test(inference): tolerate model non-determinism in bare-string embed test

test_embed_bare_string_auto_wrapped_{rest,async} called embed() twice
with the same input (once as a bare string, once wrapped in a list)
and asserted bit-exact equality of the returned values. The
multilingual-e5-large model isn't bit-deterministic across replicas —
two consecutive calls regularly drift in the 3rd–4th decimal. That
flake just took down release-readiness on run 25746418051.

The SDK contract being verified here is that the bare-string path
doesn't transform the input differently from the single-element-list
path, not that the model itself is deterministic. Switch the value
comparison to pytest.approx(abs=1e-3), matching the existing
convention from test_query_namespaces_tie_breaking_rest. Length
check is retained so a shape-changing regression still fails.

Author: jhamon

tests/integration/test_inference.py

@@ -234,8 +234,13 @@ def test_embed_bare_string_auto_wrapped_rest(client: Pinecone) -> None:
         parameters={"input_type": "passage", "truncate": "END"},
     )
     assert len(result_as_list) == 1
-    # Same string → same embedding values (deterministic model)
-    assert result.data[0].values == result_as_list.data[0].values
+    # Same string → same embedding shape and (approximately) same values.
+    # Element-wise tolerance absorbs server-side non-determinism (e.g. mixed
+    # precision across replicas); the SDK contract being verified here is
+    # that the bare-string path doesn't transform the input differently from
+    # the single-element-list path, not that the model is bit-deterministic.
+    assert len(result.data[0].values) == len(result_as_list.data[0].values)
+    assert result.data[0].values == pytest.approx(result_as_list.data[0].values, abs=1e-3)
 
 
 # ---------------------------------------------------------------------------

tests/integration/test_inference_async.py

@@ -190,8 +190,13 @@ async def test_embed_bare_string_auto_wrapped_async(async_client: AsyncPinecone)
         parameters={"input_type": "passage", "truncate": "END"},
     )
     assert len(result_as_list) == 1
-    # Same string → same embedding values (deterministic model)
-    assert result.data[0].values == result_as_list.data[0].values
+    # Same string → same embedding shape and (approximately) same values.
+    # Element-wise tolerance absorbs server-side non-determinism (e.g. mixed
+    # precision across replicas); the SDK contract being verified here is
+    # that the bare-string path doesn't transform the input differently from
+    # the single-element-list path, not that the model is bit-deterministic.
+    assert len(result.data[0].values) == len(result_as_list.data[0].values)
+    assert result.data[0].values == pytest.approx(result_as_list.data[0].values, abs=1e-3)
 
 
 # ---------------------------------------------------------------------------