increase autocomplete 'phrase:slop' setting from 2->3

[missinglink]

this PR resolves pelias/pelias#307

by increasing the phrase slop from 2->3 we are able to successfully match "52 Görlitzer Straße" with "Görlitzer Straße 52".

I will add integration tests in another PR against schema, the value of 3 was confirmed with the following test case (2 fails but 3 passes):

// test the minimum amount of slop required to retrieve documents
module.exports.tests.slop = function(test, common){
  test( 'slop', function(t){

    var suite = new elastictest.Suite( null, { schema: schema } );
    suite.action( function( done ){ setTimeout( done, 500 ); }); // wait for es to bring some shards up

    // index a document
    suite.action( function( done ){
      suite.client.index({
        index: suite.props.index,
        type: 'test',
        id: '1',
        body: { name: { default: '52 Görlitzer Straße' } }
      }, done);
    });

    // search using 'peliasQueryFullToken'
    // in this case we require a slop of 3 to return the same
    // record with the street number and street name reversed.
    // (as is common in European countries, such as Germany).
    suite.assert( function( done ){
      suite.client.search({
        index: suite.props.index,
        type: 'test',
        body: { query: { match: {
          'name.default': {
            'analyzer': 'peliasQueryFullToken',
            'query': 'Görlitzer Straße 52',
            'type': 'phrase',
            'slop': 3,
          }
        }}}
      }, function( err, res ){
        t.equal( err, undefined );
        t.equal( res.hits.total, 1, 'document found' );
        done();
      });
    });

    suite.run( t.end );
  });
};

closes pelias/pelias#307

[hannesj]

This wouldn't still fix all issues, as there are street names consisting of three parts, such as Minna Canthin katu.

[riordan]

@hannesj: yes, you're absolutely correct.

However, when combined with Libpostal's improved address parsing, we'll be able to better apply this rule based on component parts.

e.g.

In [4]: parser.parse_address(" Minna Canthin katu 32")
Out[4]: [(u'minna canthin katu', u'road'), (u'32', u'house_number')]

In [5]: parser.parse_address("katu Minna Canthin 32")
Out[5]: [(u'katu minna canthin', u'road'), (u'32', u'house_number')]

In [6]: parser.parse_address("32 Minna Cathin katu")
Out[6]: [(u'32', u'house_number'), (u'minna', u'house'), (u'cathin katu', u'road')]

In [7]: parser.parse_address("Görlitzer Straße 52")
Out[7]: [(u'goerlitzer strasse', u'road'), (u'52', u'house_number')]

It didn't get it right on 32 Minna Cathin Katu, but we can work to get some redundancy built in and add failures to the next model training.

[hannesj]

Thanks, do you have any estimate when you would start working with the libpostal integration?

[orangejulius]

This gets my 👍 after an acceptance-test and loadtest on dev or prod_build.