Word frequency in names is not sufficient to interpolate word
frequency in the address because names of towns, states etc. are
much more frequently used than, say street names.
Don't add the postcode again in the list of address details and
make sure that the result proper always comes before anything else
independently of the address rank.
Move the first cutting of the result list before reranking
by result match. This means that results with significantly
less importance are removed early and independently of the
fact how well they match the original query.
Fixes#3266.
The index for nameaddress_vector has grown so large that PostgreSQL
will resort to a sequential scan if there are too many items
in the LookupAny list.
A normal address search with housenumber will use name rankings for
the street name. This is slightly different than weighing for
address parts. Use the same ranking for the first part of the
address for housenumber-only searches to make sure that penalties
remain comparable.
This already worked for left-to-right readings and now is also
implemented for right-to-left reading. A qualifier must always be
before or after the name.
A country search result usually comes with a very high importance.
As a result only other very well known places will show up together
with country results and that means only places with lower address
ranks. Name searches for country names tend to yield a lot of POI
results because the country name is part of the name
(think "embassy of Sweden"). By excluding POIs from further searches,
the search is sped up quite a bit.
There is a bug in SQLAlchemy that assigns the wrong value to bind
parameters from closure variables when reusing lambda statements
that are later extended with other non-lambda expressions.
Thus either avoid lambda statements with closure variables or extending
them with non-lambda expressions.
A non-index version of ST_DWithin is not necessary. ST_Distance
can be used for that purpose. Index use for intersects can be
covered with a simple parameter.
Use JSON arrays which can have mixed types and therefore have
a more logical structure than separate arrays. Avoid JSON dicts
because of their verboseness.
Connecting is reasonably expensive because the spatialite extension
needs to be loaded. Disable pooling for tests because there is some
memory leak when quickly opening and closing QueuePools with sqlite
connections.
