DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.
Python ai-agentsai-tutorclawbotcli-tooldeepresearchinteractive-learninglearningmulti-agent-systemsrag
"""Runtime resolution of ``api_format`` and per-model capability overrides."""
from __future__ import annotations
from typing import Any
from deeptutor.runtime.agentic.client import can_use_native_tool_calling
from deeptutor.services.config.provider_runtime import resolve_llm_runtime_config
from deeptutor.services.llm import provider_factory
from deeptutor.services.llm.capabilities import (
catalog_capability_override,
set_catalog_capability_overrides,
supports_tools,
)
from deeptutor.services.llm.config import LLMConfig
from deeptutor.services.llm.provider_core.anthropic_provider import AnthropicProvider
from deeptutor.services.llm.provider_core.azure_openai_provider import AzureOpenAIProvider
from deeptutor.services.llm.provider_core.openai_compat_provider import OpenAICompatProvider
def _catalog(profile: dict[str, Any]) -> dict[str, Any]:
return {
"version": 1,
"services": {
"llm": {
"active_profile_id": profile["id"],
"active_model_id": profile["models"][0]["id"],
"profiles": [profile],
},
"embedding": {"active_profile_id": None, "profiles": []},
"search": {"active_profile_id": None, "profiles": []},
},
}
def _profile(**overrides: Any) -> dict[str, Any]:
profile = {
"id": "llm-p",
"name": "LLM",
"binding": "custom",
"base_url": "https://relay.example/v1",
"api_key": "test-key",
"api_version": "",
"extra_headers": {},
"models": [{"id": "llm-m", "name": "m", "model": "some-model"}],
}
profile.update(overrides)
return profile
def _runtime_config(resolved) -> LLMConfig:
return LLMConfig(
model=resolved.model,
api_key=resolved.api_key,
base_url=resolved.base_url,
effective_url=resolved.effective_url,
binding=resolved.binding,
provider_name=resolved.provider_name,
provider_mode=resolved.provider_mode,
api_version=resolved.api_version,
extra_headers=resolved.extra_headers,
wire_api=resolved.wire_api,
api_format=resolved.api_format,
)
def test_custom_endpoint_in_anthropic_format_gets_the_anthropic_backend() -> None:
resolved = resolve_llm_runtime_config(
catalog=_catalog(
_profile(api_format="anthropic", base_url="https://relay.example/anthropic")
)
)
assert resolved.provider_name == "custom"
assert resolved.api_format == "anthropic"
assert resolved.wire_api == "auto"
provider = provider_factory._build_runtime_provider(_runtime_config(resolved))
assert isinstance(provider, AnthropicProvider)
def test_legacy_custom_anthropic_binding_resolves_exactly_as_before() -> None:
resolved = resolve_llm_runtime_config(
catalog=_catalog(
_profile(binding="custom_anthropic", base_url="https://relay.example/anthropic")
)
)
assert resolved.provider_name == "custom_anthropic"
assert resolved.api_format == "anthropic"
assert isinstance(
provider_factory._build_runtime_provider(_runtime_config(resolved)), AnthropicProvider
)
def test_openai_responses_format_sets_wire_api() -> None:
resolved = resolve_llm_runtime_config(catalog=_catalog(_profile(api_format="openai_responses")))
assert resolved.wire_api == "responses"
provider = provider_factory._build_runtime_provider(_runtime_config(resolved))
assert isinstance(provider, OpenAICompatProvider)
assert provider._wire_api == "responses"
def test_legacy_wire_api_still_drives_the_protocol() -> None:
resolved = resolve_llm_runtime_config(catalog=_catalog(_profile(wire_api="responses")))
assert resolved.api_format == "openai_responses"
assert resolved.wire_api == "responses"
def test_minimax_anthropic_format_uses_the_vendor_anthropic_endpoint() -> None:
resolved = resolve_llm_runtime_config(
catalog=_catalog(_profile(binding="minimax", base_url="", api_format="anthropic"))
)
# The profile left base_url blank, so the vendor default for *this format*
# is what the runtime should fall back to.
assert resolved.effective_url == "https://api.minimax.io/anthropic"
provider = provider_factory._build_runtime_provider(_runtime_config(resolved))
assert isinstance(provider, AnthropicProvider)
assert str(provider._client.base_url).rstrip("/") == "https://api.minimax.io/anthropic"
def test_openai_profile_with_api_version_is_an_azure_deployment() -> None:
config = LLMConfig(
model="deployment",
api_key="k",
base_url="https://example.openai.azure.com",
binding="openai",
provider_name="openai",
api_version="2024-02-01",
)
assert isinstance(provider_factory._build_runtime_provider(config), AzureOpenAIProvider)
def test_model_capability_overrides_are_published_on_resolution() -> None:
profile = _profile(binding="ollama", base_url="http://localhost:11434/v1", api_key="")
profile["models"][0]["capabilities"] = {"tools": True, "vision": False}
try:
resolve_llm_runtime_config(catalog=_catalog(profile))
assert catalog_capability_override("ollama", "some-model", "supports_tools") is True
assert supports_tools("ollama", "some-model") is True
# The agentic gate normally refuses tools on local servers; the
# user's declaration is exactly what should override that.
assert can_use_native_tool_calling(binding="ollama", model="some-model") is True
assert can_use_native_tool_calling(binding="ollama", model="another-model") is False
# Removing the override in the catalog removes it at runtime too.
del profile["models"][0]["capabilities"]
resolve_llm_runtime_config(catalog=_catalog(profile))
assert catalog_capability_override("ollama", "some-model", "supports_tools") is None
assert can_use_native_tool_calling(binding="ollama", model="some-model") is False
finally:
set_catalog_capability_overrides([])