fix: only strip known provider prefixes

Author: andrewklatzke

packages/optimization/src/ldai_optimizer/client.py

@@ -91,18 +91,40 @@ def _find_model_config(
     return global_match if global_match is not None else matching[0]
 
 
+# Known provider prefixes used by the LD API (e.g. "Anthropic.claude-3").
+# Only strip the first segment when it is one of these known values so that
+# model IDs whose first dotted segment is NOT a provider — such as Bedrock
+# cross-region inference IDs like "us.amazon.nova-pro-v1:0" — are left intact.
+_KNOWN_PROVIDER_PREFIXES: frozenset = frozenset({
+    "Anthropic",
+    "Bedrock",
+    "Cohere",
+    "Google",
+    "Groq",
+    "Meta",
+    "Mistral",
+    "OpenAI",
+    "Perplexity",
+})
+
+
 def _strip_provider_prefix(model: str) -> str:
     """Strip the provider prefix from a model identifier returned by the LD API.
 
     API model keys are formatted as "Provider.model-name" (e.g. "OpenAI.gpt-5",
-    "Anthropic.claude-opus-4.6"). Only the part after the first period is needed
-    by the underlying LLM clients. If no period is present the string is returned
-    unchanged.
+    "Anthropic.claude-opus-4.6"). Only the segment before the first period is
+    stripped, and only when it is a recognised provider name. This prevents
+    incorrectly stripping region prefixes from Bedrock cross-region inference
+    IDs such as "us.amazon.nova-pro-v1:0".
 
     :param model: Raw model string from the API.
-    :return: Model name with provider prefix removed.
+    :return: Model name with provider prefix removed, or the original string if
+        the first segment is not a known provider.
     """
-    return model.split(".", 1)[-1]
+    prefix, _, rest = model.partition(".")
+    if prefix in _KNOWN_PROVIDER_PREFIXES and rest:
+        return rest
+    return model
 
 
 def _compute_validation_count(pool_size: int) -> int:

packages/optimization/tests/test_client.py

@@ -16,6 +16,7 @@
     _MAX_STANDARD_HISTORY_LENGTH,
     _compute_validation_count,
     _find_model_config,
+    _strip_provider_prefix,
     _trim_history,
 )
 from ldai_optimizer.util import judge_passed
@@ -129,6 +130,40 @@ def _make_client(ldai: MagicMock | None = None) -> OptimizationClient:
 # ---------------------------------------------------------------------------
 
 
+# ---------------------------------------------------------------------------
+# _strip_provider_prefix
+# ---------------------------------------------------------------------------
+
+
+class TestStripProviderPrefix:
+    def test_strips_known_anthropic_prefix(self):
+        assert _strip_provider_prefix("Anthropic.claude-opus-4-5") == "claude-opus-4-5"
+
+    def test_strips_known_openai_prefix(self):
+        assert _strip_provider_prefix("OpenAI.gpt-4o") == "gpt-4o"
+
+    def test_strips_known_bedrock_prefix(self):
+        # "Bedrock.us.amazon.nova-pro-v1:0" → region prefix is preserved
+        assert _strip_provider_prefix("Bedrock.us.amazon.nova-pro-v1:0") == "us.amazon.nova-pro-v1:0"
+
+    def test_does_not_strip_bedrock_region_prefix(self):
+        # Raw Bedrock cross-region ID has no provider prefix — must be unchanged
+        assert _strip_provider_prefix("us.amazon.nova-pro-v1:0") == "us.amazon.nova-pro-v1:0"
+
+    def test_does_not_strip_eu_region_prefix(self):
+        assert _strip_provider_prefix("eu.anthropic.claude-3-5-sonnet-20241022-v2:0") == "eu.anthropic.claude-3-5-sonnet-20241022-v2:0"
+
+    def test_no_period_returns_unchanged(self):
+        assert _strip_provider_prefix("gpt-4o") == "gpt-4o"
+
+    def test_empty_string_returns_unchanged(self):
+        assert _strip_provider_prefix("") == ""
+
+    def test_preserves_dots_in_model_name_after_stripping(self):
+        # Multiple dots after the provider prefix are preserved
+        assert _strip_provider_prefix("Anthropic.claude-3.5-sonnet") == "claude-3.5-sonnet"
+
+
 # ---------------------------------------------------------------------------
 # _find_model_config
 # ---------------------------------------------------------------------------