Fix: Preserve local and custom model names in parse_model_name

deepeval/models/base_model.py

@@ -1,6 +1,5 @@
 from abc import ABC, abstractmethod
 from typing import Any, Optional, List
-from deepeval.models.utils import parse_model_name
 
 
 class DeepEvalBaseModel(ABC):
@@ -32,7 +31,7 @@ def _call(self, *args, **kwargs):
 
 class DeepEvalBaseLLM(ABC):
     def __init__(self, model_name: Optional[str] = None, *args, **kwargs):
-        self.model_name = parse_model_name(model_name)
+        self.model_name = model_name
         self.model = self.load_model(*args, **kwargs)
 
     @abstractmethod
@@ -77,7 +76,7 @@ def get_model_name(self, *args, **kwargs) -> str:
 
 class DeepEvalBaseMLLM(ABC):
     def __init__(self, model_name: Optional[str] = None, *args, **kwargs):
-        self.model_name = parse_model_name(model_name)
+        self.model_name = model_name
 
     @abstractmethod
     def generate(self, *args, **kwargs) -> str:
@@ -104,7 +103,7 @@ def get_model_name(self, *args, **kwargs) -> str:
 
 class DeepEvalBaseEmbeddingModel(ABC):
     def __init__(self, model_name: Optional[str] = None, *args, **kwargs):
-        self.model_name = parse_model_name(model_name)
+        self.model_name = model_name
 
         self.model = self.load_model(*args, **kwargs)
 

deepeval/models/llms/anthropic_model.py

@@ -5,7 +5,6 @@
 
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.models.llms.utils import trim_and_load_json
-from deepeval.models.utils import parse_model_name
 
 model_pricing = {
     "claude-3-7-sonnet-latest": {"input": 3.00 / 1e6, "output": 15.00 / 1e6},
@@ -25,7 +24,7 @@ def __init__(
         temperature: float = 0,
         _anthropic_api_key: Optional[str] = None,
     ):
-        model_name = parse_model_name(model)
+        model_name = model
         self._anthropic_api_key = _anthropic_api_key
 
         if temperature < 0:

deepeval/models/llms/azure_model.py

@@ -14,7 +14,6 @@
     log_retry_error,
 )
 from deepeval.models.llms.utils import trim_and_load_json
-from deepeval.models.utils import parse_model_name
 
 retryable_exceptions = (
     openai.RateLimitError,
@@ -61,7 +60,7 @@ def __init__(
         # args and kwargs will be passed to the underlying model, in load_model function
         self.args = args
         self.kwargs = kwargs
-        super().__init__(parse_model_name(model_name))
+        super().__init__(model_name)
 
     ###############################################
     # Other generate functions

deepeval/models/llms/openai_model.py

@@ -14,7 +14,6 @@
 
 from deepeval.models import DeepEvalBaseLLM
 from deepeval.models.llms.utils import trim_and_load_json
-from deepeval.models.utils import parse_model_name
 
 
 def log_retry_error(retry_state: RetryCallState):
@@ -160,7 +159,7 @@ def __init__(
     ):
         model_name = None
         if isinstance(model, str):
-            model_name = parse_model_name(model)
+            model_name = model
             if model_name not in valid_gpt_models:
                 raise ValueError(
                     f"Invalid model. Available GPT models: {', '.join(model for model in valid_gpt_models)}"

deepeval/models/mlllms/openai_model.py

@@ -14,7 +14,6 @@
 
 from deepeval.models import DeepEvalBaseMLLM
 from deepeval.test_case import MLLMImage
-from deepeval.models.utils import parse_model_name
 
 retryable_exceptions = (
     openai.RateLimitError,
@@ -87,7 +86,7 @@ def __init__(
     ):
         model_name = None
         if isinstance(model, str):
-            model_name = parse_model_name(model)
+            model_name = model
             if model_name not in valid_multimodal_gpt_models:
                 raise ValueError(
                     f"Invalid model. Available Multimodal GPT models: {', '.join(model for model in valid_multimodal_gpt_models)}"

deepeval/models/utils.py

@@ -1,31 +1,43 @@
 from typing import Optional
 
 
-def parse_model_name(model_name: Optional[str] = None) -> str:
+def parse_model_name(model_name: Optional[str] = None) -> Optional[str]:
     """Extract base model name from provider-prefixed format.
 
-    This function is useful for extracting the actual model name from a
-    provider-prefixed format which is used by some proxies like LiteLLM.
-    LiteLLM is designed to work with many different LLM providers (OpenAI, Anthropic,
-    Cohere, etc.). To tell it which provider's API to call, you prepend the provider
-    name to the model ID, in the form "<provider>/<model>". So openai/gpt-4.1-mini
-    literally means "OpenAI's GPT-4.1 Mini via the OpenAI chat completions endpoint."
+    Some model serving systems (e.g., LiteLLM) prepend the provider to the model ID
+    in the format "<provider>/<model>". This function strips the known provider
+    prefix and returns the actual model name.
+
+    For example, "openai/gpt-4o" means GPT-4o served via OpenAI. In such cases,
+    we strip the "openai/" prefix.
+
+    However, not all prefixes are providers — for example, "local/llama-3" might refer
+    to a valid model path in a self-hosted setup. We preserve model names unless the
+    prefix matches a known provider.
 
     Args:
-        model_name: Original model identifier, potentially in
-            "<provider>/<model>" format
+        model_name: The original model identifier, potentially in "<provider>/<model>" format.
 
     Returns:
-        The model name without provider prefix
+        The model name without provider prefix for known providers.
+        Returns the full name for unknown providers (e.g., "local", "custom").
 
     Examples:
         parse_model_name("openai/gpt-4o") -> "gpt-4o"
         parse_model_name("gpt-4o") -> "gpt-4o"
+        parse_model_name("local/llama-3") -> "local/llama-3"
+        parse_model_name(None) -> None
     """
     if model_name is None:
         return None
 
+    known_providers = {"openai", "anthropic", "cohere", "mistral", "groq", "huggingface"}
+
     if "/" in model_name:
-        _, parsed_model_name = model_name.split("/", 1)
-        return parsed_model_name
+        provider, parsed_model_name = model_name.split("/", 1)
+        if provider in known_providers:
+            return parsed_model_name
+        else:
+            return model_name
+
     return model_name

tests/test_models_utils.py

@@ -1,70 +1,34 @@
 import pytest
-from deepeval.models.utils import parse_model_name
 
-
-class TestGetActualModelName:
-    """Test suite for the parse_model_name function."""
+class TestModelNameFormat:
+    """
+    Since parse_model_name function is removed, 
+    we expect model names to be used directly in full 'provider/model' format.
+    """
 
     @pytest.mark.parametrize(
-        "input_model_name,expected_output",
+        "model_name",
         [
-            # Standard provider/model format
-            ("openai/gpt-4o", "gpt-4o"),
-            ("anthropic/claude-3-opus", "claude-3-opus"),
-            ("cohere/command", "command"),
-            # No provider prefix
-            ("gpt-4o", "gpt-4o"),
-            ("claude-3-sonnet", "claude-3-sonnet"),
-            # Edge cases
-            ("", ""),  # Empty string
-            ("/", ""),  # Just a slash
-            ("openai/", ""),  # Provider with no model
-            (
-                "//model",
-                "/model",
-            ),  # Multiple slashes at start - splits only on first slash
-            (
-                "provider/model/version",
-                "model/version",
-            ),  # Multiple slashes in name
-            (
-                "provider/model-name/with/slashes",
-                "model-name/with/slashes",
-            ),  # Multiple slashes
-            # Mixed case and special characters
-            ("OpenAI/GPT-4o", "GPT-4o"),
-            ("custom-provider/model-123_test", "model-123_test"),
-            # Numerical and versioned names
-            ("openai/gpt-3.5-turbo", "gpt-3.5-turbo"),
-            ("anthropic/claude-2.1", "claude-2.1"),
+            "openai/gpt-4o",
+            "anthropic/claude-3-opus",
+            "cohere/command",
+            "local/llama-3",
+            "custom-provider/model-123_test",
+            "gpt-4o",
+            "claude-3-sonnet",
         ],
     )
-    def test_parse_model_name(self, input_model_name, expected_output):
-        """
-        Test that parse_model_name correctly extracts the model name
-        from various input formats.
-
-        Args:
-            input_model_name: The model name to process
-            expected_output: The expected result after processing
-        """
-        assert parse_model_name(input_model_name) == expected_output
-
-    def test_parse_model_name_type_preservation(self):
-        """Test that the function preserves the string type and doesn't modify the input."""
-        result = parse_model_name("provider/model")
-        assert isinstance(result, str)
-
-        result = parse_model_name("")
-        assert isinstance(result, str)
-
-    def test_parse_model_name_identity(self):
-        """Test that the function returns equal values for inputs without providers."""
-        test_cases = ["gpt-4", "claude-3", "command-r", "model-with-dashes"]
-        for model_name in test_cases:
-            assert parse_model_name(model_name) == model_name
-
-    def test_parse_model_name_none_value(self):
-        """Test that the function raises TypeError when None is passed."""
-        with pytest.raises(TypeError):
-            parse_model_name(None)
+    def test_model_name_direct_usage(self, model_name):
+        # Model names should be used as-is, with no parsing or stripping.
+        assert isinstance(model_name, str)
+        assert len(model_name) > 0
+
+    def test_model_name_format_contains_provider(self):
+        # Check that common provider/model format contains a slash
+        assert "/" in "openai/gpt-4o"
+        assert "/" in "anthropic/claude-3-opus"
+
+    def test_model_name_format_no_slash(self):
+        # Model names without provider are allowed as-is
+        model = "gpt-4o"
+        assert "/" not in model