1 # SPDX-License-Identifier: GPL-3.0-or-later
3 # This file is part of Nominatim. (https://nominatim.org)
5 # Copyright (C) 2024 by the Nominatim developer community.
6 # For a full list of authors see the git log.
8 Tokenizer implementing normalisation as used before Nominatim 4 but using
9 libICU instead of the PostgreSQL module.
11 from typing import Optional, Sequence, List, Tuple, Mapping, Any, cast, \
15 from pathlib import Path
17 from psycopg.types.json import Jsonb
18 from psycopg import sql as pysql
20 from ..db.connection import connect, Connection, Cursor, \
21 drop_tables, table_exists, execute_scalar
22 from ..config import Configuration
23 from ..db.sql_preprocessor import SQLPreprocessor
24 from ..data.place_info import PlaceInfo
25 from ..data.place_name import PlaceName
26 from .icu_rule_loader import ICURuleLoader
27 from .place_sanitizer import PlaceSanitizer
28 from .icu_token_analysis import ICUTokenAnalysis
29 from .base import AbstractAnalyzer, AbstractTokenizer
31 DBCFG_TERM_NORMALIZATION = "tokenizer_term_normalization"
33 LOG = logging.getLogger()
35 WORD_TYPES = (('country_names', 'C'),
38 ('housenumbers', 'H'))
41 def create(dsn: str, data_dir: Path) -> 'ICUTokenizer':
42 """ Create a new instance of the tokenizer provided by this module.
44 return ICUTokenizer(dsn, data_dir)
47 class ICUTokenizer(AbstractTokenizer):
48 """ This tokenizer uses libICU to convert names and queries to ASCII.
49 Otherwise it uses the same algorithms and data structures as the
50 normalization routines in Nominatim 3.
53 def __init__(self, dsn: str, data_dir: Path) -> None:
55 self.data_dir = data_dir
56 self.loader: Optional[ICURuleLoader] = None
58 def init_new_db(self, config: Configuration, init_db: bool = True) -> None:
59 """ Set up a new tokenizer for the database.
61 This copies all necessary data in the project directory to make
62 sure the tokenizer remains stable even over updates.
64 self.loader = ICURuleLoader(config)
69 self.update_sql_functions(config)
70 self._setup_db_tables(config)
71 self._create_base_indices(config, 'word')
73 def init_from_project(self, config: Configuration) -> None:
74 """ Initialise the tokenizer from the project directory.
76 self.loader = ICURuleLoader(config)
78 with connect(self.dsn) as conn:
79 self.loader.load_config_from_db(conn)
81 def finalize_import(self, config: Configuration) -> None:
82 """ Do any required postprocessing to make the tokenizer data ready
85 self._create_lookup_indices(config, 'word')
87 def update_sql_functions(self, config: Configuration) -> None:
88 """ Reimport the SQL functions for this tokenizer.
90 with connect(self.dsn) as conn:
91 sqlp = SQLPreprocessor(conn, config)
92 sqlp.run_sql_file(conn, 'tokenizer/icu_tokenizer.sql')
94 def check_database(self, config: Configuration) -> None:
95 """ Check that the tokenizer is set up correctly.
97 # Will throw an error if there is an issue.
98 self.init_from_project(config)
100 def update_statistics(self, config: Configuration, threads: int = 2) -> None:
101 """ Recompute frequencies for all name words.
103 with connect(self.dsn) as conn:
104 if not table_exists(conn, 'search_name'):
107 with conn.cursor() as cur:
108 cur.execute('ANALYSE search_name')
110 cur.execute(pysql.SQL('SET max_parallel_workers_per_gather TO {}')
111 .format(pysql.Literal(min(threads, 6),)))
113 LOG.info('Computing word frequencies')
114 drop_tables(conn, 'word_frequencies')
116 CREATE TEMP TABLE word_frequencies AS
117 WITH word_freq AS MATERIALIZED (
118 SELECT unnest(name_vector) as id, count(*)
119 FROM search_name GROUP BY id),
120 addr_freq AS MATERIALIZED (
121 SELECT unnest(nameaddress_vector) as id, count(*)
122 FROM search_name GROUP BY id)
123 SELECT coalesce(a.id, w.id) as id,
124 (CASE WHEN w.count is null THEN '{}'::JSONB
125 ELSE jsonb_build_object('count', w.count) END
127 CASE WHEN a.count is null THEN '{}'::JSONB
128 ELSE jsonb_build_object('addr_count', a.count) END) as info
129 FROM word_freq w FULL JOIN addr_freq a ON a.id = w.id;
131 cur.execute('CREATE UNIQUE INDEX ON word_frequencies(id) INCLUDE(info)')
132 cur.execute('ANALYSE word_frequencies')
133 LOG.info('Update word table with recomputed frequencies')
134 drop_tables(conn, 'tmp_word')
135 cur.execute("""CREATE TABLE tmp_word AS
136 SELECT word_id, word_token, type, word,
137 (CASE WHEN wf.info is null THEN word.info
138 ELSE coalesce(word.info, '{}'::jsonb) || wf.info
140 FROM word LEFT JOIN word_frequencies wf
141 ON word.word_id = wf.id
143 drop_tables(conn, 'word_frequencies')
145 with conn.cursor() as cur:
146 cur.execute('SET max_parallel_workers_per_gather TO 0')
148 sqlp = SQLPreprocessor(conn, config)
149 sqlp.run_string(conn,
150 'GRANT SELECT ON tmp_word TO "{{config.DATABASE_WEBUSER}}"')
152 self._create_base_indices(config, 'tmp_word')
153 self._create_lookup_indices(config, 'tmp_word')
154 self._move_temporary_word_table('tmp_word')
156 def _cleanup_housenumbers(self) -> None:
157 """ Remove unused house numbers.
159 with connect(self.dsn) as conn:
160 if not table_exists(conn, 'search_name'):
162 with conn.cursor(name="hnr_counter") as cur:
163 cur.execute("""SELECT DISTINCT word_id, coalesce(info->>'lookup', word_token)
166 AND NOT EXISTS(SELECT * FROM search_name
167 WHERE ARRAY[word.word_id] && name_vector)
168 AND (char_length(coalesce(word, word_token)) > 6
169 OR coalesce(word, word_token) not similar to '\\d+')
171 candidates = {token: wid for wid, token in cur}
172 with conn.cursor(name="hnr_counter") as cur:
173 cur.execute("""SELECT housenumber FROM placex
174 WHERE housenumber is not null
175 AND (char_length(housenumber) > 6
176 OR housenumber not similar to '\\d+')
179 for hnr in row[0].split(';'):
180 candidates.pop(hnr, None)
181 LOG.info("There are %s outdated housenumbers.", len(candidates))
182 LOG.debug("Outdated housenumbers: %s", candidates.keys())
184 with conn.cursor() as cur:
185 cur.execute("""DELETE FROM word WHERE word_id = any(%s)""",
186 (list(candidates.values()), ))
189 def update_word_tokens(self) -> None:
190 """ Remove unused tokens.
192 LOG.warning("Cleaning up housenumber tokens.")
193 self._cleanup_housenumbers()
194 LOG.warning("Tokenizer house-keeping done.")
196 def name_analyzer(self) -> 'ICUNameAnalyzer':
197 """ Create a new analyzer for tokenizing names and queries
198 using this tokinzer. Analyzers are context managers and should
202 with tokenizer.name_analyzer() as analyzer:
206 When used outside the with construct, the caller must ensure to
207 call the close() function before destructing the analyzer.
209 Analyzers are not thread-safe. You need to instantiate one per thread.
211 assert self.loader is not None
212 return ICUNameAnalyzer(self.dsn, self.loader.make_sanitizer(),
213 self.loader.make_token_analysis())
215 def most_frequent_words(self, conn: Connection, num: int) -> List[str]:
216 """ Return a list of the `num` most frequent full words
219 with conn.cursor() as cur:
220 cur.execute("""SELECT word, sum((info->>'count')::int) as count
221 FROM word WHERE type = 'W'
223 ORDER BY count DESC LIMIT %s""", (num,))
224 return list(s[0].split('@')[0] for s in cur)
226 def _save_config(self) -> None:
227 """ Save the configuration that needs to remain stable for the given
228 database as database properties.
230 assert self.loader is not None
231 with connect(self.dsn) as conn:
232 self.loader.save_config_to_db(conn)
234 def _setup_db_tables(self, config: Configuration) -> None:
235 """ Set up the word table and fill it with pre-computed word
238 with connect(self.dsn) as conn:
239 drop_tables(conn, 'word')
240 sqlp = SQLPreprocessor(conn, config)
241 sqlp.run_string(conn, """
244 word_token text NOT NULL,
248 ) {{db.tablespace.search_data}};
249 GRANT SELECT ON word TO "{{config.DATABASE_WEBUSER}}";
251 DROP SEQUENCE IF EXISTS seq_word;
252 CREATE SEQUENCE seq_word start 1;
253 GRANT SELECT ON seq_word to "{{config.DATABASE_WEBUSER}}";
257 def _create_base_indices(self, config: Configuration, table_name: str) -> None:
258 """ Set up the word table and fill it with pre-computed word
261 with connect(self.dsn) as conn:
262 sqlp = SQLPreprocessor(conn, config)
263 sqlp.run_string(conn,
264 """CREATE INDEX idx_{{table_name}}_word_token ON {{table_name}}
265 USING BTREE (word_token) {{db.tablespace.search_index}}""",
266 table_name=table_name)
267 for name, ctype in WORD_TYPES:
268 sqlp.run_string(conn,
269 """CREATE INDEX idx_{{table_name}}_{{idx_name}} ON {{table_name}}
270 USING BTREE (word) {{db.tablespace.address_index}}
271 WHERE type = '{{column_type}}'
273 table_name=table_name, idx_name=name,
277 def _create_lookup_indices(self, config: Configuration, table_name: str) -> None:
278 """ Create additional indexes used when running the API.
280 with connect(self.dsn) as conn:
281 sqlp = SQLPreprocessor(conn, config)
282 # Index required for details lookup.
286 CREATE INDEX IF NOT EXISTS idx_{{table_name}}_word_id
287 ON {{table_name}} USING BTREE (word_id) {{db.tablespace.search_index}}
289 table_name=table_name)
292 def _move_temporary_word_table(self, old: str) -> None:
293 """ Rename all tables and indexes used by the tokenizer.
295 with connect(self.dsn) as conn:
296 drop_tables(conn, 'word')
297 with conn.cursor() as cur:
298 cur.execute(f"ALTER TABLE {old} RENAME TO word")
299 for idx in ('word_token', 'word_id'):
300 cur.execute(f"""ALTER INDEX idx_{old}_{idx}
301 RENAME TO idx_word_{idx}""")
302 for name, _ in WORD_TYPES:
303 cur.execute(f"""ALTER INDEX idx_{old}_{name}
304 RENAME TO idx_word_{name}""")
308 class ICUNameAnalyzer(AbstractAnalyzer):
309 """ The ICU analyzer uses the ICU library for splitting names.
311 Each instance opens a connection to the database to request the
315 def __init__(self, dsn: str, sanitizer: PlaceSanitizer,
316 token_analysis: ICUTokenAnalysis) -> None:
317 self.conn: Optional[Connection] = connect(dsn)
318 self.conn.autocommit = True
319 self.sanitizer = sanitizer
320 self.token_analysis = token_analysis
322 self._cache = _TokenCache()
324 def close(self) -> None:
325 """ Free all resources used by the analyzer.
331 def _search_normalized(self, name: str) -> str:
332 """ Return the search token transliteration of the given name.
334 return cast(str, self.token_analysis.search.transliterate(name)).strip()
336 def _normalized(self, name: str) -> str:
337 """ Return the normalized version of the given name with all
338 non-relevant information removed.
340 return cast(str, self.token_analysis.normalizer.transliterate(name)).strip()
342 def get_word_token_info(self, words: Sequence[str]) -> List[Tuple[str, str, int]]:
343 """ Return token information for the given list of words.
344 If a word starts with # it is assumed to be a full name
345 otherwise is a partial name.
347 The function returns a list of tuples with
348 (original word, word token, word id).
350 The function is used for testing and debugging only
351 and not necessarily efficient.
353 assert self.conn is not None
357 if word.startswith('#'):
358 full_tokens[word] = self._search_normalized(word[1:])
360 partial_tokens[word] = self._search_normalized(word)
362 with self.conn.cursor() as cur:
363 cur.execute("""SELECT word_token, word_id
364 FROM word WHERE word_token = ANY(%s) and type = 'W'
365 """, (list(full_tokens.values()),))
366 full_ids = {r[0]: r[1] for r in cur}
367 cur.execute("""SELECT word_token, word_id
368 FROM word WHERE word_token = ANY(%s) and type = 'w'""",
369 (list(partial_tokens.values()),))
370 part_ids = {r[0]: r[1] for r in cur}
372 return [(k, v, full_ids.get(v, None)) for k, v in full_tokens.items()] \
373 + [(k, v, part_ids.get(v, None)) for k, v in partial_tokens.items()]
375 def normalize_postcode(self, postcode: str) -> str:
376 """ Convert the postcode to a standardized form.
378 This function must yield exactly the same result as the SQL function
379 'token_normalized_postcode()'.
381 return postcode.strip().upper()
383 def update_postcodes_from_db(self) -> None:
386 Removes all postcodes from the word table because they are not
387 needed. Postcodes are recognised by pattern.
389 assert self.conn is not None
391 with self.conn.cursor() as cur:
392 cur.execute("DELETE FROM word WHERE type = 'P'")
394 def update_special_phrases(self, phrases: Iterable[Tuple[str, str, str, str]],
395 should_replace: bool) -> None:
396 """ Replace the search index for special phrases with the new phrases.
397 If `should_replace` is True, then the previous set of will be
398 completely replaced. Otherwise the phrases are added to the
399 already existing ones.
401 assert self.conn is not None
402 norm_phrases = set(((self._normalized(p[0]), p[1], p[2], p[3])
405 with self.conn.cursor() as cur:
406 # Get the old phrases.
407 existing_phrases = set()
408 cur.execute("SELECT word, info FROM word WHERE type = 'S'")
409 for word, info in cur:
410 existing_phrases.add((word, info['class'], info['type'],
411 info.get('op') or '-'))
413 added = self._add_special_phrases(cur, norm_phrases, existing_phrases)
415 deleted = self._remove_special_phrases(cur, norm_phrases,
420 LOG.info("Total phrases: %s. Added: %s. Deleted: %s",
421 len(norm_phrases), added, deleted)
423 def _add_special_phrases(self, cursor: Cursor,
424 new_phrases: Set[Tuple[str, str, str, str]],
425 existing_phrases: Set[Tuple[str, str, str, str]]) -> int:
426 """ Add all phrases to the database that are not yet there.
428 to_add = new_phrases - existing_phrases
431 with cursor.copy('COPY word(word_token, type, word, info) FROM STDIN') as copy:
432 for word, cls, typ, oper in to_add:
433 term = self._search_normalized(word)
435 copy.write_row((term, 'S', word,
436 Jsonb({'class': cls, 'type': typ,
437 'op': oper if oper in ('in', 'near') else None})))
442 def _remove_special_phrases(self, cursor: Cursor,
443 new_phrases: Set[Tuple[str, str, str, str]],
444 existing_phrases: Set[Tuple[str, str, str, str]]) -> int:
445 """ Remove all phrases from the database that are no longer in the
448 to_delete = existing_phrases - new_phrases
453 WHERE type = 'S' and word = %s
454 and info->>'class' = %s and info->>'type' = %s
455 and %s = coalesce(info->>'op', '-')
458 return len(to_delete)
460 def add_country_names(self, country_code: str, names: Mapping[str, str]) -> None:
461 """ Add default names for the given country to the search index.
463 # Make sure any name preprocessing for country names applies.
464 info = PlaceInfo({'name': names, 'country_code': country_code,
465 'rank_address': 4, 'class': 'boundary',
466 'type': 'administrative'})
467 self._add_country_full_names(country_code,
468 self.sanitizer.process_names(info)[0],
471 def _add_country_full_names(self, country_code: str, names: Sequence[PlaceName],
472 internal: bool = False) -> None:
473 """ Add names for the given country from an already sanitized
476 assert self.conn is not None
479 norm_name = self._search_normalized(name.name)
481 word_tokens.add(norm_name)
483 with self.conn.cursor() as cur:
485 cur.execute("""SELECT word_token, coalesce(info ? 'internal', false) as is_internal
487 WHERE type = 'C' and word = %s""",
489 # internal/external names
490 existing_tokens: Dict[bool, Set[str]] = {True: set(), False: set()}
492 existing_tokens[word[1]].add(word[0])
494 # Delete names that no longer exist.
495 gone_tokens = existing_tokens[internal] - word_tokens
497 gone_tokens.update(existing_tokens[False] & word_tokens)
499 cur.execute("""DELETE FROM word
500 USING unnest(%s::text[]) as token
501 WHERE type = 'C' and word = %s
502 and word_token = token""",
503 (list(gone_tokens), country_code))
505 # Only add those names that are not yet in the list.
506 new_tokens = word_tokens - existing_tokens[True]
508 new_tokens -= existing_tokens[False]
511 sql = """INSERT INTO word (word_token, type, word, info)
512 (SELECT token, 'C', %s, '{"internal": "yes"}'
513 FROM unnest(%s::text[]) as token)
516 sql = """INSERT INTO word (word_token, type, word)
517 (SELECT token, 'C', %s
518 FROM unnest(%s::text[]) as token)
520 cur.execute(sql, (country_code, list(new_tokens)))
522 def process_place(self, place: PlaceInfo) -> Mapping[str, Any]:
523 """ Determine tokenizer information about the given place.
525 Returns a JSON-serializable structure that will be handed into
526 the database via the token_info field.
528 token_info = _TokenInfo()
530 names, address = self.sanitizer.process_names(place)
533 token_info.set_names(*self._compute_name_tokens(names))
535 if place.is_country():
536 assert place.country_code is not None
537 self._add_country_full_names(place.country_code, names)
540 self._process_place_address(token_info, address)
542 return token_info.to_dict()
544 def _process_place_address(self, token_info: '_TokenInfo',
545 address: Sequence[PlaceName]) -> None:
547 if item.kind == 'postcode':
548 token_info.set_postcode(self._add_postcode(item))
549 elif item.kind == 'housenumber':
550 token_info.add_housenumber(*self._compute_housenumber_token(item))
551 elif item.kind == 'street':
552 token_info.add_street(self._retrieve_full_tokens(item.name))
553 elif item.kind == 'place':
555 token_info.add_place(itertools.chain(*self._compute_name_tokens([item])))
556 elif (not item.kind.startswith('_') and not item.suffix and
557 item.kind not in ('country', 'full', 'inclusion')):
558 token_info.add_address_term(item.kind,
559 itertools.chain(*self._compute_name_tokens([item])))
561 def _compute_housenumber_token(self, hnr: PlaceName) -> Tuple[Optional[int], Optional[str]]:
562 """ Normalize the housenumber and return the word token and the
565 assert self.conn is not None
566 analyzer = self.token_analysis.analysis.get('@housenumber')
567 result: Tuple[Optional[int], Optional[str]] = (None, None)
570 # When no custom analyzer is set, simply normalize and transliterate
571 norm_name = self._search_normalized(hnr.name)
573 result = self._cache.housenumbers.get(norm_name, result)
574 if result[0] is None:
575 hid = execute_scalar(self.conn, "SELECT getorcreate_hnr_id(%s)", (norm_name, ))
577 result = hid, norm_name
578 self._cache.housenumbers[norm_name] = result
580 # Otherwise use the analyzer to determine the canonical name.
581 # Per convention we use the first variant as the 'lookup name', the
582 # name that gets saved in the housenumber field of the place.
583 word_id = analyzer.get_canonical_id(hnr)
585 result = self._cache.housenumbers.get(word_id, result)
586 if result[0] is None:
587 variants = analyzer.compute_variants(word_id)
589 hid = execute_scalar(self.conn, "SELECT create_analyzed_hnr_id(%s, %s)",
590 (word_id, list(variants)))
591 result = hid, variants[0]
592 self._cache.housenumbers[word_id] = result
596 def _retrieve_full_tokens(self, name: str) -> List[int]:
597 """ Get the full name token for the given name, if it exists.
598 The name is only retrieved for the standard analyser.
600 assert self.conn is not None
601 norm_name = self._search_normalized(name)
603 # return cached if possible
604 if norm_name in self._cache.fulls:
605 return self._cache.fulls[norm_name]
607 with self.conn.cursor() as cur:
608 cur.execute("SELECT word_id FROM word WHERE word_token = %s and type = 'W'",
610 full = [row[0] for row in cur]
612 self._cache.fulls[norm_name] = full
616 def _compute_name_tokens(self, names: Sequence[PlaceName]) -> Tuple[Set[int], Set[int]]:
617 """ Computes the full name and partial name tokens for the given
620 assert self.conn is not None
621 full_tokens: Set[int] = set()
622 partial_tokens: Set[int] = set()
625 analyzer_id = name.get_attr('analyzer')
626 analyzer = self.token_analysis.get_analyzer(analyzer_id)
627 word_id = analyzer.get_canonical_id(name)
628 if analyzer_id is None:
631 token_id = f'{word_id}@{analyzer_id}'
633 full, part = self._cache.names.get(token_id, (None, None))
635 variants = analyzer.compute_variants(word_id)
639 with self.conn.cursor() as cur:
640 cur.execute("SELECT * FROM getorcreate_full_word(%s, %s)",
641 (token_id, variants))
642 full, part = cast(Tuple[int, List[int]], cur.fetchone())
644 self._cache.names[token_id] = (full, part)
646 assert part is not None
648 full_tokens.add(full)
649 partial_tokens.update(part)
651 return full_tokens, partial_tokens
653 def _add_postcode(self, item: PlaceName) -> Optional[str]:
654 """ Make sure the normalized postcode is present in the word table.
656 assert self.conn is not None
657 analyzer = self.token_analysis.analysis.get('@postcode')
660 return item.name.strip().upper()
662 return analyzer.get_canonical_id(item)
666 """ Collect token information to be sent back to the database.
668 def __init__(self) -> None:
669 self.names: Optional[str] = None
670 self.housenumbers: Set[str] = set()
671 self.housenumber_tokens: Set[int] = set()
672 self.street_tokens: Optional[Set[int]] = None
673 self.place_tokens: Set[int] = set()
674 self.address_tokens: Dict[str, str] = {}
675 self.postcode: Optional[str] = None
677 def _mk_array(self, tokens: Iterable[Any]) -> str:
678 return f"{{{','.join((str(s) for s in tokens))}}}"
680 def to_dict(self) -> Dict[str, Any]:
681 """ Return the token information in database importable format.
683 out: Dict[str, Any] = {}
686 out['names'] = self.names
688 if self.housenumbers:
689 out['hnr'] = ';'.join(self.housenumbers)
690 out['hnr_tokens'] = self._mk_array(self.housenumber_tokens)
692 if self.street_tokens is not None:
693 out['street'] = self._mk_array(self.street_tokens)
695 if self.place_tokens:
696 out['place'] = self._mk_array(self.place_tokens)
698 if self.address_tokens:
699 out['addr'] = self.address_tokens
702 out['postcode'] = self.postcode
706 def set_names(self, fulls: Iterable[int], partials: Iterable[int]) -> None:
707 """ Adds token information for the normalised names.
709 self.names = self._mk_array(itertools.chain(fulls, partials))
711 def add_housenumber(self, token: Optional[int], hnr: Optional[str]) -> None:
712 """ Extract housenumber information from a list of normalised
716 assert hnr is not None
717 self.housenumbers.add(hnr)
718 self.housenumber_tokens.add(token)
720 def add_street(self, tokens: Iterable[int]) -> None:
721 """ Add addr:street match terms.
723 if self.street_tokens is None:
724 self.street_tokens = set()
725 self.street_tokens.update(tokens)
727 def add_place(self, tokens: Iterable[int]) -> None:
728 """ Add addr:place search and match terms.
730 self.place_tokens.update(tokens)
732 def add_address_term(self, key: str, partials: Iterable[int]) -> None:
733 """ Add additional address terms.
735 array = self._mk_array(partials)
737 self.address_tokens[key] = array
739 def set_postcode(self, postcode: Optional[str]) -> None:
740 """ Set the postcode to the given one.
742 self.postcode = postcode
746 """ Cache for token information to avoid repeated database queries.
748 This cache is not thread-safe and needs to be instantiated per
751 def __init__(self) -> None:
752 self.names: Dict[str, Tuple[int, List[int]]] = {}
753 self.partials: Dict[str, int] = {}
754 self.fulls: Dict[str, List[int]] = {}
755 self.housenumbers: Dict[str, Tuple[Optional[int], Optional[str]]] = {}