Freebase

» see freebase.com as it lived, on the wayback machine

~ the goodbye ~

freebase showed up in my browser tabs around 2011, mostly as a research bookmark. by then i had finished my engineering degree, was working in turkey on early computer vision contracts, and the bing and google knowledge panels that started appearing in search results that year had the same fingerprints as the freebase entity pages. you could click "edit" on a freebase topic and the typed edges underneath looked like a homework set for what we now call linked data. it was not part of my daily work, but it was a place to read about how a knowledge base actually got assembled.

the project began at metaweb technologies, founded by danny hillis and robert cook in 2005. freebase itself launched on march 3 2007 with a public beta of the site, the graphd graph database, and the mql query language. mql was a json-shaped thing where you wrote a partial template of the answer you wanted and the engine filled in the blanks. graphd was a single-tenant graph store with revision history built in.

google bought metaweb on july 16 2010 for an undisclosed price. less than a year later, the may 2012 launch of the knowledge graph used freebase as one of its three seed datasets, alongside wikipedia and the cia world factbook. freebase kept running as a public, editable site under google for another four years. on december 16 2014 google posted the shutdown notice. the final site went offline on may 2 2016, and the data dump was redirected to wikidata.

~ the chronicle ~

~ the actual product ~

freebase was a graph with typed edges. every node was a topic. every topic had one or more types, and each type carried a schema of properties. the actor type had a property for films, the film type had a property for cast, and the two pointed at each other. you queried it by writing a json document with the parts you knew filled in and the parts you wanted as nulls, and the engine returned the same json with the nulls replaced. that pattern is the direct ancestor of how the google knowledge graph api returns entity records today.

the data came from many places. the bulk imports brought in wikipedia infoboxes, musicbrainz for music metadata, geonames for places, and several government datasets. the long tail came from individual editors who filled in cast lists, fixed birth dates, and merged duplicate person records. by january 2014 freebase counted 44 million topics and 2.4 billion facts.

~ the postmortem ~

google announced the shutdown on december 16 2014 in a google plus post that linked to a transition guide. the timeline was clear: the api would freeze in march 2015, the site would go read-only on june 30 2015, and bulk dumps would continue until the final offline date. the recommended successor was wikidata, the wikimedia foundation project that had launched in october 2012 with a similar structured-claims model. google funded a tooling effort to help freebase editors migrate their topics into wikidata, and the kg apis released to developers in 2015 returned ids that overlapped with the wikidata namespace.

the final shutdown happened on may 2 2016. the freebase.com site started returning a deprecation notice instead of the entity pages, and the developer console disappeared. the last public data dump from march 2016 stayed available on the google developer site for several years and is still archived by the wayback machine and by github mirrors.

~ what it left behind ~

the cleanest legacy is the google knowledge graph itself, which still uses kg ids that begin with "/m/" and trace back to freebase topic ids. when you see a movie infobox in a google search result, the underlying entity record was probably first edited by a freebase volunteer over a decade ago. the kg apis that other developers use in 2026 inherit the schema decisions freebase editors made in 2009 about how to type an album, a politician, or a chemical compound.

the open-source pieces are quieter. google released graphd on github in september 2018 under apache 2.0, two and a half years after freebase shut down. the mql language stayed behind in google internal tooling and never grew a wider public following. wikidata, which inherited the editor community, kept its own query language called sparql and its own claims-with-references model. freebase's shape did not survive on the public web, but its data and its identifiers did.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »