fix: minor embed compat fixes

tests/scripts/test_reembed.py

@@ -99,6 +99,14 @@ class TestFaissRebuild:
         assert [d.page_content for d in docs] == ["alpha", "beta"]
         assert [d.metadata for d in docs] == [{"i": 0}, {"i": 1}]
 
+    def test_embeddings_key_comes_from_settings(self, stores):
+        """A placeholder here is sent as the server's bearer token."""
+        _, _, factory = stores
+        with patch.object(reembed.settings, "EMBEDDINGS_KEY", "sk-real", create=True):
+            reembed.reembed_faiss("s1", batch_size=8, dry_run=False)
+        keys = [call.kwargs.get("embeddings_key") for call in factory.call_args_list]
+        assert keys == ["sk-real", "sk-real"]
+
     def test_existing_index_is_not_deleted(self, stores):
         """The rebuild must not destroy the old index before the new one exists."""
         existing, rebuilt, _ = stores

tests/vectorstore/test_embeddings_local.py

@@ -165,3 +165,54 @@ class TestLengthSortedBatching:
             wrapper, _ = self._wrapper(fake_fastembed, 2)
             out = wrapper.embed_documents(["aa", "b", "aa", "ccc"])
         assert out == [[2.0], [1.0], [2.0], [3.0]]
+
+
+class TestTokenizerPadding:
+    """A fixed padding width in ``tokenizer.json`` makes mixed batches ragged.
+
+    FastEmbed calls ``enable_padding`` only when the tokenizer declares none,
+    so mpnet's fixed ``length: 128`` survives loading. Any batch mixing an
+    input longer than 128 tokens with a shorter one then produces rows of
+    different widths and ONNX rejects the tensor.
+    """
+
+    def _tokenizer(self, padding):
+        tokenizer = MagicMock()
+        tokenizer.padding = padding
+        return tokenizer
+
+    def test_fixed_width_padding_is_reset_to_batch_longest(self, fake_fastembed):
+        _, instance = fake_fastembed
+        tokenizer = self._tokenizer(
+            {
+                "length": 128,
+                "pad_id": 1,
+                "pad_token": "<pad>",
+                "pad_type_id": 0,
+                "direction": "right",
+                "pad_to_multiple_of": None,
+            }
+        )
+        instance.model.tokenizer = tokenizer
+
+        EmbeddingsWrapper(MPNET.name)
+
+        kwargs = tokenizer.enable_padding.call_args.kwargs
+        assert kwargs["length"] is None, "padding must follow the longest input"
+        # The model's own pad token must survive the reset.
+        assert kwargs["pad_id"] == 1
+        assert kwargs["pad_token"] == "<pad>"
+
+    def test_dynamic_padding_is_left_alone(self, fake_fastembed):
+        _, instance = fake_fastembed
+        tokenizer = self._tokenizer({"length": None, "pad_id": 0, "pad_token": "<pad>"})
+        instance.model.tokenizer = tokenizer
+
+        EmbeddingsWrapper(GRANITE_97M.name)
+
+        tokenizer.enable_padding.assert_not_called()
+
+    def test_tokenizer_that_cannot_be_reached_is_not_fatal(self, fake_fastembed):
+        _, instance = fake_fastembed
+        instance.model = None
+        EmbeddingsWrapper(GRANITE_97M.name)

tests/vectorstore/test_remote_embeddings_truncation.py

@@ -105,10 +105,31 @@ class TestInputLimitResolution:
         from application.vectorstore import base
 
         monkeypatch.setattr(base.settings, "EMBEDDINGS_MAX_INPUT_TOKENS", None)
+        monkeypatch.setattr(base, "_embeddings_name_is_explicit", lambda: True)
         assert self._remote("granite-311m")._resolve_input_limit() == 32768
         # mpnet genuinely stops at 384; sending more is paid for and discarded.
         assert self._remote("all-mpnet-base-v2")._resolve_input_limit() == 384
 
+    def test_default_model_name_lends_the_server_no_ceiling(self, monkeypatch):
+        """An unset EMBEDDINGS_NAME describes nothing about the remote server.
+
+        The name is only forwarded as the ``model`` field. Letting the
+        settings default contribute mpnet's 384-token window would clip every
+        chunk on a server that may well serve a 32k-context model.
+        """
+        from application.vectorstore import base
+
+        monkeypatch.setattr(base.settings, "EMBEDDINGS_MAX_INPUT_TOKENS", None)
+        monkeypatch.setattr(base, "_embeddings_name_is_explicit", lambda: False)
+        assert self._remote("all-mpnet-base-v2")._resolve_input_limit() is None
+
+    def test_explicit_setting_still_wins_over_an_unset_name(self, monkeypatch):
+        from application.vectorstore import base
+
+        monkeypatch.setattr(base.settings, "EMBEDDINGS_MAX_INPUT_TOKENS", 512)
+        monkeypatch.setattr(base, "_embeddings_name_is_explicit", lambda: False)
+        assert self._remote("all-mpnet-base-v2")._resolve_input_limit() == 512
+
     def test_unknown_model_stays_unlimited(self, monkeypatch):
         from application.vectorstore import base
 
@@ -119,6 +140,7 @@ class TestInputLimitResolution:
         from application.vectorstore import base
 
         monkeypatch.setattr(base.settings, "EMBEDDINGS_MAX_INPUT_TOKENS", 0)
+        monkeypatch.setattr(base, "_embeddings_name_is_explicit", lambda: True)
         assert self._remote("granite-97m")._resolve_input_limit() == 32768
 
     def test_dimension_is_taken_from_the_registry(self):