torbjorn/my-pal-mcp-server
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Added proper temperature constraints to the model, fixes: https://github.com/BeehiveInnovations/zen-mcp-server/issues/78

Browse Source 
Prompt tweaks
This commit is contained in:
Fahad
2025-06-19 08:30:46 +04:00
parent ec3a466b1c
commit 9f3b70d6d7
13 changed files with 435 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
providers/base.py
View File

@@ -100,6 +100,26 @@ class DiscreteTemperatureConstraint(TemperatureConstraint):
return self.default_temp
def create_temperature_constraint(constraint_type: str) -> TemperatureConstraint:
"""Create temperature constraint object from configuration string.
Args:
constraint_type: Type of constraint ("fixed", "range", "discrete")
Returns:
TemperatureConstraint object based on configuration
"""
if constraint_type == "fixed":
# Fixed temperature models (O3/O4) only support temperature=1.0
return FixedTemperatureConstraint(1.0)
elif constraint_type == "discrete":
# For models with specific allowed values - using common OpenAI values as default
return DiscreteTemperatureConstraint([0.0, 0.3, 0.7, 1.0, 1.5, 2.0], 0.7)
else:
# Default range constraint (for "range" or None)
return RangeTemperatureConstraint(0.0, 2.0, 0.7)
@dataclass
class ModelCapabilities:
"""Capabilities and constraints for a specific model."""
@@ -114,6 +134,7 @@ class ModelCapabilities:
supports_function_calling: bool = False
supports_images: bool = False # Whether model can process images
max_image_size_mb: float = 0.0 # Maximum total size for all images in MB
supports_temperature: bool = True # Whether model accepts temperature parameter in API calls
# Temperature constraint object - preferred way to define temperature limits
temperature_constraint: TemperatureConstraint = field(
@@ -245,3 +266,17 @@ class ModelProvider(ABC):
List of all model names and alias targets known by this provider
"""
pass
def _resolve_model_name(self, model_name: str) -> str:
"""Resolve model shorthand to full name.
Base implementation returns the model name unchanged.
Subclasses should override to provide alias resolution.
Args:
model_name: Model name that may be an alias
Returns:
Resolved model name
"""
return model_name

1
providers/custom.py
View File

@@ -162,6 +162,7 @@ class CustomProvider(OpenAICompatibleProvider):
supports_system_prompts=True,
supports_streaming=True,
supports_function_calling=False, # Conservative default
supports_temperature=True, # Most custom models accept temperature parameter
temperature_constraint=RangeTemperatureConstraint(0.0, 2.0, 0.7),
)

1
providers/gemini.py
View File

@@ -94,6 +94,7 @@ class GeminiModelProvider(ModelProvider):
supports_function_calling=True,
supports_images=config.get("supports_images", False),
max_image_size_mb=config.get("max_image_size_mb", 0.0),
supports_temperature=True, # Gemini models accept temperature parameter
temperature_constraint=temp_constraint,
)

32
providers/openai_compatible.py
View File

@@ -448,23 +448,41 @@ class OpenAICompatibleProvider(ModelProvider):
completion_params = {
"model": model_name,
"messages": messages,
"temperature": temperature,
}
# Add max tokens if specified
if max_output_tokens:
# Check model capabilities once to determine parameter support
resolved_model = self._resolve_model_name(model_name)
# Get model capabilities once to avoid duplicate calls
try:
capabilities = self.get_capabilities(model_name)
# Defensive check for supports_temperature field (backward compatibility)
supports_temperature = getattr(capabilities, "supports_temperature", True)
except Exception as e:
# If capability check fails, fall back to conservative behavior
# Default to including temperature for most models (backward compatibility)
logging.debug(f"Failed to check temperature support for {model_name}: {e}")
supports_temperature = True
# Add temperature parameter if supported
if supports_temperature:
completion_params["temperature"] = temperature
# Add max tokens if specified and model supports it
# O3/O4 models that don't support temperature also don't support max_tokens
if max_output_tokens and supports_temperature:
completion_params["max_tokens"] = max_output_tokens
# Add any additional OpenAI-specific parameters
# Use capabilities to filter parameters for reasoning models
for key, value in kwargs.items():
if key in ["top_p", "frequency_penalty", "presence_penalty", "seed", "stop", "stream"]:
# Reasoning models (those that don't support temperature) also don't support these parameters
if not supports_temperature and key in ["top_p", "frequency_penalty", "presence_penalty"]:
continue # Skip unsupported parameters for reasoning models
completion_params[key] = value
# Check if this is o3-pro and needs the responses endpoint
resolved_model = model_name
if hasattr(self, "_resolve_model_name"):
resolved_model = self._resolve_model_name(model_name)
if resolved_model == "o3-pro-2025-06-10":
# This model requires the /v1/responses endpoint
# If it fails, we should not fall back to chat/completions

27
providers/openai_provider.py
View File

@@ -4,11 +4,10 @@ import logging
from typing import Optional
from .base import (
FixedTemperatureConstraint,
ModelCapabilities,
ModelResponse,
ProviderType,
RangeTemperatureConstraint,
create_temperature_constraint,
)
from .openai_compatible import OpenAICompatibleProvider
@@ -25,18 +24,24 @@ class OpenAIModelProvider(OpenAICompatibleProvider):
"supports_extended_thinking": False,
"supports_images": True, # O3 models support vision
"max_image_size_mb": 20.0, # 20MB per OpenAI docs
"supports_temperature": False, # O3 models don't accept temperature parameter
"temperature_constraint": "fixed", # Fixed at 1.0
},
"o3-mini": {
"context_window": 200_000, # 200K tokens
"supports_extended_thinking": False,
"supports_images": True, # O3 models support vision
"max_image_size_mb": 20.0, # 20MB per OpenAI docs
"supports_temperature": False, # O3 models don't accept temperature parameter
"temperature_constraint": "fixed", # Fixed at 1.0
},
"o3-pro-2025-06-10": {
"context_window": 200_000, # 200K tokens
"supports_extended_thinking": False,
"supports_images": True, # O3 models support vision
"max_image_size_mb": 20.0, # 20MB per OpenAI docs
"supports_temperature": False, # O3 models don't accept temperature parameter
"temperature_constraint": "fixed", # Fixed at 1.0
},
# Aliases
"o3-pro": "o3-pro-2025-06-10",
@@ -45,18 +50,24 @@ class OpenAIModelProvider(OpenAICompatibleProvider):
"supports_extended_thinking": False,
"supports_images": True, # O4 models support vision
"max_image_size_mb": 20.0, # 20MB per OpenAI docs
"supports_temperature": False, # O4 models don't accept temperature parameter
"temperature_constraint": "fixed", # Fixed at 1.0
},
"o4-mini-high": {
"context_window": 200_000, # 200K tokens
"supports_extended_thinking": False,
"supports_images": True, # O4 models support vision
"max_image_size_mb": 20.0, # 20MB per OpenAI docs
"supports_temperature": False, # O4 models don't accept temperature parameter
"temperature_constraint": "fixed", # Fixed at 1.0
},
"gpt-4.1-2025-04-14": {
"context_window": 1_000_000, # 1M tokens
"supports_extended_thinking": False,
"supports_images": True, # GPT-4.1 supports vision
"max_image_size_mb": 20.0, # 20MB per OpenAI docs
"supports_temperature": True, # Regular models accept temperature parameter
"temperature_constraint": "range", # 0.0-2.0 range
},
# Shorthands
"mini": "o4-mini", # Default 'mini' to latest mini model
@@ -90,13 +101,10 @@ class OpenAIModelProvider(OpenAICompatibleProvider):
config = self.SUPPORTED_MODELS[resolved_name]
# Define temperature constraints per model
if resolved_name in ["o3", "o3-mini", "o3-pro", "o3-pro-2025-06-10", "o4-mini", "o4-mini-high"]:
# O3 and O4 reasoning models only support temperature=1.0
temp_constraint = FixedTemperatureConstraint(1.0)
else:
# Other OpenAI models (including GPT-4.1) support 0.0-2.0 range
temp_constraint = RangeTemperatureConstraint(0.0, 2.0, 0.7)
# Get temperature constraints and support from configuration
supports_temperature = config.get("supports_temperature", True) # Default to True for backward compatibility
temp_constraint_type = config.get("temperature_constraint", "range") # Default to range
temp_constraint = create_temperature_constraint(temp_constraint_type)
return ModelCapabilities(
provider=ProviderType.OPENAI,
@@ -109,6 +117,7 @@ class OpenAIModelProvider(OpenAICompatibleProvider):
supports_function_calling=True,
supports_images=config.get("supports_images", False),
max_image_size_mb=config.get("max_image_size_mb", 0.0),
supports_temperature=supports_temperature,
temperature_constraint=temp_constraint,
)

22
providers/openrouter_registry.py
View File

@@ -8,7 +8,12 @@ from typing import Optional
from utils.file_utils import read_json_file
from .base import ModelCapabilities, ProviderType, RangeTemperatureConstraint
from .base import (
ModelCapabilities,
ProviderType,
TemperatureConstraint,
create_temperature_constraint,
)
@dataclass
@@ -25,9 +30,21 @@ class OpenRouterModelConfig:
supports_json_mode: bool = False
supports_images: bool = False # Whether model can process images
max_image_size_mb: float = 0.0 # Maximum total size for all images in MB
supports_temperature: bool = True # Whether model accepts temperature parameter in API calls
temperature_constraint: Optional[str] = (
None # Type of temperature constraint: "fixed", "range", "discrete", or None for default range
)
is_custom: bool = False # True for models that should only be used with custom endpoints
description: str = ""
def _create_temperature_constraint(self) -> TemperatureConstraint:
"""Create temperature constraint object from configuration.
Returns:
TemperatureConstraint object based on configuration
"""
return create_temperature_constraint(self.temperature_constraint or "range")
def to_capabilities(self) -> ModelCapabilities:
"""Convert to ModelCapabilities object."""
return ModelCapabilities(
@@ -41,7 +58,8 @@ class OpenRouterModelConfig:
supports_function_calling=self.supports_function_calling,
supports_images=self.supports_images,
max_image_size_mb=self.max_image_size_mb,
temperature_constraint=RangeTemperatureConstraint(0.0, 2.0, 1.0),
supports_temperature=self.supports_temperature,
temperature_constraint=self._create_temperature_constraint(),
)