From 49c4c0f2fda1d0f6f39c90c934e92932d911c955 Mon Sep 17 00:00:00 2001
From: NIK-TIGER-BILL <nik.tiger.bill@github.com>
Date: Mon, 1 Jun 2026 23:04:22 +0000
Subject: [PATCH] =?UTF-8?q?fix:=20prevent=20N=C3=97M=20reply=20explosion?=
 =?UTF-8?q?=20in=20HuggingFaceLocalGenerator=20with=20multiple=20stop=5Fwo?=
 =?UTF-8?q?rds?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Signed-off-by: NIK-TIGER-BILL <nik.tiger.bill@github.com>
---
 .../components/generators/hugging_face_local.py  |  3 ++-
 ...top-words-cross-product-31d12d96c57c442b.yaml |  4 ++++
 .../test_hugging_face_local_generator.py         | 16 ++++++++++++++++
 3 files changed, 22 insertions(+), 1 deletion(-)
 create mode 100644 releasenotes/notes/fix-hf-local-stop-words-cross-product-31d12d96c57c442b.yaml

diff --git a/haystack/components/generators/hugging_face_local.py b/haystack/components/generators/hugging_face_local.py
index 290bc1a11b..f12b7838bc 100644
--- a/haystack/components/generators/hugging_face_local.py
+++ b/haystack/components/generators/hugging_face_local.py
@@ -260,6 +260,7 @@ def run(
 
         if self.stop_words:
             # the output of the pipeline includes the stop word
-            replies = [reply.replace(stop_word, "").rstrip() for reply in replies for stop_word in self.stop_words]
+            for stop_word in self.stop_words:
+                replies = [reply.replace(stop_word, "").rstrip() for reply in replies]
 
         return {"replies": replies}
diff --git a/releasenotes/notes/fix-hf-local-stop-words-cross-product-31d12d96c57c442b.yaml b/releasenotes/notes/fix-hf-local-stop-words-cross-product-31d12d96c57c442b.yaml
new file mode 100644
index 0000000000..312f3e1240
--- /dev/null
+++ b/releasenotes/notes/fix-hf-local-stop-words-cross-product-31d12d96c57c442b.yaml
@@ -0,0 +1,4 @@
+---
+fixes:
+  - |
+    Fixes ``HuggingFaceLocalGenerator`` so that when multiple ``stop_words`` are configured the generator returns exactly N replies instead of N×M replies. Previously the list comprehension created a cross-product, duplicating replies and leaving some stop words unstripped.
diff --git a/test/components/generators/test_hugging_face_local_generator.py b/test/components/generators/test_hugging_face_local_generator.py
index 62779e855b..836dd94684 100644
--- a/test/components/generators/test_hugging_face_local_generator.py
+++ b/test/components/generators/test_hugging_face_local_generator.py
@@ -420,6 +420,22 @@ def test_run_stop_words_removal(self):
         results = generator.run(prompt="irrelevant")
         assert results == {"replies": ["Hello"]}
 
+    def test_run_multiple_stop_words_removal(self):
+        """Test that multiple stop words are all removed without producing N*M replies."""
+        generator = HuggingFaceLocalGenerator(
+            model="Qwen/Qwen3-0.6B", task="text-generation", stop_words=["STOP", "END"]
+        )
+        generator.pipeline = Mock(
+            return_value=[
+                {"generated_text": "Paris is the capital. STOP"},
+                {"generated_text": "France is in Europe. STOP"},
+            ]
+        )
+        generator.stopping_criteria_list = Mock()
+        results = generator.run(prompt="irrelevant")
+        # Should return exactly 2 replies, both stop words stripped
+        assert results == {"replies": ["Paris is the capital.", "France is in Europe."]}
+
     @pytest.mark.integration
     def test_stop_words_criteria_using_hf_tokenizer(self):
         """