little bit of information would have to be discovered and read from every page. There are vocabs for dataset and vocab discovery (remind me of the acronyms please :) Note too that you're not limiting your client agent to a single datastore. traditional backends (SQL or NoSQL) are effectively isolated silos, triplestores are integrated with the links of the Web.
Incidentally, this is something that might be nice to express as a Web Intent, along the lines of "make me a query from this template with these parameters and apply it to this endpoint, putting the results into this widget" (that's a bit verbose for a general-purpose intent, but you get the gist). c.f. RDFAffordances.