Random lookup algorithm is now more naive, but less broken.
[zzz-pokedex.git] / pokedex / lookup.py
index dcc3e7a..8488f21 100644 (file)
 # encoding: utf8
+import os, os.path
+import random
 import re
+import shutil
+import unicodedata
 
 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.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', 'indexed_name', 'name', 'language', 'iso639', 'iso3166', 'exact',
+])
+
+class UninitializedIndex(object):
+    class UninitializedIndexError(Exception):
+        pass
 
-# 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.Pokemon,
-    ]:
-    indexed_tables[cls.__tablename__] = cls
+    def __nonzero__(self):
+        """Dummy object should identify itself as False."""
+        return False
 
-index_bits = {}
-def get_index(session):
-    """Returns (index, speller).
+    def __bool__(self):
+        """Python 3000 version of the above.  Future-proofing rules!"""
+        return False
 
-    Creates an index if one does not exist.
+    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.
     """
 
-    if index_bits:
-        return index_bits['index'], index_bits['speller']
+    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)
+
+        # 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 language == u'Roomaji':
+            # Give Roomaji a little boost; it's most likely to be searched
+            return weight * 0.9
+        else:
+            # Everything else can drop down the totem pole
+            return weight * 0.8
+
 
-    store = whoosh.store.RamStorage()
-    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,
+class PokedexLookup(object):
+    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
+    # 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.PokemonForm,
+            tables.Type,
+        )
     )
 
-    index = whoosh.index.Index(store, schema=schema, create=True)
-    writer = index.writer()
 
-    # Index every name in all our tables of interest
-    for cls in indexed_tables.values():
-        q = session.query(cls)
+    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.
 
-        # Only index base Pokémon formes
-        if hasattr(cls, 'forme_base_pokemon_id'):
-            q = q.filter_by(forme_base_pokemon_id=None)
+        `directory`
+            Directory containing the index.  Defaults to a location within the
+            `pokedex` egg directory.
 
-        for row in q.yield_per(5):
-            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)
+        `session`
+            Used for creating the index and retrieving objects.  Defaults to an
+            attempt to connect to the default SQLite database installed by
+            `pokedex setup`.
+        """
 
-    writer.commit()
+        # By the time this returns, self.index, self.speller, and self.session
+        # must be set
 
-    ### Construct a spell-checker index
-    speller = whoosh.spelling.SpellChecker(index.storage)
+        # If a directory was not given, use the default
+        if directory is None:
+            directory = get_default_index_dir()
 
-    # 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()
+        self.directory = directory
 
-    index_bits['index'] = index
-    index_bits['speller'] = speller
-    index_bits['store'] = store
-    return index_bits['index'], index_bits['speller']
+        if session:
+            self.session = session
+        else:
+            self.session = connect()
 
-def lookup(session, name, exact_only=False):
-    """Attempts to find some sort of object, given a database session and name.
+        # 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
 
-    Returns (objects, exact) where `objects` is a list of database objects, and
-    `exact` is True iff the given name matched the returned objects exactly.
+        # 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."
+            )
 
-    This function ONLY does fuzzy matching if there are no exact matches.
+        # Create speller, and done
+        spell_store = whoosh.filedb.filestore.FileStorage(directory)
+        self.speller = whoosh.spelling.SpellChecker(spell_store,
+            **self.SPELLER_OPTIONS)
 
-    Formes are not returned; "Shaymin" will return only grass Shaymin.
 
-    Currently recognizes:
-    - Pokémon names: "Eevee"
-    """
+    def rebuild_index(self):
+        """Creates the index from scratch."""
 
-    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 = []
+        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,
+            iso639=whoosh.fields.ID(stored=True),
+            iso3166=whoosh.fields.ID(stored=True),
+            display_name=whoosh.fields.STORED,  # non-lowercased name
+        )
+
+        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 = set()
+        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, iso639, iso3166):
+                    normalized_name = self.normalize_name(name)
+
+                    writer.add_document(
+                        name=normalized_name, display_name=name,
+                        language=language, iso639=iso639, iso3166=iso3166,
+                        **row_key
+                    )
+
+                    speller_entries.add(normalized_name)
+
+
+                # Add the basic English name to the index
+                if cls == tables.Pokemon:
+                    # Don't re-add alternate forms of the same Pokémon; they'll
+                    # be added as Pokémon forms instead
+                    if not row.is_base_form:
+                        continue
+                elif cls == tables.PokemonForm:
+                    if row.name:
+                        add(row.pokemon_name, None, u'en', u'us')
+                    continue
+
+                # Some things also have other languages' names
+                # XXX other language form names..?
+                seen = set()
+                for language, name in getattr(row, 'name_map', {}).items():
+                    if name in seen:
+                        # Don't add the name again as a different
+                        # language; no point and it makes spell results
+                        # confusing
+                        continue
+                    seen.add(name)
+
+                    add(name, language.name,
+                              language.iso639,
+                              language.iso3166)
+
+                    # Add Roomaji too
+                    if language.identifier == 'ja':
+                        roomaji = romanize(name)
+                        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, mingram=2,
+            **self.SPELLER_OPTIONS)
+        self.speller.add_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 = []
+            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__')
+
+        # 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.get('language'),
+                                        iso639=record['iso639'],
+                                        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, 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.
+
+        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".
+        - 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 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
+            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
+        # 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 = []
+
+            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
+
+            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)
+
+        # Truncate and return
+        return objects[:max_results]
+
+
+    def random_lookup(self, valid_types=[]):
+        """Returns a random lookup result from one of the provided
+        `valid_types`.
+        """
+
+        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.
+            # XXX The vast majority of Pokémon forms are unnamed and unindexed,
+            #     which can produce blank results.  So skip them too for now.
+            if table_name and table_name != 'pokemon_forms':
+                table_names.append(table_name)
+
+        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 instead
+            table_names = self.indexed_tables.keys()
+            table_names.remove('pokemon_forms')
+
+        # Pick a random table, then pick a random item from it.  Small tables
+        # like Type will have an unnatural bias.  The alternative is that a
+        # simple search for "random" will do some eight queries, counting the
+        # rows in every single indexed table, and that's awful.
+        # XXX Can we improve on this, reasonably?
+        table_name = random.choice(table_names)
+        count = self.session.query(self.indexed_tables[table_name]).count()
+        id, = self.session.query(self.indexed_tables[table_name].id) \
+            .offset(random.randint(0, count - 1)) \
+            .first()
+
+        return self.lookup(unicode(id), valid_types=[table_name])
+
+    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)
 
-            for suggestion in speller.suggest(name, 3):
-                results.extend( index.find(suggestion, parser=parser) )
+        query = whoosh.query.Prefix(u'name', self.normalize_name(prefix))
 
-    # 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
+        if type_term:
+            query = query & type_term
 
-        cls = indexed_tables[result['table']]
-        obj = session.query(cls).get(result['row_id'])
-        objects.append(obj)
+        searcher = self.index.searcher()
+        searcher.weighting = LanguageWeighting()
+        results = searcher.search(query)  # XXX , limit=self.MAX_LOOKUP_RESULTS)
 
-    return objects, exact
+        return self._whoosh_records_to_results(results)