diff --git a/ai_diffusion/text.py b/ai_diffusion/text.py
index eacba1ac72..66dcd0a6bb 100644
--- a/ai_diffusion/text.py
+++ b/ai_diffusion/text.py
@@ -48,6 +48,7 @@ def replace(match: re.Match[str]):
                 lora_normalized = file.name.lower()
                 if input == lora_filename or input == lora_normalized:
                     lora_file = file
+                    break
 
         if not lora_file:
             error = _("LoRA not found") + f": {input}"
@@ -64,7 +65,7 @@ def replace(match: re.Match[str]):
                 raise Exception(error)
 
         loras.append(LoraInput(lora_file.id, lora_strength))
-        return lora_file.meta("lora_triggers", "")
+        return ""
 
     prompt = _pattern_lora.sub(replace, prompt)
     return prompt.strip(), loras
diff --git a/ai_diffusion/ui/autocomplete.py b/ai_diffusion/ui/autocomplete.py
index e3571ef9a9..a77a01f8f7 100644
--- a/ai_diffusion/ui/autocomplete.py
+++ b/ai_diffusion/ui/autocomplete.py
@@ -133,9 +133,6 @@ def _blend_colors(self, from_, to, factor):
 
 
 class PromptAutoComplete:
-    _completer: QCompleter
-    _completion_prefix: str
-    _completion_suffix: str
 
     def __init__(self, widget: QLineEdit):
         self._widget = widget
@@ -235,17 +232,20 @@ def check_completion(self):
         self._completer.complete(rect)
 
     def _insert_completion(self, completion):
-        if not self._current_text().startswith("<lora:"):
+        triggers = ""
+        if self._current_text().startswith("<lora:"):
+            if file := root.files.loras.find(f"{completion}.safetensors"):
+                triggers = " " + file.meta("lora_triggers", "")
+        else:  # tag completion
             # escape () in tags so they won't be interpreted as prompt weights
             completion = completion.replace("(", "\\(").replace(")", "\\)")
         text = self._widget.text()
         pos = self._widget.cursorPosition()
-        prefix_len = len(self._completion_prefix)
-        text = text[: pos - prefix_len] + completion + self._completion_suffix + text[pos:]
+        start_pos = pos - len(self._completion_prefix)
+        fill = completion + self._completion_suffix + triggers
+        text = text[:start_pos] + fill + text[pos:]
         self._widget.setText(text)
-        self._widget.setCursorPosition(
-            pos - prefix_len + len(completion) + len(self._completion_suffix)
-        )
+        self._widget.setCursorPosition(start_pos + len(fill))
 
     @property
     def is_active(self):
diff --git a/tests/test_text.py b/tests/test_text.py
index e7ebba225e..0117cabcaa 100644
--- a/tests/test_text.py
+++ b/tests/test_text.py
@@ -79,7 +79,7 @@ def test_extract_loras_meta():
     loras.set_meta(lora, "lora_triggers", "zippity")
 
     assert extract_loras("a ship <lora:zap> zap", loras) == (
-        "a ship zippity zap",
+        "a ship  zap",  # triggers are inserted on auto-complete, not at extraction
         [LoraInput(lora.id, 0.5)],
     )