diff --git a/flow/flow/doctype/flow_model/flow_model.py b/flow/flow/doctype/flow_model/flow_model.py
index fb9e37e..ac8efcd 100644
--- a/flow/flow/doctype/flow_model/flow_model.py
+++ b/flow/flow/doctype/flow_model/flow_model.py
@@ -139,21 +139,41 @@ def test_connection(self):
 		kwargs = {
 			"model": self.model_id,
 			"api_key": api_key,
-			"messages": [{"role": "user", "content": "ping"}],
-			"max_tokens": 1,
 			"timeout": 15,
 		}
 		if base_url:
 			kwargs["api_base"] = base_url
 
 		try:
-			litellm.completion(**kwargs)
+			if _is_embedding_model(self.model_id):
+				litellm.embedding(input=["ping"], encoding_format="float", **kwargs)
+			else:
+				litellm.completion(
+					messages=[{"role": "user", "content": "ping"}],
+					max_tokens=1,
+					**kwargs,
+				)
 		except Exception as e:
 			frappe.throw(str(e)[:500] or type(e).__name__, title=_(type(e).__name__))
 
 		return {"ok": True, "message": _("Connection OK")}
 
 
+def _is_embedding_model(model_id: str) -> bool:
+	"""Whether ``model_id`` should be tested through the embeddings endpoint."""
+	import litellm
+
+	try:
+		if litellm.get_model_info(model_id).get("mode") == "embedding":
+			return True
+	except Exception:
+		# Custom OpenAI-compatible models are often absent from LiteLLM's registry.
+		pass
+
+	model_name = model_id.rsplit("/", 1)[-1].lower()
+	return "embedding" in model_name
+
+
 def _detect_context_window(model_id: str) -> int:
 	"""Max input tokens for `model_id` per litellm, or 0 if unknown/unmapped."""
 	import litellm
diff --git a/flow/flow/doctype/flow_model/test_flow_model.py b/flow/flow/doctype/flow_model/test_flow_model.py
index c40ae06..05cc5d9 100644
--- a/flow/flow/doctype/flow_model/test_flow_model.py
+++ b/flow/flow/doctype/flow_model/test_flow_model.py
@@ -7,7 +7,7 @@
 import frappe
 from frappe.tests import IntegrationTestCase
 
-from flow.flow.doctype.flow_model.flow_model import _detect_context_window
+from flow.flow.doctype.flow_model.flow_model import FlowModel, _detect_context_window, _is_embedding_model
 
 
 def _model(**overrides: Any) -> dict:
@@ -180,3 +180,64 @@ def test_keeps_existing_when_new_model_unmapped(self):
 			doc.model_id = "anthropic/made-up-model-xyz"
 			doc.save()
 		self.assertEqual(doc.context_window, 128000)
+
+
+class TestFlowModelConnection(IntegrationTestCase):
+	def tearDown(self):
+		frappe.db.rollback()
+
+	def test_detects_mapped_embedding_model(self):
+		with patch("litellm.get_model_info", return_value={"mode": "embedding"}):
+			self.assertTrue(_is_embedding_model("vendor/vector-model"))
+
+	def test_detects_unmapped_embedding_model_by_name(self):
+		with patch("litellm.get_model_info", side_effect=Exception("not mapped")):
+			self.assertTrue(_is_embedding_model("openai/text-embedding-v4"))
+
+	def test_connection_uses_embedding_endpoint_for_unmapped_embedding_model(self):
+		doc = frappe.get_doc(
+			_model(model_id="openai/text-embedding-v4", base_url="https://api.example.com/v1")
+		)
+
+		with (
+			patch.object(FlowModel, "check_permission"),
+			patch.object(FlowModel, "get_password", return_value="sk-test"),
+			patch("flow.lib.model.resolve_provider_credentials", return_value={}),
+			patch("litellm.get_model_info", side_effect=Exception("not mapped")),
+			patch("litellm.embedding") as embedding,
+			patch("litellm.completion") as completion,
+		):
+			result = doc.test_connection()
+
+		embedding.assert_called_once_with(
+			model="openai/text-embedding-v4",
+			api_key="sk-test",
+			timeout=15,
+			api_base="https://api.example.com/v1",
+			input=["ping"],
+			encoding_format="float",
+		)
+		completion.assert_not_called()
+		self.assertTrue(result["ok"])
+
+	def test_connection_keeps_chat_models_on_completion_endpoint(self):
+		doc = frappe.get_doc(_model(model_id="openai/gpt-4o-mini"))
+
+		with (
+			patch.object(FlowModel, "check_permission"),
+			patch.object(FlowModel, "get_password", return_value="sk-test"),
+			patch("flow.lib.model.resolve_provider_credentials", return_value={}),
+			patch("litellm.get_model_info", return_value={"mode": "chat"}),
+			patch("litellm.embedding") as embedding,
+			patch("litellm.completion") as completion,
+		):
+			doc.test_connection()
+
+		embedding.assert_not_called()
+		completion.assert_called_once_with(
+			model="openai/gpt-4o-mini",
+			api_key="sk-test",
+			timeout=15,
+			messages=[{"role": "user", "content": "ping"}],
+			max_tokens=1,
+		)