X-Git-Url: http://git.veekun.com/zzz-pokedex.git/blobdiff_plain/b5fde0da1952ae91f8addce211b3e967841fffd7..5ed0ed3118bcd848f4d9533fc00baf3afbef4322:/pokedex/lookup.py diff --git a/pokedex/lookup.py b/pokedex/lookup.py index 16c4e5e..0b09823 100644 --- a/pokedex/lookup.py +++ b/pokedex/lookup.py @@ -1,8 +1,9 @@ # encoding: utf8 -from collections import namedtuple import os, os.path -import pkg_resources +import random import re +import shutil +import unicodedata from sqlalchemy.sql import func import whoosh @@ -13,136 +14,38 @@ from whoosh.qparser import QueryParser import whoosh.scoring import whoosh.spelling +from pokedex.util import namedtuple + from pokedex.db import connect import pokedex.db.tables as tables from pokedex.roomaji import romanize +from pokedex.defaults import get_default_index_dir -__all__ = ['open_index', 'lookup'] - -# Dictionary of table name => table class. -# Need the table name so we can get the class from the table name after we -# retrieve something from the index -indexed_tables = {} -for cls in [ - tables.Ability, - tables.Item, - tables.Move, - tables.Pokemon, - tables.Type, - ]: - indexed_tables[cls.__tablename__] = cls - -def open_index(directory=None, session=None, recreate=False): - """Opens the whoosh index stored in the named directory and returns (index, - speller). If the index doesn't already exist, it will be created. - - `directory` - Directory containing the index. Defaults to a location within the - `pokedex` egg directory. - - `session` - If the index needs to be created, this database session will be used. - Defaults to an attempt to connect to the default SQLite database - installed by `pokedex setup`. - - `recreate` - If set to True, the whoosh index will be created even if it already - exists. - """ +__all__ = ['PokedexLookup'] - # Defaults - if not directory: - directory = pkg_resources.resource_filename('pokedex', - 'data/whoosh_index') - if not session: - session = connect() +rx_is_number = re.compile('^\d+$') - # Attempt to open or create the index - directory_exists = os.path.exists(directory) - if directory_exists and not recreate: - # Already exists; should be an index! - try: - index = whoosh.index.open_dir(directory, indexname='MAIN') - spell_store = whoosh.filedb.filestore.FileStorage(directory) - speller = whoosh.spelling.SpellChecker(spell_store) - return index, speller - except whoosh.index.EmptyIndexError as e: - # Apparently not a real index. Fall out of the if and create it - pass - - if not directory_exists: - os.mkdir(directory) - - - # Create index - schema = whoosh.fields.Schema( - name=whoosh.fields.ID(stored=True), - table=whoosh.fields.STORED, - row_id=whoosh.fields.ID(stored=True), - language=whoosh.fields.STORED, - display_name=whoosh.fields.STORED, # non-lowercased name - ) +LookupResult = namedtuple('LookupResult', + ['object', 'indexed_name', 'name', 'language', 'iso3166', 'exact']) + +class UninitializedIndex(object): + class UninitializedIndexError(Exception): + pass - index = whoosh.index.create_in(directory, schema=schema, indexname='MAIN') - writer = index.writer() - - # Index every name in all our tables of interest - # speller_entries becomes a list of (word, score) tuples; the score is 2 - # for English names, 1.5 for Roomaji, and 1 for everything else. I think - # this biases the results in the direction most people expect, especially - # when e.g. German names are very similar to English names - speller_entries = [] - for cls in indexed_tables.values(): - q = session.query(cls) - - # Only index base Pokémon formes - if hasattr(cls, 'forme_base_pokemon_id'): - q = q.filter_by(forme_base_pokemon_id=None) - - for row in q.yield_per(5): - row_key = dict(table=cls.__tablename__, row_id=unicode(row.id)) - - name = row.name - writer.add_document(name=name.lower(), - display_name=name, - **row_key) - speller_entries.append((name.lower(), 1)) - - # Pokemon also get other languages - for foreign_name in getattr(row, 'foreign_names', []): - moonspeak = foreign_name.name - if name == moonspeak: - # Don't add the English name again as a different language; - # no point and it makes spell results confusing - continue - - writer.add_document(name=moonspeak.lower(), - language=foreign_name.language.name, - display_name=moonspeak, - **row_key) - speller_entries.append((moonspeak.lower(), 3)) - - # Add Roomaji too - if foreign_name.language.name == 'Japanese': - roomaji = romanize(foreign_name.name) - writer.add_document(name=roomaji.lower(), - language='Roomaji', - display_name=roomaji, - **row_key) - speller_entries.append((roomaji.lower(), 8)) - - - writer.commit() - - # Construct and populate a spell-checker index. Quicker to do it all - # at once, as every call to add_* does a commit(), and those seem to be - # expensive - speller = whoosh.spelling.SpellChecker(index.storage) - speller.add_scored_words(speller_entries) - - return index, speller + def __nonzero__(self): + """Dummy object should identify itself as False.""" + return False + def __bool__(self): + """Python 3000 version of the above. Future-proofing rules!""" + return False + + def __getattr__(self, *args, **kwargs): + raise self.UninitializedIndexError( + "The lookup index does not exist. Please use `pokedex setup` " + "or lookup.rebuild_index() to create it." + ) class LanguageWeighting(whoosh.scoring.Weighting): """A scoring class that forces otherwise-equal English results to come @@ -155,112 +58,466 @@ class LanguageWeighting(whoosh.scoring.Weighting): # English (well, "default"); leave it at 1 return weight elif doc['language'] == u'Roomaji': - # Give Roomaji a bit of a boost, as it's most likely to be searched + # Give Roomaji a little boost; it's most likely to be searched return weight * 0.95 else: # Everything else can drop down the totem pole return weight * 0.9 -rx_is_number = re.compile('^\d+$') -LookupResult = namedtuple('LookupResult', - ['object', 'name', 'language', 'exact']) -def lookup(input, session=None, indices=None, exact_only=False): - """Attempts to find some sort of object, given a database session and name. - - Returns a list of named (object, name, language, exact) tuples. `object` - is a database object, `name` is the name under which the object was found, - `language` is the name of the language in which the name was found, and - `exact` is True iff this was an exact match. - - This function currently ONLY does fuzzy matching if there are no exact - matches. - - Formes are not returned; "Shaymin" will return only grass Shaymin. - - Recognizes: - - Names: "Eevee", "Surf", "Run Away", "Payapa Berry", etc. - - Foreign names: "Iibui", "Eivui" - - Fuzzy names in whatever language: "Evee", "Ibui" - - IDs: "pokemon 133", "move 192", "item 250" - - Dex numbers: "sinnoh 55", "133", "johto 180" - - `input` - Name of the thing to look for. - - `session` - A database session to use for retrieving objects. As with get_index, - if this is not provided, a connection to the default database will be - attempted. - - `indices` - Tuple of index, speller as returned from `open_index()`. Defaults to - a call to `open_index()`. - - `exact_only` - If True, only exact matches are returned. If set to False (the - default), and the provided `name` doesn't match anything exactly, - spelling correction will be attempted. - """ +class PokedexLookup(object): + INTERMEDIATE_LOOKUP_RESULTS = 25 + MAX_LOOKUP_RESULTS = 10 + + # Dictionary of table name => table class. + # Need the table name so we can get the class from the table name after we + # retrieve something from the index + indexed_tables = dict( + (cls.__tablename__, cls) + for cls in ( + tables.Ability, + tables.Item, + tables.Location, + tables.Move, + tables.Nature, + tables.Pokemon, + tables.Type, + ) + ) - if not session: - session = connect() - - if indices: - index, speller = indices - else: - index, speller = open_index() - - name = unicode(input).lower() - exact = True - - # If the input provided is a number, match it as an id. Otherwise, name. - # Term objects do an exact match, so we don't have to worry about a query - # parser tripping on weird characters in the input - if rx_is_number.match(name): - # Don't spell-check numbers! - exact_only = True - query = whoosh.query.Term(u'row_id', name) - else: - # Not an integer - query = whoosh.query.Term(u'name', name) - - ### Actual searching - searcher = index.searcher() - searcher.weighting = LanguageWeighting() # XXX kosher? docs say search() - # takes a weighting kw but it - # certainly does not - results = searcher.search(query) - - # Look for some fuzzy matches if necessary - if not exact_only and not results: - exact = False - results = [] - for suggestion in speller.suggest(name, 25): - query = whoosh.query.Term('name', suggestion) - results.extend(searcher.search(query)) - - ### Convert results to db objects - objects = [] - seen = {} - for result in results: - # Skip dupe results - seen_key = result['table'], result['row_id'] - if seen_key in seen: - continue - seen[seen_key] = True - - cls = indexed_tables[result['table']] - obj = session.query(cls).get(result['row_id']) - objects.append(LookupResult(object=obj, - name=result['display_name'], - language=result['language'], - exact=exact)) - - # Only return up to 10 matches; beyond that, something is wrong. - # We strip out duplicate entries above, so it's remotely possible that we - # should have more than 10 here and lost a few. The speller returns 25 to - # give us some padding, and should avoid that problem. Not a big deal if - # we lose the 25th-most-likely match anyway. - return objects[:10] + def __init__(self, directory=None, session=None): + """Opens the whoosh index stored in the named directory. If the index + doesn't already exist, it will be created. + + `directory` + Directory containing the index. Defaults to a location within the + `pokedex` egg directory. + + `session` + Used for creating the index and retrieving objects. Defaults to an + attempt to connect to the default SQLite database installed by + `pokedex setup`. + """ + + # By the time this returns, self.index, self.speller, and self.session + # must be set + + # If a directory was not given, use the default + if directory is None: + directory = get_default_index_dir() + + self.directory = directory + + if session: + self.session = session + else: + self.session = connect() + + # Attempt to open or create the index + if not os.path.exists(directory) or not os.listdir(directory): + # Directory doesn't exist OR is empty; caller needs to use + # rebuild_index before doing anything. Provide a dummy object that + # complains when used + self.index = UninitializedIndex() + self.speller = UninitializedIndex() + return + + # Otherwise, already exists; should be an index! Bam, done. + # Note that this will explode if the directory exists but doesn't + # contain an index; that's a feature + try: + self.index = whoosh.index.open_dir(directory, indexname='MAIN') + except whoosh.index.EmptyIndexError: + raise IOError( + "The index directory already contains files. " + "Please use a dedicated directory for the lookup index." + ) + + # Create speller, and done + spell_store = whoosh.filedb.filestore.FileStorage(directory) + self.speller = whoosh.spelling.SpellChecker(spell_store) + + + def rebuild_index(self): + """Creates the index from scratch.""" + + schema = whoosh.fields.Schema( + name=whoosh.fields.ID(stored=True), + table=whoosh.fields.ID(stored=True), + row_id=whoosh.fields.ID(stored=True), + language=whoosh.fields.STORED, + iso3166=whoosh.fields.STORED, + display_name=whoosh.fields.STORED, # non-lowercased name + ) + + if not os.path.exists(self.directory): + os.mkdir(self.directory) + + self.index = whoosh.index.create_in(self.directory, schema=schema, + indexname='MAIN') + writer = self.index.writer() + + # Index every name in all our tables of interest + # speller_entries becomes a list of (word, score) tuples; the score is + # 2 for English names, 1.5 for Roomaji, and 1 for everything else. I + # think this biases the results in the direction most people expect, + # especially when e.g. German names are very similar to English names + speller_entries = [] + for cls in self.indexed_tables.values(): + q = self.session.query(cls) + + for row in q.yield_per(5): + row_key = dict(table=unicode(cls.__tablename__), + row_id=unicode(row.id)) + + def add(name, language, iso3166, score): + normalized_name = self.normalize_name(name) + + writer.add_document( + name=normalized_name, display_name=name, + language=language, iso3166=iso3166, + **row_key + ) + + speller_entries.append((normalized_name, score)) + + + # Add the basic English name to the index + if cls == tables.Pokemon: + # Pokémon need their form name added + # XXX kinda kludgy + add(row.full_name, None, u'us', 1) + + # If this is a default form, ALSO add the unadorned name, + # so 'Deoxys' alone will still do the right thing + if row.forme_name and not row.forme_base_pokemon_id: + add(row.name, None, u'us', 1) + else: + add(row.name, None, u'us', 1) + + # Some things also have other languages' names + # XXX other language form names..? + for foreign_name in getattr(row, 'foreign_names', []): + moonspeak = foreign_name.name + if row.name == moonspeak: + # Don't add the English name again as a different + # language; no point and it makes spell results + # confusing + continue + + add(moonspeak, foreign_name.language.name, + foreign_name.language.iso3166, + 3) + + # Add Roomaji too + if foreign_name.language.name == 'Japanese': + roomaji = romanize(foreign_name.name) + add(roomaji, u'Roomaji', u'jp', 8) + + writer.commit() + + # Construct and populate a spell-checker index. Quicker to do it all + # at once, as every call to add_* does a commit(), and those seem to be + # expensive + self.speller = whoosh.spelling.SpellChecker(self.index.storage) + self.speller.add_scored_words(speller_entries) + + + def normalize_name(self, name): + """Strips irrelevant formatting junk from name input. + + Specifically: everything is lowercased, and accents are removed. + """ + # http://stackoverflow.com/questions/517923/what-is-the-best-way-to-remove-accents-in-a-python-unicode-string + # Makes sense to me. Decompose by Unicode rules, then remove combining + # characters, then recombine. I'm explicitly doing it this way instead + # of testing combining() because Korean characters apparently + # decompose! But the results are considered letters, not combining + # characters, so testing for Mn works well, and combining them again + # makes them look right. + nkfd_form = unicodedata.normalize('NFKD', unicode(name)) + name = u"".join(c for c in nkfd_form + if unicodedata.category(c) != 'Mn') + name = unicodedata.normalize('NFC', name) + + name = name.strip() + name = name.lower() + + return name + + + def _apply_valid_types(self, name, valid_types): + """Combines the enforced `valid_types` with any from the search string + itself and updates the query. + + For example, a name of 'a,b:foo' and valid_types of b,c will search for + only `b`s named "foo". + + Returns `(name, merged_valid_types, term)`, where `name` has had any type + prefix stripped, `merged_valid_types` combines the original + `valid_types` with the type prefix, and `term` is a query term for + limited to just the allowed types. If there are no type restrictions + at all, `term` will be None. + """ + + # Remove any type prefix (pokemon:133) first + user_valid_types = [] + if ':' in name: + prefix_chunk, name = name.split(':', 1) + name = name.strip() + + prefixes = prefix_chunk.split(',') + user_valid_types = [_.strip() for _ in prefixes] + + # Merge the valid types together. Only types that appear in BOTH lists + # may be used. + # As a special case, if the user asked for types that are explicitly + # forbidden, completely ignore what the user requested + combined_valid_types = [] + if user_valid_types and valid_types: + combined_valid_types = list( + set(user_valid_types) & set(combined_valid_types) + ) + + if not combined_valid_types: + # No overlap! Just use the enforced ones + combined_valid_types = valid_types + else: + # One list or the other was blank, so just use the one that isn't + combined_valid_types = valid_types + user_valid_types + + if not combined_valid_types: + # No restrictions + return name, [], None + + # Construct the term + type_terms = [] + final_valid_types = [] + for valid_type in combined_valid_types: + table_name = self._parse_table_name(valid_type) + + # Quietly ignore bogus valid_types; more likely to DTRT + if table_name: + final_valid_types.append(valid_type) + type_terms.append(whoosh.query.Term(u'table', table_name)) + + return name, final_valid_types, whoosh.query.Or(type_terms) + + + def _parse_table_name(self, name): + """Takes a singular table name, table name, or table object and returns + the table name. + + Returns None for a bogus name. + """ + # Table object + if hasattr(name, '__tablename__'): + return getattr(name, '__tablename__') + + # Table name + for table in self.indexed_tables.values(): + if name in (table.__tablename__, table.__singlename__): + return table.__tablename__ + + # Bogus. Be nice and return dummy + return None + + def _whoosh_records_to_results(self, records, exact=True): + """Converts a list of whoosh's indexed records to LookupResult tuples + containing database objects. + """ + # XXX this 'exact' thing is getting kinda leaky. would like a better + # way to handle it, since only lookup() cares about fuzzy results + seen = {} + results = [] + for record in records: + # Skip dupes + seen_key = record['table'], record['row_id'] + if seen_key in seen: + continue + seen[seen_key] = True + + cls = self.indexed_tables[record['table']] + obj = self.session.query(cls).get(record['row_id']) + + results.append(LookupResult(object=obj, + indexed_name=record['name'], + name=record['display_name'], + language=record['language'], + iso3166=record['iso3166'], + exact=exact)) + + return results + + + def lookup(self, input, valid_types=[], exact_only=False): + """Attempts to find some sort of object, given a name. + + Returns a list of named (object, name, language, iso3166, exact) + tuples. `object` is a database object, `name` is the name under which + the object was found, `language` and `iso3166` are the name and country + code of the language in which the name was found, and `exact` is True + iff this was an + exact match. + + This function currently ONLY does fuzzy matching if there are no exact + matches. + + Formes are not returned unless requested; "Shaymin" will return only + grass Shaymin. + + Extraneous whitespace is removed with extreme prejudice. + + Recognizes: + - Names: "Eevee", "Surf", "Run Away", "Payapa Berry", etc. + - Foreign names: "Iibui", "Eivui" + - Fuzzy names in whatever language: "Evee", "Ibui" + - IDs: "133", "192", "250" + Also: + - Type restrictions. "type:psychic" will only return the type. This + is how to make ID lookup useful. Multiple type specs can be entered + with commas, as "move,item:1". If `valid_types` are provided, any + type prefix will be ignored. + - Alternate formes can be specified merely like "wash rotom". + + `input` + Name of the thing to look for. + + `valid_types` + A list of table objects or names, e.g., `['pokemon', 'moves']`. If + this is provided, only results in one of the given tables will be + returned. + + `exact_only` + If True, only exact matches are returned. If set to False (the + default), and the provided `name` doesn't match anything exactly, + spelling correction will be attempted. + """ + + name = self.normalize_name(input) + exact = True + form = None + + # Pop off any type prefix and merge with valid_types + name, merged_valid_types, type_term = \ + self._apply_valid_types(name, valid_types) + + # Random lookup + if name == 'random': + return self.random_lookup(valid_types=merged_valid_types) + + # Do different things depending what the query looks like + # Note: Term objects do an exact match, so we don't have to worry about + # a query parser tripping on weird characters in the input + try: + # Let Python try to convert to a number, so 0xff works + name_as_number = int(name, base=0) + except ValueError: + # Oh well + name_as_number = None + + if '*' in name or '?' in name: + exact_only = True + query = whoosh.query.Wildcard(u'name', name) + elif name_as_number is not None: + # Don't spell-check numbers! + exact_only = True + query = whoosh.query.Term(u'row_id', unicode(name_as_number)) + else: + # Not an integer + query = whoosh.query.Term(u'name', name) + + if type_term: + query = query & type_term + + + ### Actual searching + searcher = self.index.searcher() + # XXX is this kosher? docs say search() takes a weighting arg, but it + # certainly does not + searcher.weighting = LanguageWeighting() + results = searcher.search(query, + limit=self.INTERMEDIATE_LOOKUP_RESULTS) + + # Look for some fuzzy matches if necessary + if not exact_only and not results: + exact = False + results = [] + + for suggestion in self.speller.suggest( + name, self.INTERMEDIATE_LOOKUP_RESULTS): + + query = whoosh.query.Term('name', suggestion) + if type_term: + query = query & type_term + results.extend(searcher.search(query)) + + ### Convert results to db objects + objects = self._whoosh_records_to_results(results, exact=exact) + + # Only return up to 10 matches; beyond that, something is wrong. We + # strip out duplicate entries above, so it's remotely possible that we + # should have more than 10 here and lost a few. The speller returns 25 + # to give us some padding, and should avoid that problem. Not a big + # deal if we lose the 25th-most-likely match anyway. + return objects[:self.MAX_LOOKUP_RESULTS] + + + def random_lookup(self, valid_types=[]): + """Returns a random lookup result from one of the provided + `valid_types`. + """ + + tables = [] + for valid_type in valid_types: + table_name = self._parse_table_name(valid_type) + if table_name: + tables.append(self.indexed_tables[table_name]) + + if not tables: + # n.b.: It's possible we got a list of valid_types and none of them + # were valid, but this function is guaranteed to return + # *something*, so it politely selects from the entire index isntead + tables = self.indexed_tables.values() + + # Rather than create an array of many hundred items and pick randomly + # from it, just pick a number up to the total number of potential + # items, then pick randomly from that, and partition the whole range + # into chunks. This also avoids the slight problem that the index + # contains more rows (for languages) for some items than others. + # XXX ought to cache this (in the index?) if possible + total = 0 + partitions = [] + for table in tables: + count = self.session.query(table).count() + total += count + partitions.append((table, count)) + + n = random.randint(1, total) + while n > partitions[0][1]: + n -= partitions[0][1] + partitions.pop(0) + + return self.lookup(unicode(n), valid_types=[ partitions[0][0] ]) + + def prefix_lookup(self, prefix, valid_types=[]): + """Returns terms starting with the given exact prefix. + + Type prefixes are recognized, but no other name munging is done. + """ + + # Pop off any type prefix and merge with valid_types + prefix, merged_valid_types, type_term = \ + self._apply_valid_types(prefix, valid_types) + + query = whoosh.query.Prefix(u'name', self.normalize_name(prefix)) + + if type_term: + query = query & type_term + + searcher = self.index.searcher() + searcher.weighting = LanguageWeighting() + results = searcher.search(query) # XXX , limit=self.MAX_LOOKUP_RESULTS) + + return self._whoosh_records_to_results(results)