X-Git-Url: http://git.veekun.com/zzz-pokedex.git/blobdiff_plain/4603adc7c77a5c04b9f4a1ffca75c94442e51752..9a3152efb04b1c90fcbe786882775ed5984f24ce:/pokedex/lookup.py diff --git a/pokedex/lookup.py b/pokedex/lookup.py index 7b10974..98e69c4 100644 --- a/pokedex/lookup.py +++ b/pokedex/lookup.py @@ -1,7 +1,5 @@ # encoding: utf8 -from collections import namedtuple import os, os.path -import pkg_resources import random import re import shutil @@ -16,39 +14,83 @@ 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__ = ['PokedexLookup'] rx_is_number = re.compile('^\d+$') -LookupResult = namedtuple('LookupResult', - ['object', 'name', 'language', 'iso3166', 'exact']) +LookupResult = namedtuple('LookupResult', [ + 'object', 'indexed_name', 'name', 'language', 'iso639', 'iso3166', 'exact', +]) + +class UninitializedIndex(object): + class UninitializedIndexError(Exception): + pass + + 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 before foreign results. """ + def __init__(self, extra_weights={}, *args, **kwargs): + """`extra_weights` may be a dictionary of weights which will be + factored in. + + Intended for use with spelling corrections, which come along with their + own weightings. + """ + self.extra_weights = extra_weights + super(LanguageWeighting, self).__init__(*args, **kwargs) + def score(self, searcher, fieldnum, text, docnum, weight, QTF=1): doc = searcher.stored_fields(docnum) - if doc['language'] == None: + + # Apply extra weight + weight = weight * self.extra_weights.get(text, 1.0) + + language = doc.get('language') + if language is None: # English (well, "default"); leave it at 1 return weight - elif doc['language'] == u'Roomaji': + elif language == u'Roomaji': # Give Roomaji a little boost; it's most likely to be searched - return weight * 0.95 + return weight * 0.9 else: # Everything else can drop down the totem pole - return weight * 0.9 + return weight * 0.8 class PokedexLookup(object): - INTERMEDIATE_LOOKUP_RESULTS = 25 - MAX_LOOKUP_RESULTS = 10 + MAX_FUZZY_RESULTS = 10 + MAX_EXACT_RESULTS = 43 + INTERMEDIATE_FACTOR = 2 + + # The speller only checks how much the input matches a word; there can be + # all manner of extra unmatched junk, and it won't affect the weighting. + # To compensate, greatly boost the weighting of matches at the beginning + # and end, so nearly-full-word-matches are much better + SPELLER_OPTIONS = dict(booststart=10.0, boostend=9.0) # Dictionary of table name => table class. # Need the table name so we can get the class from the table name after we @@ -58,14 +100,16 @@ class PokedexLookup(object): for cls in ( tables.Ability, tables.Item, + tables.Location, tables.Move, + tables.Nature, tables.Pokemon, tables.Type, ) ) - def __init__(self, directory=None, session=None, recreate=False): + 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. @@ -74,22 +118,19 @@ class PokedexLookup(object): `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. + 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 - # Defaults - if not directory: - directory = pkg_resources.resource_filename('pokedex', - 'data/whoosh-index') + # 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 @@ -97,114 +138,118 @@ class PokedexLookup(object): self.session = connect() # 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! Bam, done. - try: - self.index = whoosh.index.open_dir(directory, indexname='MAIN') - spell_store = whoosh.filedb.filestore.FileStorage(directory) - self.speller = whoosh.spelling.SpellChecker(spell_store) - return - except whoosh.index.EmptyIndexError as e: - # Apparently not a real index. Fall out and create it - pass - - # Delete and start over if we're going to bail anyway. - if directory_exists and recreate: - # Be safe and only delete if it looks like a whoosh index, i.e., - # everything starts with _ - if all(f[0] == '_' for f in os.listdir(directory)): - shutil.rmtree(directory) - directory_exists = False - - if not directory_exists: - os.mkdir(directory) - - - ### Create 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, + **self.SPELLER_OPTIONS) + + + 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, + iso639=whoosh.fields.ID(stored=True), + iso3166=whoosh.fields.ID(stored=True), display_name=whoosh.fields.STORED, # non-lowercased name - forme_name=whoosh.fields.ID, ) - self.index = whoosh.index.create_in(directory, schema=schema, - indexname='MAIN') + if os.path.exists(self.directory): + # create_in() isn't totally reliable, so just nuke whatever's there + # manually. Try to be careful about this... + for f in os.listdir(self.directory): + if re.match('^_?(MAIN|SPELL)_', f): + os.remove(os.path.join(self.directory, f)) + else: + 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 = [] + speller_entries = set() for cls in self.indexed_tables.values(): - q = session.query(cls) + q = self.session.query(cls) for row in q.yield_per(5): - # Need to give forme_name a dummy value because I can't - # search for explicitly empty fields. Boo. - row_keys = [ - dict(table=unicode(cls.__tablename__), - row_id=unicode(row.id), - forme_name=u'__empty__') - ] - - # If this is a form, mark it as such - # XXX foreign form names...? - if getattr(row, 'forme_name', None): - # ...but if it's also the *default* form, index the name - # bare too - if not getattr(row, 'forme_base_pokemon_id', None): - new_key = row_keys[0].copy() - row_keys.append(new_key) - - row_keys[0]['forme_name'] = row.forme_name - - def add(name, language, iso3166, score): + row_key = dict(table=unicode(cls.__tablename__), + row_id=unicode(row.id)) + + def add(name, language, iso639, iso3166): normalized_name = self.normalize_name(name) - for row_key in row_keys: - writer.add_document( - name=normalized_name, display_name=name, - language=language, iso3166=iso3166, - **row_key - ) - speller_entries.append((normalized_name, score)) + writer.add_document( + name=normalized_name, display_name=name, + language=language, iso639=iso639, iso3166=iso3166, + **row_key + ) + + speller_entries.add(normalized_name) - name = row.name - add(name, None, u'us', 1) + # 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'en', u'us') - # Pokemon also get other languages + # 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'en', u'us') + else: + add(row.name, None, u'en', u'us') + + # 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 name == moonspeak: + 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) + foreign_name.language.iso639, + foreign_name.language.iso3166) # Add Roomaji too if foreign_name.language.name == 'Japanese': roomaji = romanize(foreign_name.name) - add(roomaji, u'Roomaji', u'jp', 8) + add(roomaji, u'Roomaji', u'ja', u'jp') 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) + self.speller = whoosh.spelling.SpellChecker(self.index.storage, mingram=2, + **self.SPELLER_OPTIONS) + self.speller.add_words(speller_entries) def normalize_name(self, name): @@ -230,21 +275,102 @@ class PokedexLookup(object): 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 = [] + for prefix in prefixes: + prefix = prefix.strip() + if prefix: + user_valid_types.append(prefix) + + # 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. + # And, just to complicate matters: "type" and language need to be + # considered separately. + def merge_requirements(func): + user = filter(func, user_valid_types) + system = filter(func, valid_types) + + if user and system: + merged = list(set(user) & set(system)) + if merged: + return merged + else: + # No overlap; use the system restrictions + return system + else: + # One or the other is blank; use the one that's not + return user or system + + # @foo means language must be foo; otherwise it's a table name + lang_requirements = merge_requirements(lambda req: req[0] == u'@') + type_requirements = merge_requirements(lambda req: req[0] != u'@') + all_requirements = lang_requirements + type_requirements + + # Construct the term + lang_terms = [] + for lang in lang_requirements: + # Allow for either country or language codes + lang_code = lang[1:] + lang_terms.append(whoosh.query.Term(u'iso639', lang_code)) + lang_terms.append(whoosh.query.Term(u'iso3166', lang_code)) + + type_terms = [] + for type in type_requirements: + table_name = self._parse_table_name(type) + + # Quietly ignore bogus valid_types; more likely to DTRT + if table_name: + type_terms.append(whoosh.query.Term(u'table', table_name)) + + # Combine both kinds of restriction + all_terms = [] + if type_terms: + all_terms.append(whoosh.query.Or(type_terms)) + if lang_terms: + all_terms.append(whoosh.query.Or(lang_terms)) + + return name, all_requirements, whoosh.query.And(all_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__') - elif name in self.indexed_tables: - return name - elif name + 's' in self.indexed_tables: - return name + 's' - else: - # Bogus. Be nice and return dummy - return None + + # 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 @@ -265,8 +391,10 @@ class PokedexLookup(object): 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'], + language=record.get('language'), + iso639=record['iso639'], iso3166=record['iso3166'], exact=exact)) @@ -276,12 +404,11 @@ class PokedexLookup(object): 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. + Returns a list of named (object, name, language, iso639, iso3166, + exact) tuples. `object` is a database object, `name` is the name under + which the object was found, `language` and the two isos are the name + and country codes 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. @@ -299,17 +426,19 @@ class PokedexLookup(object): 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. + with commas, as "move,item:1". + - Language restrictions. "@fr:charge" will only return Tackle, which + is called "Charge" in French. These can be combined with type + restrictions, e.g., "@fr,move:charge". - 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. + A list of type or language restrictions, e.g., `['pokemon', + '@ja']`. 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 @@ -321,83 +450,96 @@ class PokedexLookup(object): exact = True form = None - # Remove any type prefix (pokemon:133) before constructing a query - if ':' in name: - prefix_chunk, name = name.split(':', 1) - name = name.strip() - - if not valid_types: - # Only use types from the query string if none were explicitly - # provided - prefixes = prefix_chunk.split(',') - valid_types = [_.strip() for _ in prefixes] + # 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=valid_types) + 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 rx_is_number.match(name): + elif name_as_number is not None: # Don't spell-check numbers! exact_only = True - query = whoosh.query.Term(u'row_id', name) + query = whoosh.query.Term(u'row_id', unicode(name_as_number)) else: # Not an integer - query = whoosh.query.Term(u'name', name) \ - & whoosh.query.Term(u'forme_name', u'__empty__') + query = whoosh.query.Term(u'name', name) - # If there's a space in the input, this might be a form - if ' ' in name: - form, formless_name = name.split(' ', 1) - form_query = whoosh.query.Term(u'name', formless_name) \ - & whoosh.query.Term(u'forme_name', form) - query = query | form_query - - ### Filter by type of object - type_terms = [] - for valid_type in valid_types: - table_name = self._parse_table_name(valid_type) - if table_name: - # Quietly ignore bogus valid_types; more likely to DTRT - type_terms.append(whoosh.query.Term(u'table', table_name)) - - if type_terms: - query = query & whoosh.query.Or(type_terms) + 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) + # Limits; result limits are constants, and intermediate results (before + # duplicate items are stripped out) are capped at the result limit + # times another constant. + # Fuzzy are capped at 10, beyond which something is probably very + # wrong. Exact matches -- that is, wildcards and ids -- are far less + # constrained. + # Also, exact matches are sorted by name, since weight doesn't matter. + sort_by = dict() + if exact_only: + max_results = self.MAX_EXACT_RESULTS + sort_by['sortedby'] = (u'table', u'name') + else: + max_results = self.MAX_FUZZY_RESULTS + + searcher = self.index.searcher(weighting=LanguageWeighting()) + results = searcher.search( + query, + limit=int(max_results * self.INTERMEDIATE_FACTOR), + **sort_by + ) # 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): + fuzzy_query_parts = [] + fuzzy_weights = {} + min_weight = [None] + for suggestion, _, weight in self.speller.suggestions_and_scores(name): + # Only allow the top 50% of scores; otherwise there will always + # be a lot of trailing junk + if min_weight[0] is None: + min_weight[0] = weight * 0.5 + elif weight < min_weight[0]: + break + + fuzzy_query_parts.append(whoosh.query.Term('name', suggestion)) + fuzzy_weights[suggestion] = weight - query = whoosh.query.Term('name', suggestion) - results.extend(searcher.search(query)) + if not fuzzy_query_parts: + # Nothing at all; don't try querying + return [] + + fuzzy_query = whoosh.query.Or(fuzzy_query_parts) + if type_term: + fuzzy_query = fuzzy_query & type_term + + searcher.weighting = LanguageWeighting(extra_weights=fuzzy_weights) + results = searcher.search(fuzzy_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] + # Truncate and return + return objects[:max_results] def random_lookup(self, valid_types=[]): @@ -405,17 +547,18 @@ class PokedexLookup(object): `valid_types`. """ - tables = [] + table_names = [] for valid_type in valid_types: table_name = self._parse_table_name(valid_type) + # Skip anything not recognized. Could be, say, a language code if table_name: - tables.append(self.indexed_tables[table_name]) + table_names.append(table_name) - if not tables: + if not table_names: # 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() + # *something*, so it politely selects from the entire index instead + table_names = self.indexed_tables.keys() # 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 @@ -425,10 +568,10 @@ class PokedexLookup(object): # XXX ought to cache this (in the index?) if possible total = 0 partitions = [] - for table in tables: - count = self.session.query(table).count() + for table_name in table_names: + count = self.session.query(self.indexed_tables[table_name]).count() total += count - partitions.append((table, count)) + partitions.append((table_name, count)) n = random.randint(1, total) while n > partitions[0][1]: @@ -437,15 +580,21 @@ class PokedexLookup(object): return self.lookup(unicode(n), valid_types=[ partitions[0][0] ]) - def prefix_lookup(self, prefix): + def prefix_lookup(self, prefix, valid_types=[]): """Returns terms starting with the given exact prefix. - No special magic is currently done with the name; type prefixes are not - recognized. + 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)