137 lines
5.3 KiB
Python
137 lines
5.3 KiB
Python
"""OpenAI model provider implementation."""
|
|
|
|
import logging
|
|
from typing import TYPE_CHECKING, ClassVar, Optional
|
|
|
|
if TYPE_CHECKING:
|
|
from tools.models import ToolModelCategory
|
|
|
|
from .openai_compatible import OpenAICompatibleProvider
|
|
from .registries.openai import OpenAIModelRegistry
|
|
from .registry_provider_mixin import RegistryBackedProviderMixin
|
|
from .shared import ModelCapabilities, ProviderType
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
class OpenAIModelProvider(RegistryBackedProviderMixin, OpenAICompatibleProvider):
|
|
"""Implementation that talks to api.openai.com using rich model metadata.
|
|
|
|
In addition to the built-in catalogue, the provider can surface models
|
|
defined in ``conf/custom_models.json`` (for organisations running their own
|
|
OpenAI-compatible gateways) while still respecting restriction policies.
|
|
"""
|
|
|
|
REGISTRY_CLASS = OpenAIModelRegistry
|
|
MODEL_CAPABILITIES: ClassVar[dict[str, ModelCapabilities]] = {}
|
|
|
|
def __init__(self, api_key: str, **kwargs):
|
|
"""Initialize OpenAI provider with API key."""
|
|
self._ensure_registry()
|
|
# Set default OpenAI base URL, allow override for regions/custom endpoints
|
|
kwargs.setdefault("base_url", "https://api.openai.com/v1")
|
|
super().__init__(api_key, **kwargs)
|
|
self._invalidate_capability_cache()
|
|
|
|
# ------------------------------------------------------------------
|
|
# Capability surface
|
|
# ------------------------------------------------------------------
|
|
|
|
def _lookup_capabilities(
|
|
self,
|
|
canonical_name: str,
|
|
requested_name: Optional[str] = None,
|
|
) -> Optional[ModelCapabilities]:
|
|
"""Look up OpenAI capabilities from built-ins or the custom registry."""
|
|
|
|
self._ensure_registry()
|
|
builtin = super()._lookup_capabilities(canonical_name, requested_name)
|
|
if builtin is not None:
|
|
return builtin
|
|
|
|
try:
|
|
from .registries.openrouter import OpenRouterModelRegistry
|
|
|
|
registry = OpenRouterModelRegistry()
|
|
config = registry.get_model_config(canonical_name)
|
|
|
|
if config and config.provider == ProviderType.OPENAI:
|
|
return config
|
|
|
|
except Exception as exc: # pragma: no cover - registry failures are non-critical
|
|
logger.debug(f"Could not resolve custom OpenAI model '{canonical_name}': {exc}")
|
|
|
|
return None
|
|
|
|
def _finalise_capabilities(
|
|
self,
|
|
capabilities: ModelCapabilities,
|
|
canonical_name: str,
|
|
requested_name: str,
|
|
) -> ModelCapabilities:
|
|
"""Ensure registry-sourced models report the correct provider type."""
|
|
|
|
if capabilities.provider != ProviderType.OPENAI:
|
|
capabilities.provider = ProviderType.OPENAI
|
|
return capabilities
|
|
|
|
def _raise_unsupported_model(self, model_name: str) -> None:
|
|
raise ValueError(f"Unsupported OpenAI model: {model_name}")
|
|
|
|
# ------------------------------------------------------------------
|
|
# Provider identity
|
|
# ------------------------------------------------------------------
|
|
|
|
def get_provider_type(self) -> ProviderType:
|
|
"""Get the provider type."""
|
|
return ProviderType.OPENAI
|
|
|
|
# ------------------------------------------------------------------
|
|
# Provider preferences
|
|
# ------------------------------------------------------------------
|
|
|
|
def get_preferred_model(self, category: "ToolModelCategory", allowed_models: list[str]) -> Optional[str]:
|
|
"""Get OpenAI's preferred model for a given category from allowed models.
|
|
|
|
Args:
|
|
category: The tool category requiring a model
|
|
allowed_models: Pre-filtered list of models allowed by restrictions
|
|
|
|
Returns:
|
|
Preferred model name or None
|
|
"""
|
|
from tools.models import ToolModelCategory
|
|
|
|
if not allowed_models:
|
|
return None
|
|
|
|
# Helper to find first available from preference list
|
|
def find_first(preferences: list[str]) -> Optional[str]:
|
|
"""Return first available model from preference list."""
|
|
for model in preferences:
|
|
if model in allowed_models:
|
|
return model
|
|
return None
|
|
|
|
if category == ToolModelCategory.EXTENDED_REASONING:
|
|
# Prefer models with extended thinking support
|
|
# GPT-5-Codex first for coding tasks
|
|
preferred = find_first(["gpt-5-codex", "gpt-5-pro", "o3", "o3-pro", "gpt-5"])
|
|
return preferred if preferred else allowed_models[0]
|
|
|
|
elif category == ToolModelCategory.FAST_RESPONSE:
|
|
# Prefer fast, cost-efficient models
|
|
# GPT-5 models for speed, GPT-5-Codex after (premium pricing but cached)
|
|
preferred = find_first(["gpt-5", "gpt-5-mini", "gpt-5-codex", "o4-mini", "o3-mini"])
|
|
return preferred if preferred else allowed_models[0]
|
|
|
|
else: # BALANCED or default
|
|
# Prefer balanced performance/cost models
|
|
# Include GPT-5-Codex for coding workflows
|
|
preferred = find_first(["gpt-5", "gpt-5-codex", "gpt-5-pro", "gpt-5-mini", "o4-mini", "o3-mini"])
|
|
return preferred if preferred else allowed_models[0]
|
|
|
|
|
|
# Load registry data at import time so dependent providers (Azure) can reuse it
|
|
OpenAIModelProvider._ensure_registry()
|