Totally overhauled lookup to use a class.
[zzz-pokedex.git] / pokedex / lookup.py
index dcc3e7a..1363c7a 100644 (file)
 # 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
+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
 
-# 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
+__all__ = ['PokedexLookup']
 
-index_bits = {}
-def get_index(session):
-    """Returns (index, speller).
 
-    Creates an index if one does not exist.
+rx_is_number = re.compile('^\d+$')
+
+LookupResult = namedtuple('LookupResult',
+                          ['object', 'name', 'language', 'iso3166', 'exact'])
+
+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 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 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
+
+
+class PokedexLookup(object):
+    INTERMEDIATE_LOOKUP_RESULTS = 25
+    MAX_LOOKUP_RESULTS = 10
 
-    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,
+    # 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.Move,
+            tables.Pokemon,
+            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, recreate=False):
+        """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`
+            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`.
 
-    writer.commit()
+        `recreate`
+            If set to True, the whoosh index will be created even if it already
+            exists.
+        """
 
-    ### Construct a spell-checker index
-    speller = whoosh.spelling.SpellChecker(index.storage)
+        # By the time this returns, self.index, self.speller, and self.session
+        # must be set
 
-    # 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()
+        # Defaults
+        if not directory:
+            directory = pkg_resources.resource_filename('pokedex',
+                                                        'data/whoosh-index')
 
-    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
+        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
 
-    Returns (objects, exact) where `objects` is a list of database objects, and
-    `exact` is True iff the given name matched the returned objects exactly.
+        # 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
 
-    This function ONLY does fuzzy matching if there are no exact matches.
+        if not directory_exists:
+            os.mkdir(directory)
 
-    Formes are not returned; "Shaymin" will return only grass Shaymin.
 
-    Currently recognizes:
-    - Pokémon names: "Eevee"
-    """
+        ### Create index
+        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
+            forme_name=whoosh.fields.ID,
+        )
 
-    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 = []
+        self.index = whoosh.index.create_in(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 = 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_key = dict(table=unicode(cls.__tablename__),
+                               row_id=unicode(row.id),
+                               forme_name=u'__empty__')
+
+                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))
+
+                # If this is a form, mark it as such
+                if getattr(row, 'forme_base_pokemon_id', None):
+                    row_key['forme_name'] = row.forme_name
+
+                name = row.name
+                add(name, None, u'us', 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
+
+                    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 _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.
+        """
+        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
+
+    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,
+                                        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
+
+        # 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]
+
+        # Random lookup
+        if name == 'random':
+            return self.random_lookup(valid_types=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
+        if '*' in name or '?' in name:
+            exact_only = True
+            query = whoosh.query.Wildcard(u'name', name)
+        elif 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) \
+                  & whoosh.query.Term(u'forme_name', u'__empty__')
+
+            # 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)
+
+
+        ### 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)
+                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):
+        """Returns terms starting with the given exact prefix.
 
-            for suggestion in speller.suggest(name, 3):
-                results.extend( index.find(suggestion, parser=parser) )
+        No special magic is currently done with the name; type prefixes are not
+        recognized.
+        """
 
-    # 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
+        query = whoosh.query.Prefix(u'name', self.normalize_name(prefix))
 
-        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)