Random lookup algorithm is now more naive, but less broken.
[zzz-pokedex.git] / pokedex / lookup.py
1 # encoding: utf8
2 import os, os.path
3 import random
4 import re
5 import shutil
6 import unicodedata
7
8 from sqlalchemy.sql import func
9 import whoosh
10 import whoosh.filedb.filestore
11 import whoosh.filedb.fileindex
12 import whoosh.index
13 from whoosh.qparser import QueryParser
14 import whoosh.scoring
15 import whoosh.spelling
16
17 from pokedex.util import namedtuple
18
19 from pokedex.db import connect
20 import pokedex.db.tables as tables
21 from pokedex.roomaji import romanize
22 from pokedex.defaults import get_default_index_dir
23
24 __all__ = ['PokedexLookup']
25
26
27 rx_is_number = re.compile('^\d+$')
28
29 LookupResult = namedtuple('LookupResult', [
30 'object', 'indexed_name', 'name', 'language', 'iso639', 'iso3166', 'exact',
31 ])
32
33 class UninitializedIndex(object):
34 class UninitializedIndexError(Exception):
35 pass
36
37 def __nonzero__(self):
38 """Dummy object should identify itself as False."""
39 return False
40
41 def __bool__(self):
42 """Python 3000 version of the above. Future-proofing rules!"""
43 return False
44
45 def __getattr__(self, *args, **kwargs):
46 raise self.UninitializedIndexError(
47 "The lookup index does not exist. Please use `pokedex setup` "
48 "or lookup.rebuild_index() to create it."
49 )
50
51 class LanguageWeighting(whoosh.scoring.Weighting):
52 """A scoring class that forces otherwise-equal English results to come
53 before foreign results.
54 """
55
56 def __init__(self, extra_weights={}, *args, **kwargs):
57 """`extra_weights` may be a dictionary of weights which will be
58 factored in.
59
60 Intended for use with spelling corrections, which come along with their
61 own weightings.
62 """
63 self.extra_weights = extra_weights
64 super(LanguageWeighting, self).__init__(*args, **kwargs)
65
66 def score(self, searcher, fieldnum, text, docnum, weight, QTF=1):
67 doc = searcher.stored_fields(docnum)
68
69 # Apply extra weight
70 weight = weight * self.extra_weights.get(text, 1.0)
71
72 language = doc.get('language')
73 if language is None:
74 # English (well, "default"); leave it at 1
75 return weight
76 elif language == u'Roomaji':
77 # Give Roomaji a little boost; it's most likely to be searched
78 return weight * 0.9
79 else:
80 # Everything else can drop down the totem pole
81 return weight * 0.8
82
83
84 class PokedexLookup(object):
85 MAX_FUZZY_RESULTS = 10
86 MAX_EXACT_RESULTS = 43
87 INTERMEDIATE_FACTOR = 2
88
89 # The speller only checks how much the input matches a word; there can be
90 # all manner of extra unmatched junk, and it won't affect the weighting.
91 # To compensate, greatly boost the weighting of matches at the beginning
92 # and end, so nearly-full-word-matches are much better
93 SPELLER_OPTIONS = dict(booststart=10.0, boostend=9.0)
94
95 # Dictionary of table name => table class.
96 # Need the table name so we can get the class from the table name after we
97 # retrieve something from the index
98 indexed_tables = dict(
99 (cls.__tablename__, cls)
100 for cls in (
101 tables.Ability,
102 tables.Item,
103 tables.Location,
104 tables.Move,
105 tables.Nature,
106 tables.Pokemon,
107 tables.PokemonForm,
108 tables.Type,
109 )
110 )
111
112
113 def __init__(self, directory=None, session=None):
114 """Opens the whoosh index stored in the named directory. If the index
115 doesn't already exist, it will be created.
116
117 `directory`
118 Directory containing the index. Defaults to a location within the
119 `pokedex` egg directory.
120
121 `session`
122 Used for creating the index and retrieving objects. Defaults to an
123 attempt to connect to the default SQLite database installed by
124 `pokedex setup`.
125 """
126
127 # By the time this returns, self.index, self.speller, and self.session
128 # must be set
129
130 # If a directory was not given, use the default
131 if directory is None:
132 directory = get_default_index_dir()
133
134 self.directory = directory
135
136 if session:
137 self.session = session
138 else:
139 self.session = connect()
140
141 # Attempt to open or create the index
142 if not os.path.exists(directory) or not os.listdir(directory):
143 # Directory doesn't exist OR is empty; caller needs to use
144 # rebuild_index before doing anything. Provide a dummy object that
145 # complains when used
146 self.index = UninitializedIndex()
147 self.speller = UninitializedIndex()
148 return
149
150 # Otherwise, already exists; should be an index! Bam, done.
151 # Note that this will explode if the directory exists but doesn't
152 # contain an index; that's a feature
153 try:
154 self.index = whoosh.index.open_dir(directory, indexname='MAIN')
155 except whoosh.index.EmptyIndexError:
156 raise IOError(
157 "The index directory already contains files. "
158 "Please use a dedicated directory for the lookup index."
159 )
160
161 # Create speller, and done
162 spell_store = whoosh.filedb.filestore.FileStorage(directory)
163 self.speller = whoosh.spelling.SpellChecker(spell_store,
164 **self.SPELLER_OPTIONS)
165
166
167 def rebuild_index(self):
168 """Creates the index from scratch."""
169
170 schema = whoosh.fields.Schema(
171 name=whoosh.fields.ID(stored=True),
172 table=whoosh.fields.ID(stored=True),
173 row_id=whoosh.fields.ID(stored=True),
174 language=whoosh.fields.STORED,
175 iso639=whoosh.fields.ID(stored=True),
176 iso3166=whoosh.fields.ID(stored=True),
177 display_name=whoosh.fields.STORED, # non-lowercased name
178 )
179
180 if os.path.exists(self.directory):
181 # create_in() isn't totally reliable, so just nuke whatever's there
182 # manually. Try to be careful about this...
183 for f in os.listdir(self.directory):
184 if re.match('^_?(MAIN|SPELL)_', f):
185 os.remove(os.path.join(self.directory, f))
186 else:
187 os.mkdir(self.directory)
188
189 self.index = whoosh.index.create_in(self.directory, schema=schema,
190 indexname='MAIN')
191 writer = self.index.writer()
192
193 # Index every name in all our tables of interest
194 speller_entries = set()
195 for cls in self.indexed_tables.values():
196 q = self.session.query(cls)
197
198 for row in q.yield_per(5):
199 row_key = dict(table=unicode(cls.__tablename__),
200 row_id=unicode(row.id))
201
202 def add(name, language, iso639, iso3166):
203 normalized_name = self.normalize_name(name)
204
205 writer.add_document(
206 name=normalized_name, display_name=name,
207 language=language, iso639=iso639, iso3166=iso3166,
208 **row_key
209 )
210
211 speller_entries.add(normalized_name)
212
213
214 # Add the basic English name to the index
215 if cls == tables.Pokemon:
216 # Don't re-add alternate forms of the same Pokémon; they'll
217 # be added as Pokémon forms instead
218 if not row.is_base_form:
219 continue
220 elif cls == tables.PokemonForm:
221 if row.name:
222 add(row.pokemon_name, None, u'en', u'us')
223 continue
224
225 # Some things also have other languages' names
226 # XXX other language form names..?
227 seen = set()
228 for language, name in getattr(row, 'name_map', {}).items():
229 if name in seen:
230 # Don't add the name again as a different
231 # language; no point and it makes spell results
232 # confusing
233 continue
234 seen.add(name)
235
236 add(name, language.name,
237 language.iso639,
238 language.iso3166)
239
240 # Add Roomaji too
241 if language.identifier == 'ja':
242 roomaji = romanize(name)
243 add(roomaji, u'Roomaji', u'ja', u'jp')
244
245 writer.commit()
246
247 # Construct and populate a spell-checker index. Quicker to do it all
248 # at once, as every call to add_* does a commit(), and those seem to be
249 # expensive
250 self.speller = whoosh.spelling.SpellChecker(self.index.storage, mingram=2,
251 **self.SPELLER_OPTIONS)
252 self.speller.add_words(speller_entries)
253
254
255 def normalize_name(self, name):
256 """Strips irrelevant formatting junk from name input.
257
258 Specifically: everything is lowercased, and accents are removed.
259 """
260 # http://stackoverflow.com/questions/517923/what-is-the-best-way-to-remove-accents-in-a-python-unicode-string
261 # Makes sense to me. Decompose by Unicode rules, then remove combining
262 # characters, then recombine. I'm explicitly doing it this way instead
263 # of testing combining() because Korean characters apparently
264 # decompose! But the results are considered letters, not combining
265 # characters, so testing for Mn works well, and combining them again
266 # makes them look right.
267 nkfd_form = unicodedata.normalize('NFKD', unicode(name))
268 name = u"".join(c for c in nkfd_form
269 if unicodedata.category(c) != 'Mn')
270 name = unicodedata.normalize('NFC', name)
271
272 name = name.strip()
273 name = name.lower()
274
275 return name
276
277
278 def _apply_valid_types(self, name, valid_types):
279 """Combines the enforced `valid_types` with any from the search string
280 itself and updates the query.
281
282 For example, a name of 'a,b:foo' and valid_types of b,c will search for
283 only `b`s named "foo".
284
285 Returns `(name, merged_valid_types, term)`, where `name` has had any type
286 prefix stripped, `merged_valid_types` combines the original
287 `valid_types` with the type prefix, and `term` is a query term for
288 limited to just the allowed types. If there are no type restrictions
289 at all, `term` will be None.
290 """
291
292 # Remove any type prefix (pokemon:133) first
293 user_valid_types = []
294 if ':' in name:
295 prefix_chunk, name = name.split(':', 1)
296 name = name.strip()
297
298 prefixes = prefix_chunk.split(',')
299 user_valid_types = []
300 for prefix in prefixes:
301 prefix = prefix.strip()
302 if prefix:
303 user_valid_types.append(prefix)
304
305 # Merge the valid types together. Only types that appear in BOTH lists
306 # may be used.
307 # As a special case, if the user asked for types that are explicitly
308 # forbidden, completely ignore what the user requested.
309 # And, just to complicate matters: "type" and language need to be
310 # considered separately.
311 def merge_requirements(func):
312 user = filter(func, user_valid_types)
313 system = filter(func, valid_types)
314
315 if user and system:
316 merged = list(set(user) & set(system))
317 if merged:
318 return merged
319 else:
320 # No overlap; use the system restrictions
321 return system
322 else:
323 # One or the other is blank; use the one that's not
324 return user or system
325
326 # @foo means language must be foo; otherwise it's a table name
327 lang_requirements = merge_requirements(lambda req: req[0] == u'@')
328 type_requirements = merge_requirements(lambda req: req[0] != u'@')
329 all_requirements = lang_requirements + type_requirements
330
331 # Construct the term
332 lang_terms = []
333 for lang in lang_requirements:
334 # Allow for either country or language codes
335 lang_code = lang[1:]
336 lang_terms.append(whoosh.query.Term(u'iso639', lang_code))
337 lang_terms.append(whoosh.query.Term(u'iso3166', lang_code))
338
339 type_terms = []
340 for type in type_requirements:
341 table_name = self._parse_table_name(type)
342
343 # Quietly ignore bogus valid_types; more likely to DTRT
344 if table_name:
345 type_terms.append(whoosh.query.Term(u'table', table_name))
346
347 # Combine both kinds of restriction
348 all_terms = []
349 if type_terms:
350 all_terms.append(whoosh.query.Or(type_terms))
351 if lang_terms:
352 all_terms.append(whoosh.query.Or(lang_terms))
353
354 return name, all_requirements, whoosh.query.And(all_terms)
355
356
357 def _parse_table_name(self, name):
358 """Takes a singular table name, table name, or table object and returns
359 the table name.
360
361 Returns None for a bogus name.
362 """
363 # Table object
364 if hasattr(name, '__tablename__'):
365 return getattr(name, '__tablename__')
366
367 # Table name
368 for table in self.indexed_tables.values():
369 if name in (table.__tablename__, table.__singlename__):
370 return table.__tablename__
371
372 # Bogus. Be nice and return dummy
373 return None
374
375 def _whoosh_records_to_results(self, records, exact=True):
376 """Converts a list of whoosh's indexed records to LookupResult tuples
377 containing database objects.
378 """
379 # XXX this 'exact' thing is getting kinda leaky. would like a better
380 # way to handle it, since only lookup() cares about fuzzy results
381 seen = {}
382 results = []
383 for record in records:
384 # Skip dupes
385 seen_key = record['table'], record['row_id']
386 if seen_key in seen:
387 continue
388 seen[seen_key] = True
389
390 cls = self.indexed_tables[record['table']]
391 obj = self.session.query(cls).get(record['row_id'])
392
393 results.append(LookupResult(object=obj,
394 indexed_name=record['name'],
395 name=record['display_name'],
396 language=record.get('language'),
397 iso639=record['iso639'],
398 iso3166=record['iso3166'],
399 exact=exact))
400
401 return results
402
403
404 def lookup(self, input, valid_types=[], exact_only=False):
405 """Attempts to find some sort of object, given a name.
406
407 Returns a list of named (object, name, language, iso639, iso3166,
408 exact) tuples. `object` is a database object, `name` is the name under
409 which the object was found, `language` and the two isos are the name
410 and country codes of the language in which the name was found, and
411 `exact` is True iff this was an exact match.
412
413 This function currently ONLY does fuzzy matching if there are no exact
414 matches.
415
416 Formes are not returned unless requested; "Shaymin" will return only
417 grass Shaymin.
418
419 Extraneous whitespace is removed with extreme prejudice.
420
421 Recognizes:
422 - Names: "Eevee", "Surf", "Run Away", "Payapa Berry", etc.
423 - Foreign names: "Iibui", "Eivui"
424 - Fuzzy names in whatever language: "Evee", "Ibui"
425 - IDs: "133", "192", "250"
426 Also:
427 - Type restrictions. "type:psychic" will only return the type. This
428 is how to make ID lookup useful. Multiple type specs can be entered
429 with commas, as "move,item:1".
430 - Language restrictions. "@fr:charge" will only return Tackle, which
431 is called "Charge" in French. These can be combined with type
432 restrictions, e.g., "@fr,move:charge".
433 - Alternate formes can be specified merely like "wash rotom".
434
435 `input`
436 Name of the thing to look for.
437
438 `valid_types`
439 A list of type or language restrictions, e.g., `['pokemon',
440 '@ja']`. If this is provided, only results in one of the given
441 tables will be returned.
442
443 `exact_only`
444 If True, only exact matches are returned. If set to False (the
445 default), and the provided `name` doesn't match anything exactly,
446 spelling correction will be attempted.
447 """
448
449 name = self.normalize_name(input)
450 exact = True
451 form = None
452
453 # Pop off any type prefix and merge with valid_types
454 name, merged_valid_types, type_term = \
455 self._apply_valid_types(name, valid_types)
456
457 # Random lookup
458 if name == 'random':
459 return self.random_lookup(valid_types=merged_valid_types)
460
461 # Do different things depending what the query looks like
462 # Note: Term objects do an exact match, so we don't have to worry about
463 # a query parser tripping on weird characters in the input
464 try:
465 # Let Python try to convert to a number, so 0xff works
466 name_as_number = int(name, base=0)
467 except ValueError:
468 # Oh well
469 name_as_number = None
470
471 if '*' in name or '?' in name:
472 exact_only = True
473 query = whoosh.query.Wildcard(u'name', name)
474 elif name_as_number is not None:
475 # Don't spell-check numbers!
476 exact_only = True
477 query = whoosh.query.Term(u'row_id', unicode(name_as_number))
478 else:
479 # Not an integer
480 query = whoosh.query.Term(u'name', name)
481
482 if type_term:
483 query = query & type_term
484
485
486 ### Actual searching
487 # Limits; result limits are constants, and intermediate results (before
488 # duplicate items are stripped out) are capped at the result limit
489 # times another constant.
490 # Fuzzy are capped at 10, beyond which something is probably very
491 # wrong. Exact matches -- that is, wildcards and ids -- are far less
492 # constrained.
493 # Also, exact matches are sorted by name, since weight doesn't matter.
494 sort_by = dict()
495 if exact_only:
496 max_results = self.MAX_EXACT_RESULTS
497 sort_by['sortedby'] = (u'table', u'name')
498 else:
499 max_results = self.MAX_FUZZY_RESULTS
500
501 searcher = self.index.searcher(weighting=LanguageWeighting())
502 results = searcher.search(
503 query,
504 limit=int(max_results * self.INTERMEDIATE_FACTOR),
505 **sort_by
506 )
507
508 # Look for some fuzzy matches if necessary
509 if not exact_only and not results:
510 exact = False
511 results = []
512
513 fuzzy_query_parts = []
514 fuzzy_weights = {}
515 min_weight = [None]
516 for suggestion, _, weight in self.speller.suggestions_and_scores(name):
517 # Only allow the top 50% of scores; otherwise there will always
518 # be a lot of trailing junk
519 if min_weight[0] is None:
520 min_weight[0] = weight * 0.5
521 elif weight < min_weight[0]:
522 break
523
524 fuzzy_query_parts.append(whoosh.query.Term('name', suggestion))
525 fuzzy_weights[suggestion] = weight
526
527 if not fuzzy_query_parts:
528 # Nothing at all; don't try querying
529 return []
530
531 fuzzy_query = whoosh.query.Or(fuzzy_query_parts)
532 if type_term:
533 fuzzy_query = fuzzy_query & type_term
534
535 searcher.weighting = LanguageWeighting(extra_weights=fuzzy_weights)
536 results = searcher.search(fuzzy_query)
537
538 ### Convert results to db objects
539 objects = self._whoosh_records_to_results(results, exact=exact)
540
541 # Truncate and return
542 return objects[:max_results]
543
544
545 def random_lookup(self, valid_types=[]):
546 """Returns a random lookup result from one of the provided
547 `valid_types`.
548 """
549
550 table_names = []
551 for valid_type in valid_types:
552 table_name = self._parse_table_name(valid_type)
553 # Skip anything not recognized. Could be, say, a language code.
554 # XXX The vast majority of Pokémon forms are unnamed and unindexed,
555 # which can produce blank results. So skip them too for now.
556 if table_name and table_name != 'pokemon_forms':
557 table_names.append(table_name)
558
559 if not table_names:
560 # n.b.: It's possible we got a list of valid_types and none of them
561 # were valid, but this function is guaranteed to return
562 # *something*, so it politely selects from the entire index instead
563 table_names = self.indexed_tables.keys()
564 table_names.remove('pokemon_forms')
565
566 # Pick a random table, then pick a random item from it. Small tables
567 # like Type will have an unnatural bias. The alternative is that a
568 # simple search for "random" will do some eight queries, counting the
569 # rows in every single indexed table, and that's awful.
570 # XXX Can we improve on this, reasonably?
571 table_name = random.choice(table_names)
572 count = self.session.query(self.indexed_tables[table_name]).count()
573 id, = self.session.query(self.indexed_tables[table_name].id) \
574 .offset(random.randint(0, count - 1)) \
575 .first()
576
577 return self.lookup(unicode(id), valid_types=[table_name])
578
579 def prefix_lookup(self, prefix, valid_types=[]):
580 """Returns terms starting with the given exact prefix.
581
582 Type prefixes are recognized, but no other name munging is done.
583 """
584
585 # Pop off any type prefix and merge with valid_types
586 prefix, merged_valid_types, type_term = \
587 self._apply_valid_types(prefix, valid_types)
588
589 query = whoosh.query.Prefix(u'name', self.normalize_name(prefix))
590
591 if type_term:
592 query = query & type_term
593
594 searcher = self.index.searcher()
595 searcher.weighting = LanguageWeighting()
596 results = searcher.search(query) # XXX , limit=self.MAX_LOOKUP_RESULTS)
597
598 return self._whoosh_records_to_results(results)