Skip to content

OpenRAG is a comprehensive, single package Retrieval-Augmented Generation platform built on Langflow, Docling, and Opensearch.

Python

fix: correct Azure AI Foundry resource health checks (#2514)

Edwin Josee02cda2801e0
diff --git a/docs/docs/reference/configuration.mdx b/docs/docs/reference/configuration.mdx
index 5b12adb69..26daa07b9 100644
--- a/docs/docs/reference/configuration.mdx
+++ b/docs/docs/reference/configuration.mdx
@@ -190,6 +190,19 @@ today — are entered in **Settings** through a form built from the provider's o
credential fields. Configuring one needs no code change beyond its row in this
file.
+#### Azure AI Foundry health checks {#azure-ai-foundry-health}
+
+For `azure_ai`, a resource endpoint such as
+`https://<resource>.services.ai.azure.com/` or its `/models` form uses the
+legacy Foundry inference route. Keep that endpoint if inference is already
+working; the periodic health check uses the resource's separate
+`/openai/v1/models` listing, not `GET /models`.
+
+Explicit `/openai/v1` endpoints, including project-scoped endpoints, keep their
+configured prefix when listing models. These lightweight checks do not run
+inference or consume model tokens, and a successful listing does not prove that
+a particular deployment supports chat or tool calling.
+
#### Add a custom OpenAI-compatible provider {#custom-openai-compatible-provider}
To offer a self-hosted OpenAI-compatible endpoint — vLLM, LM Studio, or an
diff --git a/enhancements/providers/azure/foundry.py b/enhancements/providers/azure/foundry.py
index a266b7129..7d38fb69d 100644
--- a/enhancements/providers/azure/foundry.py
+++ b/enhancements/providers/azure/foundry.py
@@ -240,14 +240,10 @@ CREDENTIAL_FIELDS: list[dict[str, Any]] = [
{
"key": "api_base",
"label": "Endpoint",
- "placeholder": "https://<resource>.services.ai.azure.com/openai/v1",
+ "placeholder": "https://<resource>.services.ai.azure.com/",
"tooltip": (
- "The resource endpoint, not a per-deployment Target URI. Copy the "
- "'Azure AI Foundry' endpoint from your resource's Overview page and keep "
- "the /openai/v1 suffix. A Target URI that names one deployment "
- "(.../openai/deployments/<name>/chat/completions) pins every model you "
- "select to that one deployment. The older /models endpoint is still "
- "accepted for deployments already using it."
+ "Copy the resource endpoint from Azure's Keys and Endpoint page. "
+ "Do not use a deployment-specific Target URI."
),
"required": True,
"field_type": "text",
@@ -269,9 +265,8 @@ CREDENTIAL_FIELDS: list[dict[str, Any]] = [
"label": "API version",
"placeholder": "2024-05-01-preview",
"tooltip": (
- "Only for the older /models endpoint, which dates its API. Leave blank "
- "for /openai/v1, which is versionless — sending one there is at best "
- "ignored."
+ "For a resource root or /models endpoint, enter the API version used by "
+ "your deployment. Leave blank for /openai/v1, which is versionless."
),
"required": False,
"field_type": "text",
@@ -514,9 +509,8 @@ def litellm_credentials(stored: Mapping[str, Any], *, kind: CallKind = "chat") -
#: Read-only listing used to check credentials. It needs no deployment and
-#: bills nothing, which is what makes it a check the provider can pass before
-#: anything has been selected in Settings. Both surfaces expose it directly
-#: under the endpoint.
+#: bills nothing. Even a resource using the legacy `/models` inference surface
+#: lists models through `/openai/v1/models`, not `GET /models`.
MODELS_PATH = "/models"
#: A human is waiting on the pre-save check, so it fails fast.
@@ -526,15 +520,15 @@ HEALTH_TIMEOUT_SECONDS = 15.0
def models_url(api_base: str) -> str:
"""The deployment/model listing URL for an endpoint.
- On the v1 surface this is `…/openai/v1/models`. On the legacy route the
- endpoint *is* `…/models`, and the listing hangs off the resource root
- instead, so the suffix is not doubled.
+ Both a resource root and its legacy `/models` inference base use the
+ resource's `/openai/v1/models` listing. Explicit v1 bases retain their
+ path, including a project prefix when present.
"""
base = normalized_api_base(api_base)
if endpoint_profile(base) == "legacy_models":
- return (
- base if base.rstrip("/").endswith(MODELS_PATH) else f"{base.rstrip('/')}{MODELS_PATH}"
- )
+ if base.lower().endswith(MODELS_PATH):
+ base = base[: -len(MODELS_PATH)]
+ base = f"{base}{_V1_SUFFIX}"
return f"{base.rstrip('/')}{MODELS_PATH}"
diff --git a/tests/unit/enhancements/providers/azure/test_foundry.py b/tests/unit/enhancements/providers/azure/test_foundry.py
index 6d0175f13..5ab2a2de3 100644
--- a/tests/unit/enhancements/providers/azure/test_foundry.py
+++ b/tests/unit/enhancements/providers/azure/test_foundry.py
@@ -374,18 +374,6 @@ class TestCredentialForm:
assert "api_key" in secret_field_keys("azure_ai")
- def test_the_form_replaces_litellms_deployment_pinning_placeholder(self) -> None:
- """LiteLLM's own `api_base` placeholder is a per-deployment Target URI.
-
- Following it produces a provider that serves exactly one deployment:
- every model selected resolves to the one baked into the URL.
- """
- from services.model_catalog import credential_fields
-
- api_base = next(f for f in credential_fields("azure_ai") if f["key"] == "api_base")
- assert "/openai/v1" in (api_base["placeholder"] or "")
- assert "deployments/" not in (api_base["placeholder"] or "")
-
def test_the_form_asks_only_for_what_this_release_consumes(self) -> None:
"""No field is offered that nothing reads.
@@ -401,19 +389,6 @@ class TestCredentialForm:
class TestModelsUrl:
- @pytest.mark.parametrize(
- "api_base, expected",
- [
- (f"{RESOURCE}/openai/v1", f"{RESOURCE}/openai/v1/models"),
- (f"{RESOURCE}/api/projects/p/openai/v1", f"{RESOURCE}/api/projects/p/openai/v1/models"),
- # The legacy endpoint already *is* the listing path.
- (f"{RESOURCE}/models", f"{RESOURCE}/models"),
- (RESOURCE, f"{RESOURCE}/models"),
- ],
- )
- def test_the_listing_path_is_not_doubled(self, api_base, expected) -> None:
- assert foundry.models_url(api_base) == expected
-
def test_both_auth_header_styles_are_sent(self) -> None:
"""litellm itself disagrees with itself here.
@@ -432,6 +407,55 @@ class TestModelsUrl:
class TestLightweightHealthCheck:
"""The pre-save probe: authenticated, read-only, and never billed."""
+ @pytest.mark.asyncio
+ @pytest.mark.parametrize(
+ "api_base, listing_path",
+ [
+ (RESOURCE, "/openai/v1/models"),
+ (f"{RESOURCE}/models/", "/openai/v1/models"),
+ (f"{RESOURCE}/openai/v1", "/openai/v1/models"),
+ (
+ f"{RESOURCE}/api/projects/p/openai/v1",
+ "/api/projects/p/openai/v1/models",
+ ),
+ ],
+ )
+ async def test_polling_uses_the_listing_api_not_the_inference_base(
+ self, monkeypatch, api_base, listing_path
+ ) -> None:
+ import httpx
+
+ from api.provider_validation import validate_provider_setup
+
+ def serve(request: httpx.Request) -> httpx.Response:
+ if (
+ request.method == "GET"
+ and request.url.path == listing_path
+ and request.headers.get("api-key") == "k"
+ and not request.url.query
+ ):
+ return httpx.Response(200, json={"object": "list", "data": []})
+ return httpx.Response(
+ 404, json={"error": {"code": "404", "message": "Resource not found"}}
+ )
+
+ client_type = httpx.AsyncClient
+ monkeypatch.setattr(
+ httpx,
+ "AsyncClient",
+ lambda **kwargs: client_type(transport=httpx.MockTransport(serve), **kwargs),
+ )
+ credentials = {"api_base": api_base, "api_key": "k", "api_version": "2024-05-01-preview"}
+
+ result = await validate_provider_setup(
+ provider="azure_ai",
+ llm_model="working-deployment",
+ credentials=foundry.litellm_credentials(credentials),
+ stored_credentials=credentials,
+ )
+
+ assert not result.model_probed
+
@pytest.mark.asyncio
async def test_a_good_endpoint_and_key_pass_without_an_inference_call(
self, monkeypatch