Sync from GitHub e41c1493d83f777145673b3f2273c84efb787b11

src/rag/input_handler.py

@@ -212,7 +212,11 @@ class InputHandler:
             and any(pattern in joined_words for pattern in InputHandler.KEYBOARD_MASH_PATTERNS)
         ):
             return True
-        if InputHandler._looks_like_nonsense_word(joined_words):
+        # Nonsense scoring on a concatenation only holds for compact input
+        # (e.g. a mash split by stray punctuation). On a real multi-word
+        # sentence it false-positives because word boundaries fabricate
+        # suspicious sequences and consonant runs.
+        if len(words) <= 2 and InputHandler._looks_like_nonsense_word(joined_words):
             return True
         if any(InputHandler._looks_like_nonsense_word(word) for word in words):
             return True
@@ -223,15 +227,18 @@ class InputHandler:
         ):
             return True
 
-        vowel_count = sum(1 for char in joined_words if char in InputHandler.VOWELS)
-        if len(joined_words) >= 8 and vowel_count / len(joined_words) <= 0.18:
-            return True
+        # Vowel-ratio is only a reliable gibberish signal for a single token;
+        # across a whole sentence it is meaningless.
+        if len(words) == 1:
+            vowel_count = sum(1 for char in joined_words if char in InputHandler.VOWELS)
+            if len(joined_words) >= 8 and vowel_count / len(joined_words) <= 0.18:
+                return True
 
+        # A long consonant run alone is a poor signal: German compounds such as
+        # "deutschsprachige" legitimately contain them. Only a complete absence
+        # of vowels in a long token is treated as gibberish here.
         for word in words:
-            if len(word) >= 8 and (
-                InputHandler._has_long_consonant_run(word)
-                or not any(char in InputHandler.VOWELS for char in word)
-            ):
+            if len(word) >= 8 and not any(char in InputHandler.VOWELS for char in word):
                 return True
 
         return False

tests/test_input_handler_gibberish.py

@@ -42,6 +42,12 @@ def test_process_input_rejects_probable_gibberish(message):
         "Zurich 8000",
         "leadership 2026",
         "Was sind die nächsten Schritte im Bewerbungsprozess?",
+        # Regression: real queries the gibberish heuristic wrongly rejected.
+        # German compound words contain long consonant runs ("deutschsprachige");
+        # long English sentences false-positived once their words were concatenated.
+        "Wie lange dauert der deutschsprachige EMBA HSG?",
+        "Give me a short overview of all three executive MBA programmes.",
+        "Welche Weiterbildungsmöglichkeiten bietet die Universität?",
         "Ich interessiere mich für ein MBA mit Schwerpunkt nachhaltige Unternehmensführung.\nberufsbegleitend.",
         "Ich habe gerade meinen Bachelor abgeschlossen und 2 Jahre Berufserfahrung. Kann ich mich für den Executive MBA bewerben?",
     ],