Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

15.9 KB
Raw Blame
"""Integration tests for get_llm() dispatch through the provider registry.

PR #3984 removed the procedural if/elif provider chain in get_llm(); the
registry path (auto-discovery -> provider class create_llm) is now the only
construction path. These tests pin the full dispatch chain end to end:

    get_llm(provider=..., settings_snapshot=...)
        -> _llm_registry (populated by discover_providers() at import)
        -> <Provider>.create_llm
        -> langchain client constructor (patched at the provider module)

Only the langchain client class itself is patched — settings resolution,
registry dispatch, URL handling, key resolution, and the max_tokens cap all
execute real production code.
"""

import pytest
from unittest.mock import patch

from langchain_core.language_models.fake_chat_models import FakeListChatModel

from local_deep_research.config.llm_config import get_llm
from local_deep_research.llm import is_llm_registered
from local_deep_research.llm.providers import get_discovered_provider_options
from local_deep_research.llm.providers.base import normalize_provider

OPENAI_CHAT = (
    "local_deep_research.llm.providers.implementations.openai.ChatOpenAI"
)
ANTHROPIC_CHAT = (
    "local_deep_research.llm.providers.implementations.anthropic.ChatAnthropic"
)
OLLAMA_CHAT = (
    "local_deep_research.llm.providers.implementations.ollama.ChatOllama"
)
# LM Studio and llama.cpp construct their client in the shared parent
# (OpenAICompatibleProvider._create_llm_instance), so the patch point is
# openai_base, not the implementation module.
OPENAI_BASE_CHAT = "local_deep_research.llm.providers.openai_base.ChatOpenAI"


@pytest.fixture(autouse=True, scope="module")
def _ensure_providers_registered():
    """Re-run auto-discovery in case an earlier test cleared the registry."""
    from local_deep_research.llm.providers import discover_providers

    discover_providers(force_refresh=True)


def _snapshot(**overrides):
    """Minimal permissive settings snapshot for the live dispatch path."""
    snap = {
        "llm.supports_max_tokens": True,
        "llm.max_tokens": 4096,
        "rate_limiting.llm_enabled": False,
        "search.tool": "searxng",
    }
    snap.update(overrides)
    return snap


def _fake_llm():
    return FakeListChatModel(responses=["ok"])


class TestRegistryPopulation:
    """Every built-in provider must be reachable through the registry,
    since get_llm() no longer has any per-provider construction code."""

    def test_all_discovered_providers_are_registered(self):
        discovered = [
            normalize_provider(option["value"])
            for option in get_discovered_provider_options()
        ]
        unregistered = [p for p in discovered if not is_llm_registered(p)]
        assert unregistered == [], (
            f"Providers missing from the LLM registry: {unregistered}. "
            "get_llm() has no fallback construction path for these anymore."
        )


class TestCloudProviderDispatch:
    def test_openai_dispatches_through_provider_class(self):
        snapshot = _snapshot(**{"llm.openai.api_key": "sk-test-key"})
        fake = _fake_llm()
        with patch(OPENAI_CHAT, return_value=fake) as mock_chat:
            result = get_llm(
                provider="openai",
                model_name="gpt-4o-mini",
                temperature=0.5,
                settings_snapshot=snapshot,
            )
        kwargs = mock_chat.call_args.kwargs
        assert kwargs["model"] == "gpt-4o-mini"
        assert kwargs["api_key"] == "sk-test-key"
        assert kwargs["temperature"] == 0.5
        assert result.base_llm is fake

    def test_anthropic_dispatches_through_provider_class(self):
        snapshot = _snapshot(**{"llm.anthropic.api_key": "sk-ant-test"})
        fake = _fake_llm()
        with patch(ANTHROPIC_CHAT, return_value=fake) as mock_chat:
            result = get_llm(
                provider="anthropic",
                model_name="claude-sonnet-4-5",
                settings_snapshot=snapshot,
            )
        kwargs = mock_chat.call_args.kwargs
        assert kwargs["model"] == "claude-sonnet-4-5"
        assert kwargs["anthropic_api_key"] == "sk-ant-test"
        assert result.base_llm is fake

    def test_anthropic_missing_key_raises_value_error(self):
        """Required-key semantics survive on the live path: no placeholder."""
        with pytest.raises(ValueError, match="API key not configured"):
            get_llm(
                provider="anthropic",
                model_name="claude-sonnet-4-5",
                settings_snapshot=_snapshot(),
            )


class TestOpenAIEndpointOverride:
    def test_explicit_url_overrides_snapshot_without_mutating_it(self):
        snapshot = _snapshot(
            **{
                "llm.openai_endpoint.url": "http://localhost:9000/v1",
            }
        )

        with patch(OPENAI_BASE_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="openai_endpoint",
                model_name="custom-model",
                openai_endpoint_url="http://localhost:8000/v1",
                settings_snapshot=snapshot,
            )

        assert (
            mock_chat.call_args.kwargs["base_url"] == "http://localhost:8000/v1"
        )
        assert snapshot["llm.openai_endpoint.url"] == "http://localhost:9000/v1"

    def test_explicit_local_url_works_without_snapshot(self):
        with patch(OPENAI_BASE_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="openai_endpoint",
                model_name="custom-model",
                openai_endpoint_url="http://localhost:8000/v1",
                settings_snapshot=None,
            )

        assert (
            mock_chat.call_args.kwargs["base_url"] == "http://localhost:8000/v1"
        )

    def test_snapshot_url_is_used_when_override_is_omitted(self):
        snapshot = _snapshot(
            **{
                "llm.openai_endpoint.url": "http://localhost:9000/v1",
            }
        )

        with patch(OPENAI_BASE_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="openai_endpoint",
                model_name="custom-model",
                settings_snapshot=snapshot,
            )

        assert (
            mock_chat.call_args.kwargs["base_url"] == "http://localhost:9000/v1"
        )

    def test_empty_explicit_url_is_rejected(self):
        with pytest.raises(
            ValueError, match="openai_endpoint_url must be a non-empty string"
        ):
            get_llm(
                provider="openai_endpoint",
                model_name="custom-model",
                openai_endpoint_url="  ",
                settings_snapshot=_snapshot(),
            )

    def test_non_string_explicit_url_is_rejected(self):
        with pytest.raises(
            ValueError, match="openai_endpoint_url must be a non-empty string"
        ):
            get_llm(
                provider="openai_endpoint",
                model_name="custom-model",
                openai_endpoint_url=123,
                settings_snapshot=_snapshot(),
            )

    def test_explicit_url_is_ignored_for_non_endpoint_provider(self):
        snapshot = _snapshot(**{"llm.ollama.url": "http://localhost:11434"})
        with patch(OLLAMA_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="ollama",
                model_name="llama3.1:8b",
                openai_endpoint_url=123,
                settings_snapshot=snapshot,
            )

        assert mock_chat.call_args.kwargs["base_url"] == (
            "http://localhost:11434"
        )


class TestLocalProviderDispatch:
    def test_lmstudio_appends_v1_suffix(self):
        """#4532: a URL without /v1 gets the suffix on the live path."""
        snapshot = _snapshot(**{"llm.lmstudio.url": "http://localhost:1234"})
        with patch(OPENAI_BASE_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="lmstudio",
                model_name="qwen2.5-7b",
                settings_snapshot=snapshot,
            )
        kwargs = mock_chat.call_args.kwargs
        assert kwargs["base_url"] == "http://localhost:1234/v1"
        assert kwargs["api_key"] == "not-required"

    def test_lmstudio_does_not_double_v1_suffix(self):
        snapshot = _snapshot(**{"llm.lmstudio.url": "http://localhost:1234/v1"})
        with patch(OPENAI_BASE_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="lmstudio",
                model_name="qwen2.5-7b",
                settings_snapshot=snapshot,
            )
        assert (
            mock_chat.call_args.kwargs["base_url"] == "http://localhost:1234/v1"
        )

    def test_lmstudio_real_api_key_passed_through(self):
        snapshot = _snapshot(
            **{
                "llm.lmstudio.url": "http://localhost:1234/v1",
                "llm.lmstudio.api_key": "lms-real-key",
            }
        )
        with patch(OPENAI_BASE_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="lmstudio",
                model_name="qwen2.5-7b",
                settings_snapshot=snapshot,
            )
        assert mock_chat.call_args.kwargs["api_key"] == "lms-real-key"

    def test_llamacpp_uses_placeholder_key_and_verbatim_url(self):
        """llama.cpp does NOT force /v1 — the user-provided URL is used as-is."""
        snapshot = _snapshot(**{"llm.llamacpp.url": "http://localhost:8080"})
        with patch(OPENAI_BASE_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="llamacpp",
                model_name="llama-3.1-8b",
                settings_snapshot=snapshot,
            )
        kwargs = mock_chat.call_args.kwargs
        assert kwargs["base_url"] == "http://localhost:8080"
        assert kwargs["api_key"] == "not-required"

    def test_ollama_enable_thinking_false_reaches_chatollama(self):
        """#3984's headline bug fix: llm.ollama.enable_thinking was silently
        ignored on the live path before this PR."""
        snapshot = _snapshot(
            **{
                "llm.ollama.url": "http://localhost:11434",
                "llm.ollama.enable_thinking": False,
            }
        )
        fake = _fake_llm()
        with patch(OLLAMA_CHAT, return_value=fake) as mock_chat:
            result = get_llm(
                provider="ollama",
                model_name="deepseek-r1:14b",
                settings_snapshot=snapshot,
            )
        kwargs = mock_chat.call_args.kwargs
        assert kwargs["reasoning"] is False
        assert kwargs["model"] == "deepseek-r1:14b"
        assert result.base_llm is fake

    def test_ollama_enable_thinking_default_true(self):
        snapshot = _snapshot(**{"llm.ollama.url": "http://localhost:11434"})
        with patch(OLLAMA_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="ollama",
                model_name="deepseek-r1:14b",
                settings_snapshot=snapshot,
            )
        assert mock_chat.call_args.kwargs["reasoning"] is True


class TestMaxTokensCap:
    """The 80%-of-context-window cap previously lived only in dead code;
    these pin it on the live path end to end through get_llm()."""

    def test_cloud_max_tokens_capped_at_80_percent_of_window(self):
        snapshot = _snapshot(
            **{
                "llm.openai.api_key": "sk-test-key",
                "llm.context_window_unrestricted": False,
                "llm.context_window_size": 10000,
                "llm.max_tokens": 200000,
            }
        )
        with patch(OPENAI_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="openai",
                model_name="gpt-4o-mini",
                settings_snapshot=snapshot,
            )
        assert mock_chat.call_args.kwargs["max_tokens"] == 8000

    def test_cloud_unrestricted_window_passes_max_tokens_uncapped(self):
        snapshot = _snapshot(
            **{
                "llm.openai.api_key": "sk-test-key",
                "llm.context_window_unrestricted": True,
                "llm.max_tokens": 200000,
            }
        )
        with patch(OPENAI_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="openai",
                model_name="gpt-4o-mini",
                settings_snapshot=snapshot,
            )
        assert mock_chat.call_args.kwargs["max_tokens"] == 200000

    def test_ollama_num_ctx_from_shared_helper_no_max_tokens(self):
        """Ollama's window resolves through the shared helper (same source
        as context_limit bookkeeping); no max_tokens kwarg is passed since
        ChatOllama silently ignores it."""
        snapshot = _snapshot(
            **{
                "llm.ollama.url": "http://localhost:11434",
                "llm.local_context_window_size": 4096,
                "llm.max_tokens": 200000,
            }
        )
        with patch(OLLAMA_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="ollama",
                model_name="llama3.1:8b",
                settings_snapshot=snapshot,
            )
        kwargs = mock_chat.call_args.kwargs
        assert kwargs["num_ctx"] == 4096
        assert "max_tokens" not in kwargs

    def test_lmstudio_capped_by_local_window_not_cloud_window(self):
        """LM Studio resolves its window via provider_key through the
        LOCAL_PROVIDERS branch — the cloud llm.context_window_size must
        play no role."""
        snapshot = _snapshot(
            **{
                "llm.lmstudio.url": "http://localhost:1234/v1",
                "llm.context_window_unrestricted": False,
                "llm.context_window_size": 128000,
                "llm.local_context_window_size": 4096,
                "llm.max_tokens": 200000,
            }
        )
        with patch(OPENAI_BASE_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="lmstudio",
                model_name="qwen2.5-7b",
                settings_snapshot=snapshot,
            )
        assert mock_chat.call_args.kwargs["max_tokens"] == int(4096 * 0.8)

    def test_unset_max_tokens_omits_kwarg(self):
        """Partial snapshots without llm.max_tokens must not inject a
        hardcoded default — the provider SDK's own default applies."""
        snapshot = _snapshot(**{"llm.openai.api_key": "sk-test-key"})
        del snapshot["llm.max_tokens"]
        with patch(OPENAI_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="openai",
                model_name="gpt-4o-mini",
                settings_snapshot=snapshot,
            )
        assert "max_tokens" not in mock_chat.call_args.kwargs

    def test_supports_max_tokens_false_omits_kwarg(self):
        snapshot = _snapshot(
            **{
                "llm.openai.api_key": "sk-test-key",
                "llm.supports_max_tokens": False,
            }
        )
        with patch(OPENAI_CHAT, return_value=_fake_llm()) as mock_chat:
            get_llm(
                provider="openai",
                model_name="gpt-4o-mini",
                settings_snapshot=snapshot,
            )
        assert "max_tokens" not in mock_chat.call_args.kwargs


class TestContextLimitBookkeeping:
    def test_context_limit_set_in_research_context(self):
        """Overflow detection relies on context_limit being populated even
        though the registered path returns before get_llm's own bookkeeping."""
        snapshot = _snapshot(
            **{
                "llm.openai.api_key": "sk-test-key",
                "llm.context_window_unrestricted": False,
                "llm.context_window_size": 10000,
            }
        )
        research_context = {}
        with patch(OPENAI_CHAT, return_value=_fake_llm()):
            get_llm(
                provider="openai",
                model_name="gpt-4o-mini",
                settings_snapshot=snapshot,
                research_context=research_context,
            )
        assert research_context["context_limit"] == 10000