Introduce word usage statistics for addresses

nominatim/api/search/db_search_builder.py

@@ -226,27 +226,74 @@ def yield_lookups(self, name: TokenRange, address: List[TokenRange])\
         name_fulls = self.query.get_tokens(name, TokenType.WORD)
         if name_fulls:
             fulls_count = sum(t.count for t in name_fulls)
-            # At this point drop unindexed partials from the address.
-            # This might yield wrong results, nothing we can do about that.
-            if not partials_indexed:
-                addr_tokens = [t.token for t in addr_partials if t.is_indexed]
+            if len(name_partials) == 1:
+                penalty += min(0.5, max(0, (exp_count - 50 * fulls_count) / (2000 * fulls_count)))
+            if partials_indexed:
                 penalty += 1.2 * sum(t.penalty for t in addr_partials if not t.is_indexed)
-            # Any of the full names applies with all of the partials from the address
-            yield penalty, fulls_count / (2**len(addr_tokens)),\
-                  dbf.lookup_by_any_name([t.token for t in name_fulls],
-                                         addr_tokens,
-                                         fulls_count > 30000 / max(1, len(addr_tokens)))
+
+            yield penalty,fulls_count / (2**len(addr_tokens)), \
+                  self.get_full_name_ranking(name_fulls, addr_partials,
+                                             fulls_count > 30000 / max(1, len(addr_tokens)))
 
         # To catch remaining results, lookup by name and address
         # We only do this if there is a reasonable number of results expected.
         exp_count = exp_count / (2**len(addr_tokens)) if addr_tokens else exp_count
         if exp_count < 10000 and all(t.is_indexed for t in name_partials.values()):
-            lookup = [dbf.FieldLookup('name_vector', list(name_partials.keys()), lookups.LookupAll)]
-            if addr_tokens:
-                lookup.append(dbf.FieldLookup('nameaddress_vector', addr_tokens, lookups.LookupAll))
             penalty += 0.35 * max(1 if name_fulls else 0.1,
                                   5 - len(name_partials) - len(addr_tokens))
-            yield penalty, exp_count, lookup
+            yield penalty, exp_count,\
+                  self.get_name_address_ranking(list(name_partials.keys()), addr_partials)
+
+
+    def get_name_address_ranking(self, name_tokens: List[int],
+                                 addr_partials: List[Token]) -> List[dbf.FieldLookup]:
+        """ Create a ranking expression looking up by name and address.
+        """
+        lookup = [dbf.FieldLookup('name_vector', name_tokens, lookups.LookupAll)]
+
+        addr_restrict_tokens = []
+        addr_lookup_tokens = []
+        for t in addr_partials:
+            if t.is_indexed:
+                if t.addr_count > 20000:
+                    addr_restrict_tokens.append(t.token)
+                else:
+                    addr_lookup_tokens.append(t.token)
+
+        if addr_restrict_tokens:
+            lookup.append(dbf.FieldLookup('nameaddress_vector',
+                                          addr_restrict_tokens, lookups.Restrict))
+        if addr_lookup_tokens:
+            lookup.append(dbf.FieldLookup('nameaddress_vector',
+                                          addr_lookup_tokens, lookups.LookupAll))
+
+        return lookup
+
+
+    def get_full_name_ranking(self, name_fulls: List[Token], addr_partials: List[Token],
+                              use_lookup: bool) -> List[dbf.FieldLookup]:
+        """ Create a ranking expression with full name terms and
+            additional address lookup. When 'use_lookup' is true, then
+            address lookups will use the index, when the occurences are not
+            too many.
+        """
+        # At this point drop unindexed partials from the address.
+        # This might yield wrong results, nothing we can do about that.
+        if use_lookup:
+            addr_restrict_tokens = []
+            addr_lookup_tokens = []
+            for t in addr_partials:
+                if t.is_indexed:
+                    if t.addr_count > 20000:
+                        addr_restrict_tokens.append(t.token)
+                    else:
+                        addr_lookup_tokens.append(t.token)
+        else:
+            addr_restrict_tokens = [t.token for t in addr_partials if t.is_indexed]
+            addr_lookup_tokens = []
+
+        return dbf.lookup_by_any_name([t.token for t in name_fulls],
+                                      addr_restrict_tokens, addr_lookup_tokens)
 
 
     def get_name_ranking(self, trange: TokenRange,

nominatim/api/search/db_search_fields.py

@@ -231,16 +231,17 @@ def lookup_by_names(name_tokens: List[int], addr_tokens: List[int]) -> List[Fiel
     return lookup
 
 
-def lookup_by_any_name(name_tokens: List[int], addr_tokens: List[int],
-                       use_index_for_addr: bool) -> List[FieldLookup]:
+def lookup_by_any_name(name_tokens: List[int], addr_restrict_tokens: List[int],
+                       addr_lookup_tokens: List[int]) -> List[FieldLookup]:
     """ Create a lookup list where name tokens are looked up via index
         and only one of the name tokens must be present.
         Potential address tokens are used to restrict the search further.
     """
     lookup = [FieldLookup('name_vector', name_tokens, lookups.LookupAny)]
-    if addr_tokens:
-        lookup.append(FieldLookup('nameaddress_vector', addr_tokens,
-                                  lookups.LookupAll if use_index_for_addr else lookups.Restrict))
+    if addr_restrict_tokens:
+        lookup.append(FieldLookup('nameaddress_vector', addr_restrict_tokens, lookups.Restrict))
+    if addr_lookup_tokens:
+        lookup.append(FieldLookup('nameaddress_vector', addr_lookup_tokens, lookups.LookupAll))
 
     return lookup
 

nominatim/api/search/icu_tokenizer.py

@@ -97,6 +97,7 @@ def from_db_row(row: SaRow) -> 'ICUToken':
         """ Create a ICUToken from the row of the word table.
         """
         count = 1 if row.info is None else row.info.get('count', 1)
+        addr_count = 1 if row.info is None else row.info.get('addr_count', 1)
 
         penalty = 0.0
         if row.type == 'w':
@@ -123,7 +124,8 @@ def from_db_row(row: SaRow) -> 'ICUToken':
 
         return ICUToken(penalty=penalty, token=row.word_id, count=count,
                         lookup_word=lookup_word, is_indexed=True,
-                        word_token=row.word_token, info=row.info)
+                        word_token=row.word_token, info=row.info,
+                        addr_count=addr_count)
 
 
 
@@ -257,7 +259,7 @@ def add_extra_tokens(self, query: qmod.QueryStruct, parts: QueryParts) -> None:
             if len(part.token) <= 4 and part[0].isdigit()\
                and not node.has_tokens(i+1, qmod.TokenType.HOUSENUMBER):
                 query.add_token(qmod.TokenRange(i, i+1), qmod.TokenType.HOUSENUMBER,
-                                ICUToken(0.5, 0, 1, part.token, True, part.token, None))
+                                ICUToken(0.5, 0, 1, 1, part.token, True, part.token, None))
 
 
     def rerank_tokens(self, query: qmod.QueryStruct, parts: QueryParts) -> None:

nominatim/api/search/legacy_tokenizer.py

@@ -210,6 +210,7 @@ def make_token(self, row: SaRow) -> Tuple[LegacyToken, qmod.TokenType]:
 
         return LegacyToken(penalty=penalty, token=row.word_id,
                            count=row.search_name_count or 1,
+                           addr_count=1, # not supported
                            lookup_word=lookup_word,
                            word_token=row.word_token.strip(),
                            category=(rowclass, row.type) if rowclass is not None else None,
@@ -226,7 +227,7 @@ def add_extra_tokens(self, query: qmod.QueryStruct, parts: List[str]) -> None:
             if len(part) <= 4 and part.isdigit()\
                and not node.has_tokens(i+1, qmod.TokenType.HOUSENUMBER):
                 query.add_token(qmod.TokenRange(i, i+1), qmod.TokenType.HOUSENUMBER,
-                                LegacyToken(penalty=0.5, token=0, count=1,
+                                LegacyToken(penalty=0.5, token=0, count=1, addr_count=1,
                                             lookup_word=part, word_token=part,
                                             category=None, country=None,
                                             operator=None, is_indexed=True))

nominatim/api/search/query.py

@@ -99,6 +99,7 @@ class Token(ABC):
     penalty: float
     token: int
     count: int
+    addr_count: int
     lookup_word: str
     is_indexed: bool
 

nominatim/clicmd/refresh.py

@@ -110,7 +110,8 @@ def run(self, args: NominatimArgs) -> int: #pylint: disable=too-many-branches, t
 
         if args.word_counts:
             LOG.warning('Recompute word statistics')
-            self._get_tokenizer(args.config).update_statistics(args.config)
+            self._get_tokenizer(args.config).update_statistics(args.config,
+                                                               threads=args.threads or 1)
 
         if args.address_levels:
             LOG.warning('Updating address levels')

nominatim/clicmd/setup.py

@@ -168,7 +168,7 @@ def run(self, args: NominatimArgs) -> int: # pylint: disable=too-many-statements
         tokenizer.finalize_import(args.config)
 
         LOG.warning('Recompute word counts')
-        tokenizer.update_statistics(args.config)
+        tokenizer.update_statistics(args.config, threads=num_threads)
 
         webdir = args.project_dir / 'website'
         LOG.warning('Setup website at %s', webdir)

nominatim/tokenizer/base.py

@@ -201,7 +201,7 @@ def check_database(self, config: Configuration) -> Optional[str]:
 
 
     @abstractmethod
-    def update_statistics(self, config: Configuration) -> None:
+    def update_statistics(self, config: Configuration, threads: int = 1) -> None:
         """ Recompute any tokenizer statistics necessary for efficient lookup.
             This function is meant to be called from time to time by the user
             to improve performance. However, the tokenizer must not depend on

nominatim/tokenizer/icu_tokenizer.py

@@ -104,30 +104,97 @@ def check_database(self, config: Configuration) -> None:
         self.init_from_project(config)
 
 
-    def update_statistics(self, config: Configuration) -> None:
+    def update_statistics(self, config: Configuration, threads: int = 2) -> None:
         """ Recompute frequencies for all name words.
         """
         with connect(self.dsn) as conn:
             if not conn.table_exists('search_name'):
                 return
 
             with conn.cursor() as cur:
-                LOG.info('Computing word frequencies')
-                cur.drop_table('word_frequencies')
-                cur.execute("""CREATE TEMP TABLE word_frequencies AS
-                                 SELECT unnest(name_vector) as id, count(*)
-                                 FROM search_name GROUP BY id""")
-                cur.execute('CREATE INDEX ON word_frequencies(id)')
-                LOG.info('Update word table with recomputed frequencies')
-                cur.drop_table('tmp_word')
-                cur.execute("""CREATE TABLE tmp_word AS
-                                SELECT word_id, word_token, type, word,
-                                       (CASE WHEN wf.count is null THEN info
-                                          ELSE info || jsonb_build_object('count', wf.count)
-                                        END) as info
-                                FROM word LEFT JOIN word_frequencies wf
-                                  ON word.word_id = wf.id""")
-                cur.drop_table('word_frequencies')
+                cur.execute('ANALYSE search_name')
+                if threads > 1:
+                    cur.execute('SET max_parallel_workers_per_gather TO %s',
+                                (min(threads, 6),))
+
+                if conn.server_version_tuple() < (12, 0):
+                    LOG.info('Computing word frequencies')
+                    cur.drop_table('word_frequencies')
+                    cur.drop_table('addressword_frequencies')
+                    cur.execute("""CREATE TEMP TABLE word_frequencies AS
+                                     SELECT unnest(name_vector) as id, count(*)
+                                     FROM search_name GROUP BY id""")
+                    cur.execute('CREATE INDEX ON word_frequencies(id)')
+                    cur.execute("""CREATE TEMP TABLE addressword_frequencies AS
+                                     SELECT unnest(nameaddress_vector) as id, count(*)
+                                     FROM search_name GROUP BY id""")
+                    cur.execute('CREATE INDEX ON addressword_frequencies(id)')
+                    cur.execute("""CREATE OR REPLACE FUNCTION word_freq_update(wid INTEGER,
+                                                                               INOUT info JSONB)
+                                   AS $$
+                                   DECLARE rec RECORD;
+                                   BEGIN
+                                   IF info is null THEN
+                                     info = '{}'::jsonb;
+                                   END IF;
+                                   FOR rec IN SELECT count FROM word_frequencies WHERE id = wid
+                                   LOOP
+                                     info = info || jsonb_build_object('count', rec.count);
+                                   END LOOP;
+                                   FOR rec IN SELECT count FROM addressword_frequencies WHERE id = wid
+                                   LOOP
+                                     info = info || jsonb_build_object('addr_count', rec.count);
+                                   END LOOP;
+                                   IF info = '{}'::jsonb THEN
+                                     info = null;
+                                   END IF;
+                                   END;
+                                   $$ LANGUAGE plpgsql IMMUTABLE;
+                                """)
+                    LOG.info('Update word table with recomputed frequencies')
+                    cur.drop_table('tmp_word')
+                    cur.execute("""CREATE TABLE tmp_word AS
+                                    SELECT word_id, word_token, type, word,
+                                           word_freq_update(word_id, info) as info
+                                    FROM word
+                                """)
+                    cur.drop_table('word_frequencies')
+                    cur.drop_table('addressword_frequencies')
+                else:
+                    LOG.info('Computing word frequencies')
+                    cur.drop_table('word_frequencies')
+                    cur.execute("""
+                      CREATE TEMP TABLE word_frequencies AS
+                      WITH word_freq AS MATERIALIZED (
+                               SELECT unnest(name_vector) as id, count(*)
+                                     FROM search_name GROUP BY id),
+                           addr_freq AS MATERIALIZED (
+                               SELECT unnest(nameaddress_vector) as id, count(*)
+                                     FROM search_name GROUP BY id)
+                      SELECT coalesce(a.id, w.id) as id,
+                             (CASE WHEN w.count is null THEN '{}'::JSONB
+                                  ELSE jsonb_build_object('count', w.count) END
+                              ||
+                              CASE WHEN a.count is null THEN '{}'::JSONB
+                                  ELSE jsonb_build_object('addr_count', a.count) END) as info
+                      FROM word_freq w FULL JOIN addr_freq a ON a.id = w.id;
+                      """)
+                    cur.execute('CREATE UNIQUE INDEX ON word_frequencies(id) INCLUDE(info)')
+                    cur.execute('ANALYSE word_frequencies')
+                    LOG.info('Update word table with recomputed frequencies')
+                    cur.drop_table('tmp_word')
+                    cur.execute("""CREATE TABLE tmp_word AS
+                                    SELECT word_id, word_token, type, word,
+                                           (CASE WHEN wf.info is null THEN word.info
+                                            ELSE coalesce(word.info, '{}'::jsonb) || wf.info
+                                            END) as info
+                                    FROM word LEFT JOIN word_frequencies wf
+                                         ON word.word_id = wf.id
+                                """)
+                    cur.drop_table('word_frequencies')
+
+            with conn.cursor() as cur:
+                cur.execute('SET max_parallel_workers_per_gather TO 0')
 
             sqlp = SQLPreprocessor(conn, config)
             sqlp.run_string(conn,

nominatim/tokenizer/legacy_tokenizer.py

@@ -210,7 +210,7 @@ def migrate_database(self, config: Configuration) -> None:
             self._save_config(conn, config)
 
 
-    def update_statistics(self, _: Configuration) -> None:
+    def update_statistics(self, config: Configuration, threads: int = 1) -> None:
         """ Recompute the frequency of full words.
         """
         with connect(self.dsn) as conn:

test/python/api/search/test_api_search_query.py

@@ -18,7 +18,8 @@ def get_category(self):
 
 
 def mktoken(tid: int):
-    return MyToken(3.0, tid, 1, 'foo', True)
+    return MyToken(penalty=3.0, token=tid, count=1, addr_count=1,
+                   lookup_word='foo', is_indexed=True)
 
 
 @pytest.mark.parametrize('ptype,ttype', [('NONE', 'WORD'),

test/python/api/search/test_db_search_builder.py

@@ -31,7 +31,9 @@ def make_query(*args):
         for end, ttype, tinfo in tlist:
             for tid, word in tinfo:
                 q.add_token(TokenRange(start, end), ttype,
-                            MyToken(0.5 if ttype == TokenType.PARTIAL else 0.0, tid, 1, word, True))
+                            MyToken(penalty=0.5 if ttype == TokenType.PARTIAL else 0.0,
+                                    token=tid, count=1, addr_count=1,
+                                    lookup_word=word, is_indexed=True))
 
 
     return q
@@ -395,14 +397,14 @@ def make_counted_searches(name_part, name_full, address_part, address_full,
     q.add_node(BreakType.END, PhraseType.NONE)
 
     q.add_token(TokenRange(0, 1), TokenType.PARTIAL,
-                MyToken(0.5, 1, name_part, 'name_part', True))
+                MyToken(0.5, 1, name_part, 1, 'name_part', True))
     q.add_token(TokenRange(0, 1), TokenType.WORD,
-                MyToken(0, 101, name_full, 'name_full', True))
+                MyToken(0, 101, name_full, 1, 'name_full', True))
     for i in range(num_address_parts):
         q.add_token(TokenRange(i + 1, i + 2), TokenType.PARTIAL,
-                    MyToken(0.5, 2, address_part, 'address_part', True))
+                    MyToken(0.5, 2, address_part, 1, 'address_part', True))
         q.add_token(TokenRange(i + 1, i + 2), TokenType.WORD,
-                    MyToken(0, 102, address_full, 'address_full', True))
+                    MyToken(0, 102, address_full, 1, 'address_full', True))
 
     builder = SearchBuilder(q, SearchDetails())
 

test/python/api/search/test_token_assignment.py

@@ -19,7 +19,8 @@ def get_category(self):
 
 def make_query(*args):
     q = QueryStruct([Phrase(args[0][1], '')])
-    dummy = MyToken(3.0, 45, 1, 'foo', True)
+    dummy = MyToken(penalty=3.0, token=45, count=1, addr_count=1,
+                    lookup_word='foo', is_indexed=True)
 
     for btype, ptype, _ in args[1:]:
         q.add_node(btype, ptype)

test/python/cli/conftest.py

@@ -32,16 +32,16 @@ def __init__(self, *args, **kwargs):
         self.update_statistics_called = False
         self.update_word_tokens_called = False
 
-    def update_sql_functions(self, *args):
+    def update_sql_functions(self, *args, **kwargs):
         self.update_sql_functions_called = True
 
-    def finalize_import(self, *args):
+    def finalize_import(self, *args, **kwargs):
         self.finalize_import_called = True
 
-    def update_statistics(self, *args):
+    def update_statistics(self, *args, **kwargs):
         self.update_statistics_called = True
 
-    def update_word_tokens(self, *args):
+    def update_word_tokens(self, *args, **kwargs):
         self.update_word_tokens_called = True