nvidia-haystack 1.1.1__tar.gz → 1.3.0__tar.gz

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/CHANGELOG.md

@@ -1,5 +1,24 @@
 # Changelog
 
+## [integrations/nvidia-v1.2.0] - 2026-06-05
+
+### 🧹 Chores
+
+- NVIDIA - deprecate generator (#3391)
+
+
+## [integrations/nvidia-v1.1.1] - 2026-04-14
+
+### 🐛 Bug Fixes
+
+- Replace in-place dataclass mutations with dataclasses.replace() (#3112)
+
+### 🧪 Testing
+
+- Track test coverage for all integrations (#3065)
+- Add unit tests for Nvidia integration (#3162)
+
+
 ## [integrations/nvidia-v1.1.0] - 2026-03-30
 
 ### 🐛 Bug Fixes

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nvidia-haystack
-Version: 1.1.1
+Version: 1.3.0
 Project-URL: Documentation, https://github.com/deepset-ai/haystack-core-integrations/tree/main/integrations/nvidia#readme
 Project-URL: Issues, https://github.com/deepset-ai/haystack-core-integrations/issues
 Project-URL: Source, https://github.com/deepset-ai/haystack-core-integrations/tree/main/integrations/nvidia

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/pyproject.toml

@@ -133,6 +133,7 @@ ignore = [
   "PLR0912",
   "PLR0913",
   "PLR0915",
+  "PLR0917",
   # Misc
   "B008",
   "S101",

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/src/haystack_integrations/components/generators/nvidia/generator.py

@@ -75,6 +75,12 @@ class NvidiaGenerator:
             Timeout for request calls, if not set it is inferred from the `NVIDIA_TIMEOUT` environment variable
             or set to 60 by default.
         """
+        warnings.warn(
+            "The `NvidiaGenerator` component is deprecated and will be removed in a future version. "
+            "Use `NvidiaChatGenerator` instead, which now also supports string inputs.",
+            FutureWarning,
+            stacklevel=2,
+        )
         self._model = model
         self.api_url = url_validation(api_url)
         self._api_key = api_key

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/src/haystack_integrations/utils/nvidia/nim_backend.py

@@ -42,6 +42,11 @@ class NimBackend:
         self.session.headers.update(headers)
 
         self.api_url = api_url
+        # Hosted ranking models are assigned a per-model endpoint that already includes the
+        # full path (e.g. ".../nv-rerankqa-mistral-4b-v3/reranking"), unlike other hosted
+        # endpoints which only override the host. `rank()` uses this to decide whether it
+        # still needs to append the `/ranking` path itself.
+        self._has_custom_ranking_endpoint = False
         if isinstance(client, str):
             client = Client.from_str(client)
         validated_model: Model | None = None
@@ -64,6 +69,8 @@ class NimBackend:
                 # we override the endpoint to use the custom endpoint
                 self.api_url = validated_model.endpoint
                 self.model_type = validated_model.model_type
+                if validated_model.model_type == "ranking":
+                    self._has_custom_ranking_endpoint = True
 
         self.model = validated_model.id if validated_model else model
         self.model_kwargs = model_kwargs or {}
@@ -89,8 +96,13 @@ class NimBackend:
             )
             res.raise_for_status()
         except requests.HTTPError as e:
-            logger.error("Error when calling NIM embedding endpoint: Error - {error}", error=e.response.text)
-            msg = f"Failed to query embedding endpoint: Error - {e.response.text}"
+            response_text = e.response.text.strip() if e.response is not None else ""
+            error_details = f": {response_text}" if response_text else ""
+            logger.error(
+                "Error when calling NIM embedding endpoint{error}",
+                error=error_details,
+            )
+            msg = f"Failed to query embedding endpoint{error_details}"
             raise ValueError(msg) from e
 
         data = res.json()
@@ -123,8 +135,13 @@ class NimBackend:
             )
             res.raise_for_status()
         except requests.HTTPError as e:
-            logger.error("Error when calling NIM chat completion endpoint: Error - {error}", error=e.response.text)
-            msg = f"Failed to query chat completion endpoint: Error - {e.response.text}"
+            response_text = e.response.text.strip() if e.response is not None else ""
+            error_details = f": {response_text}" if response_text else ""
+            logger.error(
+                "Error when calling NIM chat completion endpoint{error}",
+                error=error_details,
+            )
+            msg = f"Failed to query chat completion endpoint{error_details}"
             raise ValueError(msg) from e
 
         completions = res.json()
@@ -179,7 +196,10 @@ class NimBackend:
 
     def rank(self, query_text: str, document_texts: list[str]) -> list[dict[str, Any]]:
         """Rank documents by relevance to a query via the NIM API."""
-        url = self.api_url
+        # Hosted ranking models already carry the full path in their custom endpoint;
+        # everything else (local NIM containers, non-ranking-table models) needs the
+        # `/ranking` suffix appended explicitly, like the other endpoints below do.
+        url = self.api_url if self._has_custom_ranking_endpoint else f"{self.api_url}/ranking"
 
         try:
             res = self.session.post(
@@ -194,8 +214,13 @@ class NimBackend:
             )
             res.raise_for_status()
         except requests.HTTPError as e:
-            logger.error("Error when calling NIM ranking endpoint: Error - {error}", error=e.response.text)
-            msg = f"Failed to rank endpoint: Error - {e.response.text}"
+            response_text = e.response.text.strip() if e.response is not None else ""
+            error_details = f": {response_text}" if response_text else ""
+            logger.error(
+                "Error when calling NIM ranking endpoint{error}",
+                error=error_details,
+            )
+            msg = f"Failed to query ranking endpoint{error_details}"
             raise ValueError(msg) from e
 
         data = res.json()

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/tests/conftest.py

@@ -47,3 +47,13 @@ def mock_local_models(requests_mock: Mocker) -> None:
             ]
         },
     )
+
+
+@pytest.fixture(autouse=True)
+def allow_deserialization_of_test_modules(monkeypatch):
+    """
+    haystack-ai >= 3.0 refuses to deserialize classes and callables from modules outside its
+    trusted-module allowlist. Tools and callbacks defined in the test modules live outside that
+    allowlist, so trust them explicitly; haystack-ai 2.x ignores this environment variable.
+    """
+    monkeypatch.setenv("HAYSTACK_DESERIALIZATION_ALLOWLIST", "tests,test_*")

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/tests/test_base_url.py

@@ -28,7 +28,8 @@ from haystack_integrations.components.rankers.nvidia import NvidiaRanker
 def test_base_url_invalid_not_hosted(base_url: str, component) -> None:
     with pytest.warns(UserWarning) as msg:
         component(api_url=base_url, model="x")
-    assert "you may have inference and listing issues" in str(msg[0].message)
+    user_warnings = [w for w in msg if issubclass(w.category, UserWarning)]
+    assert any("you may have inference and listing issues" in str(w.message) for w in user_warnings)
 
 
 @pytest.mark.parametrize(

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/tests/test_generator.py

@@ -193,7 +193,7 @@ class TestNvidiaGenerator:
     @pytest.mark.integration
     def test_run_integration_with_api_catalog(self):
         generator = NvidiaGenerator(
-            model="meta/llama3-8b-instruct",
+            model="meta/llama-3.1-8b-instruct",
             api_url="https://integrate.api.nvidia.com/v1",
             api_key=Secret.from_env_var("NVIDIA_API_KEY"),
             model_arguments={
@@ -256,7 +256,7 @@ class TestNvidiaGenerator:
 
     def test_warm_up_is_idempotent(self, monkeypatch):
         monkeypatch.setenv("NVIDIA_API_KEY", "fake-api-key")
-        generator = NvidiaGenerator("meta/llama3-8b-instruct")
+        generator = NvidiaGenerator("meta/llama-3.1-8b-instruct")
         generator.warm_up()
         backend = generator.backend
         generator.warm_up()
@@ -264,7 +264,7 @@ class TestNvidiaGenerator:
 
     def test_available_models_without_backend(self, monkeypatch):
         monkeypatch.setenv("NVIDIA_API_KEY", "fake-api-key")
-        generator = NvidiaGenerator("meta/llama3-8b-instruct")
+        generator = NvidiaGenerator("meta/llama-3.1-8b-instruct")
         assert generator.available_models == []
 
     @pytest.mark.usefixtures("mock_local_models")
@@ -283,12 +283,12 @@ class TestNvidiaGenerator:
             "init_parameters": {
                 "api_key": {"env_vars": ["NVIDIA_API_KEY"], "strict": True, "type": "env_var"},
                 "api_url": "https://my.url.com/v1",
-                "model": "meta/llama3-8b-instruct",
+                "model": "meta/llama-3.1-8b-instruct",
                 "model_arguments": {"temperature": 0.5},
             },
         }
         generator = NvidiaGenerator.from_dict(data)
-        assert generator._model == "meta/llama3-8b-instruct"
+        assert generator._model == "meta/llama-3.1-8b-instruct"
         assert generator.api_url == "https://my.url.com/v1"
         assert generator._model_arguments == {"temperature": 0.5}
 

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/tests/test_nim_backend.py

@@ -289,6 +289,28 @@ class TestNimBackend:
                 timeout=60.0,
             )
 
+    def test_rank_local_nim_appends_ranking_path(self, monkeypatch):
+        # Local/self-hosted NIM deployments have no model-specific endpoint override,
+        # so `rank()` must append the `/ranking` path itself instead of relying on
+        # `validate_hosted_model` having already baked it into `self.api_url`.
+        with patch("requests.sessions.Session.post", side_effect=mock_rank_post_response) as mock_post:
+            monkeypatch.setenv("NVIDIA_API_KEY", "fake-api-key")
+            backend = NimBackend(model="custom-rerank-model", api_url="http://localhost:8000/v1", client="NvidiaRanker")
+            query_text = "query"
+            document_texts = ["text1", "text2"]
+            backend.rank(query_text=query_text, document_texts=document_texts)
+
+            expected_url = "http://localhost:8000/v1/ranking"
+            mock_post.assert_called_once_with(
+                expected_url,
+                json={
+                    "model": "custom-rerank-model",
+                    "query": {"text": query_text},
+                    "passages": [{"text": text} for text in document_texts],
+                },
+                timeout=60.0,
+            )
+
     def test_rank_raises_on_http_error(self, monkeypatch):
         error_response = requests.Response()
         error_response.status_code = 500
@@ -298,7 +320,7 @@ class TestNimBackend:
             backend = NimBackend(
                 model="nvidia/llama-3.2-nv-rerankqa-1b-v2", api_url=DEFAULT_API_URL, client="NvidiaRanker"
             )
-            with pytest.raises(ValueError, match="Failed to rank endpoint"):
+            with pytest.raises(ValueError, match="Failed to query ranking endpoint"):
                 backend.rank(query_text="q", document_texts=["a"])
 
     def test_rank_raises_when_rankings_missing(self, monkeypatch):

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/tests/test_nvidia_chat_generator.py

@@ -119,15 +119,23 @@ class TestNvidiaChatGenerator:
     def test_init_default(self, monkeypatch):
         monkeypatch.setenv("NVIDIA_API_KEY", "test-api-key")
         component = NvidiaChatGenerator()
-        assert component.client.api_key == "test-api-key"
+        assert component.api_key.resolve_value() == "test-api-key"
         assert component.model == "meta/llama-3.1-8b-instruct"
         assert component.streaming_callback is None
         assert not component.generation_kwargs
 
+    def test_warm_up(self, monkeypatch):
+        monkeypatch.setenv("NVIDIA_API_KEY", "test-api-key")
+        component = NvidiaChatGenerator()
+        component.warm_up()  # with haystack-ai >= 3.0 the client is created during warm-up
+        assert component.client.api_key == "test-api-key"
+
     def test_init_fail_wo_api_key(self, monkeypatch):
         monkeypatch.delenv("NVIDIA_API_KEY", raising=False)
         with pytest.raises(ValueError, match=r"None of the .* environment variables are set"):
-            NvidiaChatGenerator()
+            # haystack-ai 2.x raises at init; haystack-ai >= 3.0 raises when the client is created in warm_up
+            component = NvidiaChatGenerator()
+            component.warm_up()
 
     def test_init_with_parameters(self):
         component = NvidiaChatGenerator(
@@ -137,7 +145,7 @@ class TestNvidiaChatGenerator:
             api_base_url="test-base-url",
             generation_kwargs={"max_tokens": 10, "some_test_param": "test-params"},
         )
-        assert component.client.api_key == "test-api-key"
+        assert component.api_key.resolve_value() == "test-api-key"
         assert component.model == "meta/llama-3.1-8b-instruct"
         assert component.streaming_callback is print_streaming_chunk
         assert component.generation_kwargs == {"max_tokens": 10, "some_test_param": "test-params"}
@@ -463,13 +471,16 @@ class TestNvidiaChatGenerator:
 
 
 class TestNvidiaChatGeneratorAsync:
-    def test_init_default_async(self, monkeypatch):
+    @pytest.mark.asyncio
+    async def test_warm_up_async(self, monkeypatch):
         monkeypatch.setenv("NVIDIA_API_KEY", "test-api-key")
         component = NvidiaChatGenerator()
+        if hasattr(component, "warm_up_async"):
+            # haystack-ai >= 3.0 creates the async client during async warm-up
+            await component.warm_up_async()
 
         assert isinstance(component.async_client, AsyncOpenAI)
         assert component.async_client.api_key == "test-api-key"
-        assert not component.generation_kwargs
 
     @pytest.mark.asyncio
     async def test_run_async(self, chat_messages, mock_async_chat_completion, monkeypatch):  # noqa: ARG002

{nvidia_haystack-1.1.1 → nvidia_haystack-1.3.0}/tests/test_ranker.py

@@ -120,6 +120,36 @@ class TestNvidiaRanker:
         assert response[0].content == documents[1].content
         assert response[0].score == 4.2
 
+    @pytest.mark.parametrize(
+        ("model", "api_url", "expected_url"),
+        [
+            (
+                _DEFAULT_MODEL,
+                None,
+                "https://ai.api.nvidia.com/v1/retrieval/nvidia/nv-rerankqa-mistral-4b-v3/reranking",
+            ),
+            ("custom-rerank-model", "http://localhost:8000/v1", "http://localhost:8000/v1/ranking"),
+        ],
+        ids=["hosted", "local-nim"],
+    )
+    def test_ranking_url(self, requests_mock, monkeypatch, model, api_url, expected_url) -> None:
+        # Hosted models are assigned a per-model endpoint that already includes the full path,
+        # while local/self-hosted NIM deployments get no such override and need `/ranking`
+        # appended to the base URL. Registering only the exact expected URL means any other
+        # URL raises NoMockAddress instead of silently passing.
+        monkeypatch.setenv("NVIDIA_API_KEY", "fake-api-key")
+        requests_mock.post(expected_url, json={"rankings": [{"index": 0, "logit": 1.0}]}, complete_qs=True)
+
+        api_url_param = {} if api_url is None else {"api_url": api_url}
+        client = NvidiaRanker(model=model, top_k=1, **api_url_param)
+        client.warm_up()
+
+        response = client.run(query="q", documents=[Document(content="doc")])["documents"]
+
+        assert len(response) == 1
+        assert requests_mock.last_request is not None
+        assert requests_mock.last_request.url == expected_url
+
     @pytest.mark.parametrize("truncate", [True, False, 1, 0, 1.0, "START", "BOGUS"])
     def test_truncate_invalid(self, truncate: Any) -> None:
         with pytest.raises(ValueError) as e:
@@ -181,6 +211,7 @@ class TestNvidiaRanker:
         client = NvidiaRanker(
             model=os.environ["NVIDIA_NIM_RANKER_MODEL"],
             api_url=os.environ["NVIDIA_NIM_RANKER_ENDPOINT_URL"],
+            api_key=None,
             top_k=2,
         )
         client.warm_up()