X-Git-Url: http://git.veekun.com/zzz-pokedex.git/blobdiff_plain/4fe48baffb5156027c6d186308998412554a6906..929cfc5f8a786c162a0cbdbb30153fe8d1984d7b:/pokedex/lookup.py diff --git a/pokedex/lookup.py b/pokedex/lookup.py index dcc3e7a..0fa91cb 100644 --- a/pokedex/lookup.py +++ b/pokedex/lookup.py @@ -1,45 +1,97 @@ # encoding: utf8 +from collections import namedtuple +import os, os.path +import pkg_resources import re from sqlalchemy.sql import func import whoosh +import whoosh.filedb.filestore +import whoosh.filedb.fileindex +import whoosh.index from whoosh.qparser import QueryParser +import whoosh.scoring import whoosh.spelling +from pokedex.db import connect import pokedex.db.tables as tables +from pokedex.roomaji import romanize + +__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 -index_bits = {} -def get_index(session): - """Returns (index, speller). +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. - Creates an index if one does not exist. + `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. """ - if index_bits: - return index_bits['index'], index_bits['speller'] + # Defaults + if not directory: + directory = pkg_resources.resource_filename('pokedex', + 'data/whoosh_index') + + if not session: + 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! + 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) + - store = whoosh.store.RamStorage() + # Create index schema = whoosh.fields.Schema( name=whoosh.fields.ID(stored=True), - spelling_name=whoosh.fields.ID(stored=True), table=whoosh.fields.STORED, - row_id=whoosh.fields.STORED, - language_id=whoosh.fields.STORED, + row_id=whoosh.fields.ID(stored=True), + language=whoosh.fields.STORED, ) - index = whoosh.index.Index(store, schema=schema, create=True) + 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) @@ -48,77 +100,142 @@ def get_index(session): 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.lower() - spelling_name = re.sub('[^a-z]', '', name) - writer.add_document(name=name, - spelling_name=spelling_name, - table=cls.__tablename__, - row_id=row.id) + writer.add_document(name=name, **row_key) + speller_entries.append((name, 1)) + + # Pokemon also get other languages + for foreign_name in getattr(row, 'foreign_names', []): + moonspeak = foreign_name.name.lower() + 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, + language=foreign_name.language.name, + **row_key) + speller_entries.append((moonspeak, 3)) + + # Add Roomaji too + if foreign_name.language.name == 'Japanese': + roomaji = romanize(foreign_name.name).lower() + writer.add_document(name=roomaji, language='Roomaji', + **row_key) + speller_entries.append((roomaji, 8)) + writer.commit() - ### Construct a spell-checker index + # 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 - # Can't use speller.add_field because it tries to intuit a frequency, and - # names are in an ID field, which seems to be immune to frequency. - # Not hard to add everything ourselves, though - reader = index.doc_reader() - speller.add_words([ _['spelling_name'] for _ in reader ]) - reader.close() - index_bits['index'] = index - index_bits['speller'] = speller - index_bits['store'] = store - return index_bits['index'], index_bits['speller'] +class LanguageWeighting(whoosh.scoring.Weighting): + """A scoring class that forces otherwise-equal English results to come + before foreign results. + """ + + def score(self, searcher, fieldnum, text, docnum, weight, QTF=1): + doc = searcher.stored_fields(docnum) + if doc['language'] == None: + # 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 + return weight * 0.95 + else: + # Everything else can drop down the totem pole + return weight * 0.9 -def lookup(session, name, exact_only=False): +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 (objects, exact) where `objects` is a list of database objects, and - `exact` is True iff the given name matched the returned objects exactly. + 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 ONLY does fuzzy matching if there are no exact matches. + This function currently ONLY does fuzzy matching if there are no exact + matches. Formes are not returned; "Shaymin" will return only grass Shaymin. - Currently recognizes: - - Pokémon names: "Eevee" + 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. """ + if not session: + session = connect() + + if indices: + index, speller = indices + else: + index, speller = open_index() + + name = unicode(input).lower() exact = True - # Alas! We have to make three attempts to find anything with this index. - # First: Try an exact match for a name in the index. - # Second: Try an exact match for a stripped-down name in the index. - # Third: Get spelling suggestions. - # The spelling module apparently only indexes *words* -- that is, [a-z]+. - # So we have a separate field that contains the same name, stripped down to - # just [a-z]+. - # Unfortunately, exact matches aren't returned as spelling suggestions, so - # we also have to do a regular index match against this separate field. - # Otherwise, 'nidoran' will never match anything - index, speller = get_index(session) - - # Look for exact name - parser = QueryParser('name', schema=index.schema) - results = index.find(name.lower(), parser=parser) - - if not exact_only: - # Look for a match with a reduced a-z name - if not results: - parser = QueryParser('spelling_name', schema=index.schema) - results = index.find(name.lower(), parser=parser) - - # Look for some fuzzy matches - if not results: - results = [] - exact = False - - for suggestion in speller.suggest(name, 3): - results.extend( index.find(suggestion, parser=parser) ) - - # Convert results to db objects + # 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: @@ -130,6 +247,14 @@ def lookup(session, name, exact_only=False): cls = indexed_tables[result['table']] obj = session.query(cls).get(result['row_id']) - objects.append(obj) + objects.append(LookupResult(object=obj, + name=result['name'], + language=result['language'], + exact=exact)) - return objects, 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]