Thanks to some recent extra hours by our team we're ready to release our latest labs prototype to the public in a hipness compliant alpha-state:

Infolust is a new context search engine. It's for all those moments when you sit in front of a website and think "Now that's interesting" and would like to know more. Now this "More" is just a single click away. It compares existing web-pages with all Wikipedia pages and fetches up to 10 related ones. It's a first step of bringing our Similarity Engine to consumers for free.

While we're still working hard on improving things like the IE6 bookmarklet capabilities, especially for pages that contain frames, it's necessary to gain as much statistics (of course no personal or address related whatsoever) as possible to optimize it for additional content sources than Wikipedia and improve the backend efficiency.

So far the results are surprisingly good for a third of the queries, quite OK for another third and not so good for the last third. The main issue isn't so much to determine the similarity itself but distinguish between the actual content on a query page and things like navigation, footers etc. especially because we don't do any "lookup wikipedia page titles in text" tricks but rather rely on a full semantic analysis and comparison of pages.

That said we'd be more than happy to hear your suggestions /sightings / comments and hope you enjoy what Infolust is doing so far!

Something new from the Labs: "Wikipedia³ is a conversion of the monthly wikipedia database dump into RDF. While it has always been our test dataset of choice, its wealth of information created by thousands of people all over the world also opens up new perspectives when put into a Semantic Web format: How to not (again) put structure or ontology, but people and their data first. How to meaningfully infer from the already known. How to scale real-world graph platforms. It's a foundation for a lot of interesting work."

Special thanks to Andreas Bolka (aka earl) & lowi for the hard polishing and automation work. We'll be releasing more labs stuff based on that dataset during the summer (especially visualizations). As always, feedback strongly welcome and hopefully this triggers loads of other good projects!

A few days ago we silently released our first contribution to System One Labs, the place where we test and play with new ideas that might become features of System One. Retrievr lets you find Flickr images by drawing rough sketches, and this must have hit the mark according to the feedback so far:

John Battelle thinks "Oooo, i like this...", Tara Calishain got it with "You won't be able to find electronics schema screenshots with this, but it's an awesome way to browse", Matt Haughey is over-enthusiastic "...retrievr is so good, I'd venture to say it's getting close to Turing Test territory" and Jason Kottke says "Retrievr is a simple, amazing use of the Flickr API". Special gratitude for the "All my wasted years of practicing drawing Goatse Man on MS Paint has just paid off." comment on the metafilter post, this also made our day. Thanks for all the appreciation, we'll try to continue what we started here!

Deepest thanks to Chris Langreiter for the idea to merge a 1995 paper with 2005 flickr and of course then executing it brilliantly.