allow base_url to work with proxies, e.g. http://host/proxy/path/v1

libs/ai-endpoints/langchain_nvidia_ai_endpoints/_common.py

@@ -126,6 +126,17 @@ class _NVIDIAClient(BaseModel):
 
     @field_validator("base_url")
     def _validate_base_url(cls, v: str) -> str:
+        """
+        validate the base_url.
+
+        if the base_url is not a url, raise an error
+
+        if the base_url does not end in /v1, e.g. /embeddings, /completions, /rankings,
+        or /reranking, emit a warning. old documentation told users to pass in the full
+        inference url, which is incorrect and prevents model listing from working.
+
+        normalize base_url to end in /v1
+        """
         ## Making sure /v1 in added to the url
         if v is not None:
             parsed = urlparse(v)
@@ -135,12 +146,17 @@ def _validate_base_url(cls, v: str) -> str:
                 expected_format = "Expected format is: http://host:port"
                 raise ValueError(f"Invalid base_url format. {expected_format} Got: {v}")
 
-            if v.strip("/").endswith(
-                ("/embeddings", "/completions", "/rankings", "/reranking")
-            ):
-                warnings.warn(f"Using {v}, ignoring the rest")
+            normalized_path = parsed.path.rstrip("/")
+            if not normalized_path.endswith("/v1"):
+                warnings.warn(
+                    f"{v} does not end in /v1, you may "
+                    "have inference and listing issues"
+                )
+                normalized_path += "/v1"
 
-            v = urlunparse((parsed.scheme, parsed.netloc, "v1", None, None, None))
+            v = urlunparse(
+                (parsed.scheme, parsed.netloc, normalized_path, None, None, None)
+            )
 
         return v
 

libs/ai-endpoints/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "langchain-nvidia-ai-endpoints"
-version = "0.3.1"
+version = "0.3.2"
 description = "An integration package connecting NVIDIA AI Endpoints and LangChain"
 authors = []
 readme = "README.md"

libs/ai-endpoints/tests/unit_tests/test_base_url.py

@@ -1,5 +1,6 @@
 import os
 import re
+import warnings
 from typing import Any
 
 import pytest
@@ -98,6 +99,7 @@ def test_param_base_url_hosted(public_class: type, base_url: str) -> None:
     ],
 )
 def test_param_base_url_not_hosted(public_class: type, base_url: str) -> None:
+    warnings.filterwarnings("ignore", r".*does not end in /v1.*")
     with no_env_var("NVIDIA_BASE_URL"):
         client = public_class(model="model1", base_url=base_url)
         assert not client._client.is_hosted
@@ -119,9 +121,42 @@ def test_expect_warn(public_class: type, base_url: str) -> None:
     with pytest.warns(UserWarning) as record:
         public_class(model="model1", base_url=base_url)
     assert len(record) == 1
-    assert "ignoring the rest" in str(record[0].message)
+    assert "does not end in /v1" in str(record[0].message)
 
 
 def test_default_hosted(public_class: type) -> None:
     x = public_class(api_key="BOGUS")
     assert x._client.is_hosted
+
+
+@pytest.mark.parametrize(
+    "base_url",
+    [
+        "http://host/path0/path1/path2/v1",
+        "http://host:123/path0/path1/path2/v1/",
+    ],
+)
+def test_proxy_base_url(
+    public_class: type, base_url: str, requests_mock: Mocker
+) -> None:
+    with no_env_var("NVIDIA_BASE_URL"):
+        client = public_class(model="model1", base_url=base_url)
+        assert base_url.startswith(client.base_url)
+
+
+@pytest.mark.parametrize(
+    "base_url",
+    [
+        "http://host/path0/path1/path2/v1",
+        "http://host:123/path0/path1/path2/v1/",
+    ],
+)
+def test_proxy_base_url_models(
+    public_class: type, base_url: str, requests_mock: Mocker
+) -> None:
+    with no_env_var("NVIDIA_BASE_URL"):
+        client = public_class(model="model1", base_url=base_url)
+        client.available_models
+        models_url = base_url.rstrip("/") + "/models"
+        assert requests_mock.last_request
+        assert requests_mock.last_request.url == models_url