pokedex/lookup.py

   1 # encoding: utf8
   2 from collections import namedtuple
   3 import os, os.path
   4 import pkg_resources
   5 import random
   6 import re
   7 import shutil
   8 import unicodedata
   9
  10 from sqlalchemy.sql import func
  11 import whoosh
  12 import whoosh.filedb.filestore
  13 import whoosh.filedb.fileindex
  14 import whoosh.index
  15 from whoosh.qparser import QueryParser
  16 import whoosh.scoring
  17 import whoosh.spelling
  18
  19 from pokedex.db import connect
  20 import pokedex.db.tables as tables
  21 from pokedex.roomaji import romanize
  22
  23 __all__ = ['PokedexLookup']
  24
  25
  26 rx_is_number = re.compile('^\d+$')
  27
  28 LookupResult = namedtuple('LookupResult',
  29     ['object', 'indexed_name', 'name', 'language', 'iso3166', 'exact'])
  30
  31 class LanguageWeighting(whoosh.scoring.Weighting):
  32     """A scoring class that forces otherwise-equal English results to come
  33     before foreign results.
  34     """
  35
  36     def score(self, searcher, fieldnum, text, docnum, weight, QTF=1):
  37         doc = searcher.stored_fields(docnum)
  38         if doc['language'] == None:
  39             # English (well, "default"); leave it at 1
  40             return weight
  41         elif doc['language'] == u'Roomaji':
  42             # Give Roomaji a little boost; it's most likely to be searched
  43             return weight * 0.95
  44         else:
  45             # Everything else can drop down the totem pole
  46             return weight * 0.9
  47
  48
  49 class PokedexLookup(object):
  50     INTERMEDIATE_LOOKUP_RESULTS = 25
  51     MAX_LOOKUP_RESULTS = 10
  52
  53     # Dictionary of table name => table class.
  54     # Need the table name so we can get the class from the table name after we
  55     # retrieve something from the index
  56     indexed_tables = dict(
  57         (cls.__tablename__, cls)
  58         for cls in (
  59             tables.Ability,
  60             tables.Item,
  61             tables.Location,
  62             tables.Move,
  63             tables.Nature,
  64             tables.Pokemon,
  65             tables.Type,
  66         )
  67     )
  68
  69
  70     def __init__(self, directory=None, session=None, recreate=False):
  71         """Opens the whoosh index stored in the named directory.  If the index
  72         doesn't already exist, it will be created.
  73
  74         `directory`
  75             Directory containing the index.  Defaults to a location within the
  76             `pokedex` egg directory.
  77
  78         `session`
  79             If the index needs to be created, this database session will be
  80             used.  Defaults to an attempt to connect to the default SQLite
  81             database installed by `pokedex setup`.
  82
  83         `recreate`
  84             If set to True, the whoosh index will be created even if it already
  85             exists.
  86         """
  87
  88         # By the time this returns, self.index, self.speller, and self.session
  89         # must be set
  90
  91         # Defaults
  92         if not directory:
  93             directory = pkg_resources.resource_filename('pokedex',
  94                                                         'data/whoosh-index')
  95
  96         if session:
  97             self.session = session
  98         else:
  99             self.session = connect()
 100
 101         # Attempt to open or create the index
 102         directory_exists = os.path.exists(directory)
 103         if directory_exists and not recreate:
 104             # Already exists; should be an index!  Bam, done.
 105             try:
 106                 self.index = whoosh.index.open_dir(directory, indexname='MAIN')
 107                 spell_store = whoosh.filedb.filestore.FileStorage(directory)
 108                 self.speller = whoosh.spelling.SpellChecker(spell_store)
 109                 return
 110             except whoosh.index.EmptyIndexError as e:
 111                 # Apparently not a real index.  Fall out and create it
 112                 pass
 113
 114         # Delete and start over if we're going to bail anyway.
 115         if directory_exists and recreate:
 116             # Be safe and only delete if it looks like a whoosh index, i.e.,
 117             # everything starts with _
 118             if all(f[0] == '_' for f in os.listdir(directory)):
 119                 shutil.rmtree(directory)
 120                 directory_exists = False
 121
 122         if not directory_exists:
 123             os.mkdir(directory)
 124
 125
 126         ### Create index
 127         schema = whoosh.fields.Schema(
 128             name=whoosh.fields.ID(stored=True),
 129             table=whoosh.fields.ID(stored=True),
 130             row_id=whoosh.fields.ID(stored=True),
 131             language=whoosh.fields.STORED,
 132             iso3166=whoosh.fields.STORED,
 133             display_name=whoosh.fields.STORED,  # non-lowercased name
 134         )
 135
 136         self.index = whoosh.index.create_in(directory, schema=schema,
 137                                             indexname='MAIN')
 138         writer = self.index.writer()
 139
 140         # Index every name in all our tables of interest
 141         # speller_entries becomes a list of (word, score) tuples; the score is
 142         # 2 for English names, 1.5 for Roomaji, and 1 for everything else.  I
 143         # think this biases the results in the direction most people expect,
 144         # especially when e.g. German names are very similar to English names
 145         speller_entries = []
 146         for cls in self.indexed_tables.values():
 147             q = self.session.query(cls)
 148
 149             for row in q.yield_per(5):
 150                 row_key = dict(table=unicode(cls.__tablename__),
 151                                row_id=unicode(row.id))
 152
 153                 def add(name, language, iso3166, score):
 154                     normalized_name = self.normalize_name(name)
 155
 156                     writer.add_document(
 157                         name=normalized_name, display_name=name,
 158                         language=language, iso3166=iso3166,
 159                         **row_key
 160                     )
 161
 162                     speller_entries.append((normalized_name, score))
 163
 164
 165                 # Add the basic English name to the index
 166                 if cls == tables.Pokemon:
 167                     # Pokémon need their form name added
 168                     # XXX kinda kludgy
 169                     add(row.full_name, None, u'us', 1)
 170
 171                     # If this is a default form, ALSO add the unadorned name,
 172                     # so 'Deoxys' alone will still do the right thing
 173                     if row.forme_name and not row.forme_base_pokemon_id:
 174                         add(row.name, None, u'us', 1)
 175                 else:
 176                     add(row.name, None, u'us', 1)
 177
 178                 # Some things also have other languages' names
 179                 # XXX other language form names..?
 180                 for foreign_name in getattr(row, 'foreign_names', []):
 181                     moonspeak = foreign_name.name
 182                     if row.name == moonspeak:
 183                         # Don't add the English name again as a different
 184                         # language; no point and it makes spell results
 185                         # confusing
 186                         continue
 187
 188                     add(moonspeak, foreign_name.language.name,
 189                                    foreign_name.language.iso3166,
 190                                    3)
 191
 192                     # Add Roomaji too
 193                     if foreign_name.language.name == 'Japanese':
 194                         roomaji = romanize(foreign_name.name)
 195                         add(roomaji, u'Roomaji', u'jp', 8)
 196
 197         writer.commit()
 198
 199         # Construct and populate a spell-checker index.  Quicker to do it all
 200         # at once, as every call to add_* does a commit(), and those seem to be
 201         # expensive
 202         self.speller = whoosh.spelling.SpellChecker(self.index.storage)
 203         self.speller.add_scored_words(speller_entries)
 204
 205
 206     def normalize_name(self, name):
 207         """Strips irrelevant formatting junk from name input.
 208
 209         Specifically: everything is lowercased, and accents are removed.
 210         """
 211         # http://stackoverflow.com/questions/517923/what-is-the-best-way-to-remove-accents-in-a-python-unicode-string
 212         # Makes sense to me.  Decompose by Unicode rules, then remove combining
 213         # characters, then recombine.  I'm explicitly doing it this way instead
 214         # of testing combining() because Korean characters apparently
 215         # decompose!  But the results are considered letters, not combining
 216         # characters, so testing for Mn works well, and combining them again
 217         # makes them look right.
 218         nkfd_form = unicodedata.normalize('NFKD', unicode(name))
 219         name = u"".join(c for c in nkfd_form
 220                         if unicodedata.category(c) != 'Mn')
 221         name = unicodedata.normalize('NFC', name)
 222
 223         name = name.strip()
 224         name = name.lower()
 225
 226         return name
 227
 228
 229     def _parse_table_name(self, name):
 230         """Takes a singular table name, table name, or table object and returns
 231         the table name.
 232
 233         Returns None for a bogus name.
 234         """
 235         if hasattr(name, '__tablename__'):
 236             return getattr(name, '__tablename__')
 237         elif name in self.indexed_tables:
 238             return name
 239         elif name + 's' in self.indexed_tables:
 240             return name + 's'
 241         else:
 242             # Bogus.  Be nice and return dummy
 243             return None
 244
 245     def _whoosh_records_to_results(self, records, exact=True):
 246         """Converts a list of whoosh's indexed records to LookupResult tuples
 247         containing database objects.
 248         """
 249         # XXX this 'exact' thing is getting kinda leaky.  would like a better
 250         # way to handle it, since only lookup() cares about fuzzy results
 251         seen = {}
 252         results = []
 253         for record in records:
 254             # Skip dupes
 255             seen_key = record['table'], record['row_id']
 256             if seen_key in seen:
 257                 continue
 258             seen[seen_key] = True
 259
 260             cls = self.indexed_tables[record['table']]
 261             obj = self.session.query(cls).get(record['row_id'])
 262
 263             results.append(LookupResult(object=obj,
 264                                         indexed_name=record['name'],
 265                                         name=record['display_name'],
 266                                         language=record['language'],
 267                                         iso3166=record['iso3166'],
 268                                         exact=exact))
 269
 270         return results
 271
 272
 273     def lookup(self, input, valid_types=[], exact_only=False):
 274         """Attempts to find some sort of object, given a name.
 275
 276         Returns a list of named (object, name, language, iso3166, exact)
 277         tuples.  `object` is a database object, `name` is the name under which
 278         the object was found, `language` and `iso3166` are the name and country
 279         code of the language in which the name was found, and `exact` is True
 280         iff this was an
 281         exact match.
 282
 283         This function currently ONLY does fuzzy matching if there are no exact
 284         matches.
 285
 286         Formes are not returned unless requested; "Shaymin" will return only
 287         grass Shaymin.
 288
 289         Extraneous whitespace is removed with extreme prejudice.
 290
 291         Recognizes:
 292         - Names: "Eevee", "Surf", "Run Away", "Payapa Berry", etc.
 293         - Foreign names: "Iibui", "Eivui"
 294         - Fuzzy names in whatever language: "Evee", "Ibui"
 295         - IDs: "133", "192", "250"
 296         Also:
 297         - Type restrictions.  "type:psychic" will only return the type.  This
 298           is how to make ID lookup useful.  Multiple type specs can be entered
 299           with commas, as "move,item:1".  If `valid_types` are provided, any
 300           type prefix will be ignored.
 301         - Alternate formes can be specified merely like "wash rotom".
 302
 303         `input`
 304             Name of the thing to look for.
 305
 306         `valid_types`
 307             A list of table objects or names, e.g., `['pokemon', 'moves']`.  If
 308             this is provided, only results in one of the given tables will be
 309             returned.
 310
 311         `exact_only`
 312             If True, only exact matches are returned.  If set to False (the
 313             default), and the provided `name` doesn't match anything exactly,
 314             spelling correction will be attempted.
 315         """
 316
 317         name = self.normalize_name(input)
 318         exact = True
 319         form = None
 320
 321         # Remove any type prefix (pokemon:133) before constructing a query
 322         if ':' in name:
 323             prefix_chunk, name = name.split(':', 1)
 324             name = name.strip()
 325
 326             if not valid_types:
 327                 # Only use types from the query string if none were explicitly
 328                 # provided
 329                 prefixes = prefix_chunk.split(',')
 330                 valid_types = [_.strip() for _ in prefixes]
 331
 332         # Random lookup
 333         if name == 'random':
 334             return self.random_lookup(valid_types=valid_types)
 335
 336         # Do different things depending what the query looks like
 337         # Note: Term objects do an exact match, so we don't have to worry about
 338         # a query parser tripping on weird characters in the input
 339         if '*' in name or '?' in name:
 340             exact_only = True
 341             query = whoosh.query.Wildcard(u'name', name)
 342         elif rx_is_number.match(name):
 343             # Don't spell-check numbers!
 344             exact_only = True
 345             query = whoosh.query.Term(u'row_id', name)
 346         else:
 347             # Not an integer
 348             query = whoosh.query.Term(u'name', name)
 349
 350         ### Filter by type of object
 351         type_terms = []
 352         for valid_type in valid_types:
 353             table_name = self._parse_table_name(valid_type)
 354             if table_name:
 355                 # Quietly ignore bogus valid_types; more likely to DTRT
 356                 type_terms.append(whoosh.query.Term(u'table', table_name))
 357
 358         if type_terms:
 359             query = query & whoosh.query.Or(type_terms)
 360
 361
 362         ### Actual searching
 363         searcher = self.index.searcher()
 364         # XXX is this kosher?  docs say search() takes a weighting arg, but it
 365         # certainly does not
 366         searcher.weighting = LanguageWeighting()
 367         results = searcher.search(query,
 368                                   limit=self.INTERMEDIATE_LOOKUP_RESULTS)
 369
 370         # Look for some fuzzy matches if necessary
 371         if not exact_only and not results:
 372             exact = False
 373             results = []
 374
 375             for suggestion in self.speller.suggest(
 376                 name, self.INTERMEDIATE_LOOKUP_RESULTS):
 377
 378                 query = whoosh.query.Term('name', suggestion)
 379                 results.extend(searcher.search(query))
 380
 381         ### Convert results to db objects
 382         objects = self._whoosh_records_to_results(results, exact=exact)
 383
 384         # Only return up to 10 matches; beyond that, something is wrong.  We
 385         # strip out duplicate entries above, so it's remotely possible that we
 386         # should have more than 10 here and lost a few.  The speller returns 25
 387         # to give us some padding, and should avoid that problem.  Not a big
 388         # deal if we lose the 25th-most-likely match anyway.
 389         return objects[:self.MAX_LOOKUP_RESULTS]
 390
 391
 392     def random_lookup(self, valid_types=[]):
 393         """Returns a random lookup result from one of the provided
 394         `valid_types`.
 395         """
 396
 397         tables = []
 398         for valid_type in valid_types:
 399             table_name = self._parse_table_name(valid_type)
 400             if table_name:
 401                 tables.append(self.indexed_tables[table_name])
 402
 403         if not tables:
 404             # n.b.: It's possible we got a list of valid_types and none of them
 405             # were valid, but this function is guaranteed to return
 406             # *something*, so it politely selects from the entire index isntead
 407             tables = self.indexed_tables.values()
 408
 409         # Rather than create an array of many hundred items and pick randomly
 410         # from it, just pick a number up to the total number of potential
 411         # items, then pick randomly from that, and partition the whole range
 412         # into chunks.  This also avoids the slight problem that the index
 413         # contains more rows (for languages) for some items than others.
 414         # XXX ought to cache this (in the index?) if possible
 415         total = 0
 416         partitions = []
 417         for table in tables:
 418             count = self.session.query(table).count()
 419             total += count
 420             partitions.append((table, count))
 421
 422         n = random.randint(1, total)
 423         while n > partitions[0][1]:
 424             n -= partitions[0][1]
 425             partitions.pop(0)
 426
 427         return self.lookup(unicode(n), valid_types=[ partitions[0][0] ])
 428
 429     def prefix_lookup(self, prefix):
 430         """Returns terms starting with the given exact prefix.
 431
 432         No special magic is currently done with the name; type prefixes are not
 433         recognized.
 434         """
 435
 436         query = whoosh.query.Prefix(u'name', self.normalize_name(prefix))
 437
 438         searcher = self.index.searcher()
 439         searcher.weighting = LanguageWeighting()
 440         results = searcher.search(query)  # XXX , limit=self.MAX_LOOKUP_RESULTS)
 441
 442         return self._whoosh_records_to_results(results)