Fixed the hell out of all remaining form problems.

[zzz-pokedex.git] / pokedex / lookup.py

diff --git a/pokedex/lookup.py b/pokedex/lookup.py
index 37a50f3..69eedc4 100644 (file)
--- a/pokedex/lookup.py
+++ b/pokedex/lookup.py
@@ -1,7 +1,10 @@
 # encoding: utf8
 # encoding: utf8

+from collections import namedtuple

 import os, os.path
 import pkg_resources
 import os, os.path
 import pkg_resources

+import random

 import re
 import re

+import shutil

 
 from sqlalchemy.sql import func
 import whoosh
 
 from sqlalchemy.sql import func
 import whoosh


@@ -9,10 +12,17 @@ import whoosh.filedb.filestore
 import whoosh.filedb.fileindex
 import whoosh.index
 from whoosh.qparser import QueryParser
 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
 import whoosh.spelling
 
 from pokedex.db import connect
 import pokedex.db.tables as tables

+from pokedex.roomaji import romanize
+
+__all__ = ['open_index', 'lookup', 'random_lookup']
+
+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
 
 # Dictionary of table name => table class.
 # Need the table name so we can get the class from the table name after we


@@ -27,17 +37,6 @@ for cls in [
     ]:
     indexed_tables[cls.__tablename__] = cls
 
     ]:
     indexed_tables[cls.__tablename__] = cls
 

-# Dictionary of extra keys to file types of objects under, e.g. Pokémon can
-# also be looked up purely by number
-extra_keys = {
-    tables.Move: [
-        lambda row: u"move %d" % row.id,
-    ],
-    tables.Pokemon: [
-        lambda row: unicode(row.id),
-    ],
-}
-

 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.
 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.


@@ -59,7 +58,7 @@ def open_index(directory=None, session=None, recreate=False):
     # Defaults
     if not directory:
         directory = pkg_resources.resource_filename('pokedex',
     # Defaults
     if not directory:
         directory = pkg_resources.resource_filename('pokedex',

-                                                    'data/whoosh_index')
+                                                    'data/whoosh-index')

 
     if not session:
         session = connect()
 
     if not session:
         session = connect()


@@ -69,87 +68,209 @@ def open_index(directory=None, session=None, recreate=False):
     if directory_exists and not recreate:
         # Already exists; should be an index!
         try:
     if directory_exists and not recreate:
         # Already exists; should be an index!
         try:

-            index = whoosh.index.open_dir(directory, indexname='pokedex')
-            speller = whoosh.index.open_dir(directory, indexname='spelling')
+            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
 
             return index, speller
         except whoosh.index.EmptyIndexError as e:
             # Apparently not a real index.  Fall out of the if 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)
 
 
     if not directory_exists:
         os.mkdir(directory)
 
 

-    # Create index
+    ### Create index

     schema = whoosh.fields.Schema(
         name=whoosh.fields.ID(stored=True),
     schema = whoosh.fields.Schema(
         name=whoosh.fields.ID(stored=True),

-        table=whoosh.fields.STORED,
-        row_id=whoosh.fields.STORED,
+        table=whoosh.fields.ID(stored=True),
+        row_id=whoosh.fields.ID(stored=True),

         language=whoosh.fields.STORED,
         language=whoosh.fields.STORED,

+        iso3166=whoosh.fields.STORED,
+        display_name=whoosh.fields.STORED,  # non-lowercased name
+        forme_name=whoosh.fields.ID,

     )
 
     )
 

-    index = whoosh.index.create_in(directory, schema=schema,
-                                              indexname='pokedex')
+    index = whoosh.index.create_in(directory, schema=schema, indexname='MAIN')

     writer = index.writer()
 
     # Index every name in all our tables of interest
     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)
 
     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):
         for row in q.yield_per(5):

-            row_key = dict(table=cls.__tablename__, row_id=row.id)
-
-            # Spelling index only indexes strings of letters, alas, so we
-            # reduce every name to this to make the index work.  However, exact
-            # matches are not returned, so e.g. 'nidoran' would neither match
-            # exactly nor fuzzy-match.  Solution: add the spelling-munged name
-            # as a regular index row too.
-            name = row.name.lower()
-            writer.add_document(name=name, **row_key)
-
-            speller_entries.append(name)
-
-            for extra_key_func in extra_keys.get(cls, []):
-                extra_key = extra_key_func(row)
-                writer.add_document(name=extra_key, **row_key)
+            # XXX 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'XXX')
+
+            def add(name, language, iso3166, score):
+                writer.add_document(name=name.lower(), display_name=name,
+                                    language=language,
+                                    iso3166=iso3166,
+                                    **row_key)
+                speller_entries.append((name.lower(), 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
 
     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, indexname='spelling')
-    speller.add_words(speller_entries)
+    speller = whoosh.spelling.SpellChecker(index.storage)
+    speller.add_scored_words(speller_entries)

 
     return index, speller
 
 
 
     return index, speller
 
 

-def lookup(name, session=None, exact_only=False):
-    """Attempts to find some sort of object, given a database session and name.
+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

 
 

-    Returns (objects, exact) where `objects` is a list of database objects, and
-    `exact` is True iff the given name matched the returned objects exactly.
+rx_is_number = re.compile('^\d+$')

 
 

-    This function ONLY does fuzzy matching if there are no exact matches.
+LookupResult = namedtuple('LookupResult',
+                          ['object', 'name', 'language', 'iso3166', 'exact'])

 
 

-    Formes are not returned; "Shaymin" will return only grass Shaymin.
+def _parse_table_name(name):
+    """Takes a singular table name, table name, or table object and returns the
+    table name.

 
 

-    Currently recognizes:
-    - Pokémon names: "Eevee"
+    Returns None for a bogus name.
+    """
+    if hasattr(name, '__tablename__'):
+        return getattr(name, '__tablename__')
+    elif name in indexed_tables:
+        return name
+    elif name + 's' in indexed_tables:
+        return name + 's'
+    else:
+        # Bogus.  Be nice and return dummy
+        return None
+
+def _whoosh_records_to_results(records, session, 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

 
 

-    `name`
+        cls = indexed_tables[record['table']]
+        obj = 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(input, valid_types=[], 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, 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.
 
         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.
+

     `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.
 
     `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,
     `exact_only`
         If True, only exact matches are returned.  If set to False (the
         default), and the provided `name` doesn't match anything exactly,


@@ -159,38 +280,144 @@ def lookup(name, session=None, exact_only=False):
     if not session:
         session = connect()
 
     if not session:
         session = connect()
 

-    index, speller = open_index()
+    if indices:
+        index, speller = indices
+    else:
+        index, speller = open_index()

 
 

+    name = unicode(input).strip().lower()

     exact = True
     exact = True

-
-    # Look for exact name.  A Term object does an exact match, so we don't have
-    # to worry about a query parser tripping on weird characters in the input
+    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 random_lookup(indices=(index, speller),
+                             session=session,
+                             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'XXX')
+
+        # 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 = _parse_table_name(valid_type)
+        type_terms.append(whoosh.query.Term(u'table', table_name))
+
+    if type_terms:
+        query = query & whoosh.query.Or(type_terms)
+
+
+    ### Actual searching

     searcher = index.searcher()
     searcher = index.searcher()

-    query = whoosh.query.Term('name', name.lower())
-    results = searcher.search(query)
+    searcher.weighting = LanguageWeighting()  # XXX kosher?  docs say search()
+                                              # takes a weighting kw but it
+                                              # certainly does not
+    results = searcher.search(query, limit=INTERMEDIATE_LOOKUP_RESULTS)
+
+    # Look for some fuzzy matches if necessary
+    if not exact_only and not results:
+        exact = False
+        results = []
+
+        for suggestion in speller.suggest(name, INTERMEDIATE_LOOKUP_RESULTS):
+            query = whoosh.query.Term('name', suggestion)
+            results.extend(searcher.search(query))
+
+    ### Convert results to db objects
+    objects = _whoosh_records_to_results(results, session, 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[:MAX_LOOKUP_RESULTS]
+
+
+def random_lookup(valid_types=[], session=None, indices=None):
+    """Takes similar arguments as `lookup()`, but returns a random lookup
+    result from one of the provided `valid_types`.
+    """

 
 

-    if not exact_only:
-        # Look for some fuzzy matches
-        if not results:
-            exact = False
-            results = []
+    tables = []
+    for valid_type in valid_types:
+        table_name = _parse_table_name(valid_type)
+        if table_name:
+            tables.append(indexed_tables[table_name])

 
 

-            for suggestion in speller.suggest(name, 3):
-                query = whoosh.query.Term('name', suggestion)
-                results.extend(searcher.search(query))
+    if not tables:
+        tables = indexed_tables.values()

 
 

-    # 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
+    # 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
+    total = 0
+    partitions = []
+    for table in tables:
+        count = 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 lookup(unicode(n), valid_types=[ partitions[0][0] ],
+                  indices=indices, session=session)

 
 

-        cls = indexed_tables[result['table']]
-        obj = session.query(cls).get(result['row_id'])
-        objects.append(obj)
+def prefix_lookup(prefix, session=None, indices=None):
+    """Returns terms starting with the given exact prefix.
+
+    No special magic is currently done with the name; type prefixes are not
+    recognized.
+
+    `session` and `indices` are treated as with `lookup()`.
+    """
+
+    if not session:
+        session = connect()
+
+    if indices:
+        index, speller = indices
+    else:
+        index, speller = open_index()
+
+    query = whoosh.query.Prefix(u'name', prefix.lower())
+
+    searcher = index.searcher()
+    searcher.weighting = LanguageWeighting()
+    results = searcher.search(query)  # XXX , limit=MAX_LOOKUP_RESULTS)

 
 

-    return objects, exact
+    return _whoosh_records_to_results(results, session)