Improve handling of non-ASCII characters in user directory search (#15143)

* Fix a long-standing bug where non-ASCII characters in search terms,
  including accented letters, would not match characters in a different
  case.
* Fix a long-standing bug where search terms using combining accents
  would not match display names using precomposed accents and vice
  versa.

To fully take effect, the user directory must be rebuilt after this
change.

Fixes #14630.

Signed-off-by: Sean Quah <seanq@matrix.org>

1 files changed, 50 insertions, 2 deletions

diff --git a/synapse/storage/databases/main/user_directory.py b/synapse/storage/databases/main/user_directory.py
index c3f2b61bd5..f16a509ac4 100644
--- a/synapse/storage/databases/main/user_directory.py
+++ b/synapse/storage/databases/main/user_directory.py
@@ -14,6 +14,7 @@
 
 import logging
 import re
+import unicodedata
 from typing import (
     TYPE_CHECKING,
     Iterable,
@@ -490,6 +491,11 @@ class UserDirectoryBackgroundUpdateStore(StateDeltasStore):
                 values={"display_name": display_name, "avatar_url": avatar_url},
             )
 
+            # The display name that goes into the database index.
+            index_display_name = display_name
+            if index_display_name is not None:
+                index_display_name = _filter_text_for_index(index_display_name)
+
             if isinstance(self.database_engine, PostgresEngine):
                 # We weight the localpart most highly, then display name and finally
                 # server name
@@ -507,11 +513,15 @@ class UserDirectoryBackgroundUpdateStore(StateDeltasStore):
                         user_id,
                         get_localpart_from_id(user_id),
                         get_domain_from_id(user_id),
-                        display_name,
+                        index_display_name,
                     ),
                 )
             elif isinstance(self.database_engine, Sqlite3Engine):
-                value = "%s %s" % (user_id, display_name) if display_name else user_id
+                value = (
+                    "%s %s" % (user_id, index_display_name)
+                    if index_display_name
+                    else user_id
+                )
                 self.db_pool.simple_upsert_txn(
                     txn,
                     table="user_directory_search",
@@ -896,6 +906,41 @@ class UserDirectoryStore(UserDirectoryBackgroundUpdateStore):
         return {"limited": limited, "results": results[0:limit]}
 
 
+def _filter_text_for_index(text: str) -> str:
+    """Transforms text before it is inserted into the user directory index, or searched
+    for in the user directory index.
+
+    Note that the user directory search table needs to be rebuilt whenever this function
+    changes.
+    """
+    # Lowercase the text, to make searches case-insensitive.
+    # This is necessary for both PostgreSQL and SQLite. PostgreSQL's
+    # `to_tsquery/to_tsvector` functions don't lowercase non-ASCII characters when using
+    # the "C" collation, while SQLite just doesn't lowercase non-ASCII characters at
+    # all.
+    text = text.lower()
+
+    # Normalize the text. NFKC normalization has two effects:
+    #  1. It canonicalizes the text, ie. maps all visually identical strings to the same
+    #     string. For example, ["e", "◌́"] is mapped to ["é"].
+    #  2. It maps strings that are roughly equivalent to the same string.
+    #     For example, ["dž"] is mapped to ["d", "ž"], ["①"] to ["1"] and ["i⁹"] to
+    #     ["i", "9"].
+    text = unicodedata.normalize("NFKC", text)
+
+    # Note that nothing is done to make searches accent-insensitive.
+    # That could be achieved by converting to NFKD form instead (with combining accents
+    # split out) and filtering out combining accents using `unicodedata.combining(c)`.
+    # The downside of this may be noisier search results, since search terms with
+    # explicit accents will match characters with no accents, or completely different
+    # accents.
+    #
+    # text = unicodedata.normalize("NFKD", text)
+    # text = "".join([c for c in text if not unicodedata.combining(c)])
+
+    return text
+
+
 def _parse_query_sqlite(search_term: str) -> str:
     """Takes a plain unicode string from the user and converts it into a form
     that can be passed to database.
@@ -905,6 +950,7 @@ def _parse_query_sqlite(search_term: str) -> str:
     We specifically add both a prefix and non prefix matching term so that
     exact matches get ranked higher.
     """
+    search_term = _filter_text_for_index(search_term)
 
     # Pull out the individual words, discarding any non-word characters.
     results = _parse_words(search_term)
@@ -917,6 +963,8 @@ def _parse_query_postgres(search_term: str) -> Tuple[str, str, str]:
     We use this so that we can add prefix matching, which isn't something
     that is supported by default.
     """
+    search_term = _filter_text_for_index(search_term)
+
     escaped_words = []
     for word in _parse_words(search_term):
         # Postgres tsvector and tsquery quoting rules: