The fairly complex where condition of idx_placex_geometry_placenode
won't always be matched by the query planner if the condition
part doesn't appear verbatim in the query.
Fixes#2480.
Running the warm-up search requests requires querying
the most frequent words. This must be done via the tokenizer
to honor the different formats of the word table.
Additional penalty for special terms with operator None
should only go to near searches. To reduce the number
of produced searches, restrict the none operator to
appear only in conjunction with the name.
The new icu tokenizer is now no longer compatible with the old
legacy tokenizer in terms of data structures. Therefore there
is also no longer a need to refer to the legacy tokenizer in the
name.
The hack for IL, AL and LA is only needed because these abbreviations
are removed by the legacy tokenizer as a stop word. There is no need
to keep the hack for future tokenizers. Move it therefore to the
token extraction function.
This separates the logic of creating word sets from the Phrase
class. A tokenizer may now derived the word sets any way they
like. The SimpleWordList class provides a standard implementation
for splitting phrases on spaces.
Restricting tokens due to the search context is better done in
the generic search part instead of repeating the same test in
every tokenizer implementation.
Postgresql is very bad at creating statistics for jsonb
columns. The result is that the query planer tends to
use JIT for queries with a where over 'info' even when
there is an index.
The BDD tests cannot make assumptions about the structure of the
word table anymore because it depends on the tokenizer. Use more
abstract descriptions instead that ask for specific kinds of
tokens.
Moving the logic for extending the SearchDescription into the
token classes splits up the code and makes it more readable.
More importantly: it allows tokenizer to define custom token
classes in the future.
The token string is only required by the PartialToken type, so
it can simply save the token string internally. No need to pass
it to every type.
Also moves the check for multi-word partials to the token loader
code in the tokenizer. Multi-word partials can only happen with
the legacy tokenizer and when the database was loaded with an
older version of Nominatim. No need to keep the check for
everybody.
Moves token and phrase position and phrase type into a separate
class that is handed in when assembling the search description.
This drastically reduces the number of parameters for the function
to extend the search descriptions and gives us more flexibility
in the future for more complex positional analysis.
Full-word tokens are no longer marked by a space at the
beginning of the token. Use the new Partial token category
instead. This removes a couple of special casing, we don't
really need.
The word table still has the space for compatibility reasons,
so the tokenizer code needs to get rid of it when loading the
tokens.
This adds precomputation of abbreviated terms for names and removes
abbreviation of terms in the query. Basic import works but still
needs some thorough testing as well as speed improvements during
import.
New dependency for python library datrie.
Now that mutli-word partials no longer exist, multi-word full
words need to be used to search in addresses and therefore no
longer should have a penalty.
Also changes the condition when a full word is included into
the address. It is no longer relevant if an equivalent partial
exists but only if the term consists of more than one word.
When searching for house numbers in the name (for place-only
terms) then the same penalties need to apply as for the
regular house number search.
Change the code to first compute the penalties and then create
the new search variants.
