nominatim/api/search/db_search_builder.py

   1 # SPDX-License-Identifier: GPL-3.0-or-later
   2 #
   3 # This file is part of Nominatim. (https://nominatim.org)
   4 #
   5 # Copyright (C) 2023 by the Nominatim developer community.
   6 # For a full list of authors see the git log.
   7 """
   8 Convertion from token assignment to an abstract DB search.
   9 """
  10 from typing import Optional, List, Tuple, Iterator
  11 import heapq
  12
  13 from nominatim.api.types import SearchDetails, DataLayer
  14 from nominatim.api.search.query import QueryStruct, Token, TokenType, TokenRange, BreakType
  15 from nominatim.api.search.token_assignment import TokenAssignment
  16 import nominatim.api.search.db_search_fields as dbf
  17 import nominatim.api.search.db_searches as dbs
  18
  19
  20 def wrap_near_search(categories: List[Tuple[str, str]],
  21                      search: dbs.AbstractSearch) -> dbs.NearSearch:
  22     """ Create a new search that wraps the given search in a search
  23         for near places of the given category.
  24     """
  25     return dbs.NearSearch(penalty=search.penalty,
  26                           categories=dbf.WeightedCategories(categories,
  27                                                             [0.0] * len(categories)),
  28                           search=search)
  29
  30
  31 def build_poi_search(category: List[Tuple[str, str]],
  32                      countries: Optional[List[str]]) -> dbs.PoiSearch:
  33     """ Create a new search for places by the given category, possibly
  34         constraint to the given countries.
  35     """
  36     if countries:
  37         ccs = dbf.WeightedStrings(countries, [0.0] * len(countries))
  38     else:
  39         ccs = dbf.WeightedStrings([], [])
  40
  41     class _PoiData(dbf.SearchData):
  42         penalty = 0.0
  43         qualifiers = dbf.WeightedCategories(category, [0.0] * len(category))
  44         countries=ccs
  45
  46     return dbs.PoiSearch(_PoiData())
  47
  48
  49 class SearchBuilder:
  50     """ Build the abstract search queries from token assignments.
  51     """
  52
  53     def __init__(self, query: QueryStruct, details: SearchDetails) -> None:
  54         self.query = query
  55         self.details = details
  56
  57
  58     @property
  59     def configured_for_country(self) -> bool:
  60         """ Return true if the search details are configured to
  61             allow countries in the result.
  62         """
  63         return self.details.min_rank <= 4 and self.details.max_rank >= 4 \
  64                and self.details.layer_enabled(DataLayer.ADDRESS)
  65
  66
  67     @property
  68     def configured_for_postcode(self) -> bool:
  69         """ Return true if the search details are configured to
  70             allow postcodes in the result.
  71         """
  72         return self.details.min_rank <= 5 and self.details.max_rank >= 11\
  73                and self.details.layer_enabled(DataLayer.ADDRESS)
  74
  75
  76     @property
  77     def configured_for_housenumbers(self) -> bool:
  78         """ Return true if the search details are configured to
  79             allow addresses in the result.
  80         """
  81         return self.details.max_rank >= 30 \
  82                and self.details.layer_enabled(DataLayer.ADDRESS)
  83
  84
  85     def build(self, assignment: TokenAssignment) -> Iterator[dbs.AbstractSearch]:
  86         """ Yield all possible abstract searches for the given token assignment.
  87         """
  88         sdata = self.get_search_data(assignment)
  89         if sdata is None:
  90             return
  91
  92         categories = self.get_search_categories(assignment)
  93
  94         if assignment.name is None:
  95             if categories and not sdata.postcodes:
  96                 sdata.qualifiers = categories
  97                 categories = None
  98                 builder = self.build_poi_search(sdata)
  99             elif assignment.housenumber:
 100                 hnr_tokens = self.query.get_tokens(assignment.housenumber,
 101                                                    TokenType.HOUSENUMBER)
 102                 builder = self.build_housenumber_search(sdata, hnr_tokens, assignment.address)
 103             else:
 104                 builder = self.build_special_search(sdata, assignment.address,
 105                                                     bool(categories))
 106         else:
 107             builder = self.build_name_search(sdata, assignment.name, assignment.address,
 108                                              bool(categories))
 109
 110         if categories:
 111             penalty = min(categories.penalties)
 112             categories.penalties = [p - penalty for p in categories.penalties]
 113             for search in builder:
 114                 yield dbs.NearSearch(penalty, categories, search)
 115         else:
 116             yield from builder
 117
 118
 119     def build_poi_search(self, sdata: dbf.SearchData) -> Iterator[dbs.AbstractSearch]:
 120         """ Build abstract search query for a simple category search.
 121             This kind of search requires an additional geographic constraint.
 122         """
 123         if not sdata.housenumbers \
 124            and ((self.details.viewbox and self.details.bounded_viewbox) or self.details.near):
 125             yield dbs.PoiSearch(sdata)
 126
 127
 128     def build_special_search(self, sdata: dbf.SearchData,
 129                              address: List[TokenRange],
 130                              is_category: bool) -> Iterator[dbs.AbstractSearch]:
 131         """ Build abstract search queries for searches that do not involve
 132             a named place.
 133         """
 134         if sdata.qualifiers:
 135             # No special searches over qualifiers supported.
 136             return
 137
 138         if sdata.countries and not address and not sdata.postcodes \
 139            and self.configured_for_country:
 140             yield dbs.CountrySearch(sdata)
 141
 142         if sdata.postcodes and (is_category or self.configured_for_postcode):
 143             penalty = 0.0 if sdata.countries else 0.1
 144             if address:
 145                 sdata.lookups = [dbf.FieldLookup('nameaddress_vector',
 146                                                  [t.token for r in address
 147                                                   for t in self.query.get_partials_list(r)],
 148                                                  'restrict')]
 149                 penalty += 0.2
 150             yield dbs.PostcodeSearch(penalty, sdata)
 151
 152
 153     def build_housenumber_search(self, sdata: dbf.SearchData, hnrs: List[Token],
 154                                  address: List[TokenRange]) -> Iterator[dbs.AbstractSearch]:
 155         """ Build a simple address search for special entries where the
 156             housenumber is the main name token.
 157         """
 158         partial_tokens: List[int] = []
 159         for trange in address:
 160             partial_tokens.extend(t.token for t in self.query.get_partials_list(trange))
 161
 162         sdata.lookups = [dbf.FieldLookup('name_vector', [t.token for t in hnrs], 'lookup_any'),
 163                          dbf.FieldLookup('nameaddress_vector', partial_tokens, 'lookup_all')
 164                         ]
 165         sdata.housenumbers = dbf.WeightedStrings([], [])
 166         yield dbs.PlaceSearch(0.05, sdata, sum(t.count for t in hnrs))
 167
 168
 169     def build_name_search(self, sdata: dbf.SearchData,
 170                           name: TokenRange, address: List[TokenRange],
 171                           is_category: bool) -> Iterator[dbs.AbstractSearch]:
 172         """ Build abstract search queries for simple name or address searches.
 173         """
 174         if is_category or not sdata.housenumbers or self.configured_for_housenumbers:
 175             ranking = self.get_name_ranking(name)
 176             name_penalty = ranking.normalize_penalty()
 177             if ranking.rankings:
 178                 sdata.rankings.append(ranking)
 179             for penalty, count, lookup in self.yield_lookups(name, address):
 180                 sdata.lookups = lookup
 181                 yield dbs.PlaceSearch(penalty + name_penalty, sdata, count)
 182
 183
 184     def yield_lookups(self, name: TokenRange, address: List[TokenRange])\
 185                           -> Iterator[Tuple[float, int, List[dbf.FieldLookup]]]:
 186         """ Yield all variants how the given name and address should best
 187             be searched for. This takes into account how frequent the terms
 188             are and tries to find a lookup that optimizes index use.
 189         """
 190         penalty = 0.0 # extra penalty
 191         name_partials = self.query.get_partials_list(name)
 192         name_tokens = [t.token for t in name_partials]
 193
 194         addr_partials = [t for r in address for t in self.query.get_partials_list(r)]
 195         addr_tokens = [t.token for t in addr_partials]
 196
 197         partials_indexed = all(t.is_indexed for t in name_partials) \
 198                            and all(t.is_indexed for t in addr_partials)
 199         exp_count = min(t.count for t in name_partials)
 200
 201         if (len(name_partials) > 3 or exp_count < 1000) and partials_indexed:
 202             yield penalty, exp_count, dbf.lookup_by_names(name_tokens, addr_tokens)
 203             return
 204
 205         exp_count = min(exp_count, min(t.count for t in addr_partials)) \
 206                     if addr_partials else exp_count
 207         if exp_count < 1000 and partials_indexed:
 208             # Lookup by address partials and restrict results through name terms.
 209             # Give this a small penalty because lookups in the address index are
 210             # more expensive
 211             yield penalty + exp_count/5000, exp_count,\
 212                   dbf.lookup_by_addr(name_tokens, addr_tokens)
 213             return
 214
 215         # Partial term to frequent. Try looking up by rare full names first.
 216         name_fulls = self.query.get_tokens(name, TokenType.WORD)
 217         rare_names = list(filter(lambda t: t.count < 10000, name_fulls))
 218         # At this point drop unindexed partials from the address.
 219         # This might yield wrong results, nothing we can do about that.
 220         if not partials_indexed:
 221             addr_tokens = [t.token for t in addr_partials if t.is_indexed]
 222             penalty += 1.2 * sum(t.penalty for t in addr_partials if not t.is_indexed)
 223         if rare_names:
 224             # Any of the full names applies with all of the partials from the address
 225             yield penalty, sum(t.count for t in rare_names),\
 226                   dbf.lookup_by_any_name([t.token for t in rare_names], addr_tokens)
 227
 228         # To catch remaining results, lookup by name and address
 229         # We only do this if there is a reasonable number of results expected.
 230         if exp_count < 10000:
 231             if all(t.is_indexed for t in name_partials):
 232                 lookup = [dbf.FieldLookup('name_vector', name_tokens, 'lookup_all')]
 233             else:
 234                 # we don't have the partials, try with the non-rare names
 235                 non_rare_names = [t.token for t in name_fulls if t.count >= 10000]
 236                 if not non_rare_names:
 237                     return
 238                 lookup = [dbf.FieldLookup('name_vector', non_rare_names, 'lookup_any')]
 239             if addr_tokens:
 240                 lookup.append(dbf.FieldLookup('nameaddress_vector', addr_tokens, 'lookup_all'))
 241             penalty += 0.1 * max(0, 5 - len(name_partials) - len(addr_tokens))
 242             if len(rare_names) == len(name_fulls):
 243                 # if there already was a search for all full tokens,
 244                 # avoid this if anything has been found
 245                 penalty += 0.25
 246             yield penalty, exp_count, lookup
 247
 248
 249     def get_name_ranking(self, trange: TokenRange) -> dbf.FieldRanking:
 250         """ Create a ranking expression for a name term in the given range.
 251         """
 252         name_fulls = self.query.get_tokens(trange, TokenType.WORD)
 253         ranks = [dbf.RankedTokens(t.penalty, [t.token]) for t in name_fulls]
 254         ranks.sort(key=lambda r: r.penalty)
 255         # Fallback, sum of penalty for partials
 256         name_partials = self.query.get_partials_list(trange)
 257         default = sum(t.penalty for t in name_partials) + 0.2
 258         return dbf.FieldRanking('name_vector', default, ranks)
 259
 260
 261     def get_addr_ranking(self, trange: TokenRange) -> dbf.FieldRanking:
 262         """ Create a list of ranking expressions for an address term
 263             for the given ranges.
 264         """
 265         todo: List[Tuple[int, int, dbf.RankedTokens]] = []
 266         heapq.heappush(todo, (0, trange.start, dbf.RankedTokens(0.0, [])))
 267         ranks: List[dbf.RankedTokens] = []
 268
 269         while todo: # pylint: disable=too-many-nested-blocks
 270             neglen, pos, rank = heapq.heappop(todo)
 271             for tlist in self.query.nodes[pos].starting:
 272                 if tlist.ttype in (TokenType.PARTIAL, TokenType.WORD):
 273                     if tlist.end < trange.end:
 274                         chgpenalty = PENALTY_WORDCHANGE[self.query.nodes[tlist.end].btype]
 275                         if tlist.ttype == TokenType.PARTIAL:
 276                             penalty = rank.penalty + chgpenalty \
 277                                       + max(t.penalty for t in tlist.tokens)
 278                             heapq.heappush(todo, (neglen - 1, tlist.end,
 279                                                   dbf.RankedTokens(penalty, rank.tokens)))
 280                         else:
 281                             for t in tlist.tokens:
 282                                 heapq.heappush(todo, (neglen - 1, tlist.end,
 283                                                       rank.with_token(t, chgpenalty)))
 284                     elif tlist.end == trange.end:
 285                         if tlist.ttype == TokenType.PARTIAL:
 286                             ranks.append(dbf.RankedTokens(rank.penalty
 287                                                           + max(t.penalty for t in tlist.tokens),
 288                                                           rank.tokens))
 289                         else:
 290                             ranks.extend(rank.with_token(t, 0.0) for t in tlist.tokens)
 291                         if len(ranks) >= 10:
 292                             # Too many variants, bail out and only add
 293                             # Worst-case Fallback: sum of penalty of partials
 294                             name_partials = self.query.get_partials_list(trange)
 295                             default = sum(t.penalty for t in name_partials) + 0.2
 296                             ranks.append(dbf.RankedTokens(rank.penalty + default, []))
 297                             # Bail out of outer loop
 298                             todo.clear()
 299                             break
 300
 301         ranks.sort(key=lambda r: len(r.tokens))
 302         default = ranks[0].penalty + 0.3
 303         del ranks[0]
 304         ranks.sort(key=lambda r: r.penalty)
 305
 306         return dbf.FieldRanking('nameaddress_vector', default, ranks)
 307
 308
 309     def get_search_data(self, assignment: TokenAssignment) -> Optional[dbf.SearchData]:
 310         """ Collect the tokens for the non-name search fields in the
 311             assignment.
 312         """
 313         sdata = dbf.SearchData()
 314         sdata.penalty = assignment.penalty
 315         if assignment.country:
 316             tokens = self.query.get_tokens(assignment.country, TokenType.COUNTRY)
 317             if self.details.countries:
 318                 tokens = [t for t in tokens if t.lookup_word in self.details.countries]
 319                 if not tokens:
 320                     return None
 321             sdata.set_strings('countries', tokens)
 322         elif self.details.countries:
 323             sdata.countries = dbf.WeightedStrings(self.details.countries,
 324                                                   [0.0] * len(self.details.countries))
 325         if assignment.housenumber:
 326             sdata.set_strings('housenumbers',
 327                               self.query.get_tokens(assignment.housenumber,
 328                                                     TokenType.HOUSENUMBER))
 329         if assignment.postcode:
 330             sdata.set_strings('postcodes',
 331                               self.query.get_tokens(assignment.postcode,
 332                                                     TokenType.POSTCODE))
 333         if assignment.qualifier:
 334             sdata.set_qualifiers(self.query.get_tokens(assignment.qualifier,
 335                                                        TokenType.QUALIFIER))
 336
 337         if assignment.address:
 338             sdata.set_ranking([self.get_addr_ranking(r) for r in assignment.address])
 339         else:
 340             sdata.rankings = []
 341
 342         return sdata
 343
 344
 345     def get_search_categories(self,
 346                               assignment: TokenAssignment) -> Optional[dbf.WeightedCategories]:
 347         """ Collect tokens for category search or use the categories
 348             requested per parameter.
 349             Returns None if no category search is requested.
 350         """
 351         if assignment.category:
 352             tokens = [t for t in self.query.get_tokens(assignment.category,
 353                                                        TokenType.CATEGORY)
 354                       if not self.details.categories
 355                          or t.get_category() in self.details.categories]
 356             return dbf.WeightedCategories([t.get_category() for t in tokens],
 357                                           [t.penalty for t in tokens])
 358
 359         if self.details.categories:
 360             return dbf.WeightedCategories(self.details.categories,
 361                                           [0.0] * len(self.details.categories))
 362
 363         return None
 364
 365
 366 PENALTY_WORDCHANGE = {
 367     BreakType.START: 0.0,
 368     BreakType.END: 0.0,
 369     BreakType.PHRASE: 0.0,
 370     BreakType.WORD: 0.1,
 371     BreakType.PART: 0.2,
 372     BreakType.TOKEN: 0.4
 373 }