diff --git a/py/autotag.py b/py/autotag.py
index 33170526..0c807f06 100644
--- a/py/autotag.py
+++ b/py/autotag.py
@@ -16,6 +16,12 @@ from PIL import Image
 # A loader takes (model_type, use_gpu) and returns ("gguf" | "wd14" | "hf", backend)
 ModelLoader = Callable[[str, bool], Tuple[str, Any]]
 
+# Suggestions keep only names a Save accepts (the vocabulary tag rules).
+try:
+    from ..database.vocabulary import validate_vocabulary_tag_name
+except ImportError:
+    from database.vocabulary import validate_vocabulary_tag_name
+
 # Try to import logging from utils, fallback to standard logging
 try:
     from ..utils.logging_config import get_logger
@@ -795,8 +801,10 @@ class AutoTagService:
         WD14 pairs carry the real probability, best first; language models
         (GGUF, HF) have no score, so their tags keep the output order with
         ``None``. Names are distinct case-insensitively (first one wins).
-        ``prompt`` is used for this call only; ``custom_prompt`` is never
-        changed.
+        A name Save would reject (over 50 characters, a control character
+        such as a newline, a comma) is skipped before the dedupe and the
+        ``limit`` count. ``prompt`` is used for this call only;
+        ``custom_prompt`` is never changed.
         """
         if type(limit) is not int or not 1 <= limit <= MAX_SUGGESTION_LIMIT:
             raise ValueError(
@@ -807,7 +815,11 @@ class AutoTagService:
         )
         suggestions: List[Tuple[str, Optional[float]]] = []
         seen = set()
-        for name, score in scored:
+        for raw, score in scored:
+            try:
+                name = validate_vocabulary_tag_name(raw)
+            except ValueError:
+                continue
             key = name.casefold()
             if key in seen:
                 continue
diff --git a/tests/test_autotag_engine.py b/tests/test_autotag_engine.py
index a16f8ba2..c76217ab 100644
--- a/tests/test_autotag_engine.py
+++ b/tests/test_autotag_engine.py
@@ -387,6 +387,22 @@ class TestTagSuggestions(EngineTestCase):
             suggestions, [("zebra", None), ("apple", None), ("mango_tree", None)]
         )
 
+    def test_suggestions_skip_names_save_would_reject_before_the_limit(self):
+        from database.annotations import normalize_image_tags
+
+        too_long = "x" * 60
+        service = self._gguf(f"{too_long}, sky\ncloud, zebra, apple, mango, kiwi, plum")
+
+        suggestions = service.generate_tag_suggestions(
+            str(self.image_path), limit=3, prompt="tags please"
+        )
+
+        # The 60-character part and the part holding a newline are dropped
+        # before the count, so three valid chips still come back.
+        names = [name for name, _ in suggestions]
+        self.assertEqual(names, ["zebra", "apple", "mango"])
+        self.assertEqual(normalize_image_tags(names), names)
+
     def test_generate_tags_return_shape_unchanged(self):
         service = self._wd14()
         self.assertEqual(