lightrag-cn-git/tests/llm/bedrock_impl/test_bedrock_llm.py

import importlib
import logging
import os
import sys
from types import SimpleNamespace
from unittest.mock import AsyncMock, Mock, patch

import pytest
from fastapi import APIRouter
from fastapi.testclient import TestClient

from lightrag.llm.bedrock import (
    bedrock_complete,
    bedrock_complete_if_cache,
    bedrock_embed,
)


def _reload_api_modules_if_mocked() -> None:
    """Drop Mock-replaced lightrag.api entries so importlib reloads the real modules.

    Other test files (e.g. test_token_auto_renewal.py) replace
    ``sys.modules["lightrag.api.config"]`` with a Mock at import time. When
    pytest collects those files before ours, any subsequent
    ``from .config import global_args`` inside lightrag_server picks up the
    Mock, which breaks ``create_app`` in create_app_* tests below.
    """
    for modname in (
        "lightrag.api.lightrag_server",
        "lightrag.api.auth",
        "lightrag.api.config",
    ):
        if isinstance(sys.modules.get(modname), Mock):
            sys.modules.pop(modname, None)


class _FakeBedrockClient:
    def __init__(self, captured_calls: list[dict]):
        self._captured_calls = captured_calls

    async def __aenter__(self):
        return self

    async def __aexit__(self, exc_type, exc, tb):
        return False

    async def converse(self, **kwargs):
        self._captured_calls.append(kwargs)
        return {
            "output": {
                "message": {
                    "content": [
                        {
                            "text": '{"high_level_keywords":["AI"],"low_level_keywords":["RAG"]}'
                        }
                    ]
                }
            }
        }


class _FakeSession:
    def __init__(self, captured_calls: list[dict], client_kwargs_calls: list[dict]):
        self._captured_calls = captured_calls
        self._client_kwargs_calls = client_kwargs_calls

    def client(self, *_args, **kwargs):
        self._client_kwargs_calls.append(dict(kwargs))
        return _FakeBedrockClient(self._captured_calls)


class _FakeReasoningClient(_FakeBedrockClient):
    async def converse(self, **kwargs):
        self._captured_calls.append(kwargs)
        return {
            "output": {
                "message": {
                    "content": [
                        {
                            "reasoningContent": {
                                "reasoningText": {"text": "internal thought"}
                            }
                        },
                        {"text": "final answer"},
                    ]
                }
            }
        }


class _FakeReasoningSession(_FakeSession):
    def client(self, *_args, **kwargs):
        self._client_kwargs_calls.append(dict(kwargs))
        return _FakeReasoningClient(self._captured_calls)


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_complete_skips_reasoning_content_block(monkeypatch):
    monkeypatch.delenv("AWS_REGION", raising=False)
    captured_calls: list[dict] = []

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeReasoningSession(captured_calls, []),
    ):
        result = await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            extra_fields={"reasoning_config": {"type": "enabled"}},
        )

    assert result == "final answer"


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_complete_forwards_keyword_extraction_to_if_cache():
    hashing_kv = SimpleNamespace(global_config={"llm_model_name": "bedrock-model"})

    with patch(
        "lightrag.llm.bedrock.bedrock_complete_if_cache",
        AsyncMock(return_value="{}"),
    ) as mocked_complete:
        await bedrock_complete(
            prompt="hello",
            hashing_kv=hashing_kv,
            keyword_extraction=True,
        )

    assert mocked_complete.await_args.kwargs["keyword_extraction"] is True


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_keyword_extraction_does_not_inject_system_prompt(monkeypatch):
    captured_calls: list[dict] = []
    client_kwargs_calls: list[dict] = []
    monkeypatch.delenv("AWS_REGION", raising=False)

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeSession(captured_calls, client_kwargs_calls),
    ):
        result = await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            response_format={"type": "json_object"},
        )

    assert result == '{"high_level_keywords":["AI"],"low_level_keywords":["RAG"]}'
    assert len(captured_calls) == 1
    assert "system" not in captured_calls[0]
    assert client_kwargs_calls[-1] == {"region_name": None}


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_default_endpoint_sentinel_uses_sdk_default(monkeypatch):
    captured_calls: list[dict] = []
    client_kwargs_calls: list[dict] = []
    monkeypatch.delenv("AWS_REGION", raising=False)

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeSession(captured_calls, client_kwargs_calls),
    ):
        await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            endpoint_url="DEFAULT_BEDROCK_ENDPOINT",
        )

    assert client_kwargs_calls[-1] == {"region_name": None}


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_empty_endpoint_url_uses_sdk_default(monkeypatch):
    captured_calls: list[dict] = []
    client_kwargs_calls: list[dict] = []
    monkeypatch.delenv("AWS_REGION", raising=False)

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeSession(captured_calls, client_kwargs_calls),
    ):
        await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            endpoint_url="",
        )

    assert client_kwargs_calls[-1] == {"region_name": None}


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_custom_endpoint_url_is_forwarded(monkeypatch):
    captured_calls: list[dict] = []
    client_kwargs_calls: list[dict] = []
    monkeypatch.delenv("AWS_REGION", raising=False)

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeSession(captured_calls, client_kwargs_calls),
    ):
        await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            endpoint_url="https://proxy.example.com",
        )

    assert client_kwargs_calls[-1] == {
        "region_name": None,
        "endpoint_url": "https://proxy.example.com",
    }


class _FakeEmbeddingBody:
    async def json(self):
        return {"embedding": [0.1] * 1024}


class _FakeEmbeddingResponse:
    def get(self, key):
        assert key == "body"
        return _FakeEmbeddingBody()


class _FakeEmbeddingClient(_FakeBedrockClient):
    async def invoke_model(self, **_kwargs):
        return _FakeEmbeddingResponse()


class _FakeEmbeddingSession(_FakeSession):
    def client(self, *_args, **kwargs):
        self._client_kwargs_calls.append(dict(kwargs))
        return _FakeEmbeddingClient(self._captured_calls)


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_embed_custom_endpoint_url_is_forwarded(monkeypatch):
    captured_calls: list[dict] = []
    client_kwargs_calls: list[dict] = []
    monkeypatch.delenv("AWS_REGION", raising=False)

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeEmbeddingSession(captured_calls, client_kwargs_calls),
    ):
        await bedrock_embed(
            texts=["hello"],
            endpoint_url="https://proxy.example.com",
        )

    assert client_kwargs_calls[-1] == {
        "region_name": None,
        "endpoint_url": "https://proxy.example.com",
    }


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_embed_default_endpoint_sentinel_uses_sdk_default(monkeypatch):
    captured_calls: list[dict] = []
    client_kwargs_calls: list[dict] = []
    monkeypatch.delenv("AWS_REGION", raising=False)

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeEmbeddingSession(captured_calls, client_kwargs_calls),
    ):
        await bedrock_embed(
            texts=["hello"],
            endpoint_url="DEFAULT_BEDROCK_ENDPOINT",
        )

    assert client_kwargs_calls[-1] == {"region_name": None}


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_embed_empty_endpoint_url_uses_sdk_default(monkeypatch):
    captured_calls: list[dict] = []
    client_kwargs_calls: list[dict] = []
    monkeypatch.delenv("AWS_REGION", raising=False)

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeEmbeddingSession(captured_calls, client_kwargs_calls),
    ):
        await bedrock_embed(
            texts=["hello"],
            endpoint_url="",
        )

    assert client_kwargs_calls[-1] == {"region_name": None}


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_complete_forwards_explicit_sigv4_client_kwargs(monkeypatch):
    monkeypatch.delenv("AWS_REGION", raising=False)
    captured_calls: list[dict] = []
    client_kwargs_calls: list[dict] = []

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeSession(captured_calls, client_kwargs_calls),
    ):
        await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            aws_region="us-west-2",
            aws_access_key_id="akid",
            aws_secret_access_key="secret",
            aws_session_token="session",
            endpoint_url="https://proxy.example.com",
        )

    assert client_kwargs_calls[-1] == {
        "region_name": "us-west-2",
        "endpoint_url": "https://proxy.example.com",
        "aws_access_key_id": "akid",
        "aws_secret_access_key": "secret",
        "aws_session_token": "session",
    }


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_extra_fields_maps_to_additional_model_request_fields(
    monkeypatch,
):
    monkeypatch.delenv("AWS_REGION", raising=False)
    captured_calls: list[dict] = []

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeSession(captured_calls, []),
    ):
        await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            extra_fields={"reasoning_config": {"type": "enabled"}},
        )

    assert captured_calls[-1]["additionalModelRequestFields"] == {
        "reasoning_config": {"type": "enabled"}
    }


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_empty_extra_fields_is_dropped(monkeypatch):
    monkeypatch.delenv("AWS_REGION", raising=False)
    captured_calls: list[dict] = []

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeSession(captured_calls, []),
    ):
        await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            extra_fields=None,
        )
        await bedrock_complete_if_cache(
            model="bedrock-model",
            prompt="hello",
            extra_fields={},
        )

    for call in captured_calls:
        assert "additionalModelRequestFields" not in call


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_api_key_is_ignored_and_does_not_mutate_env(monkeypatch):
    monkeypatch.delenv("AWS_REGION", raising=False)
    monkeypatch.setenv("AWS_BEARER_TOKEN_BEDROCK", "absk-from-env")
    monkeypatch.delenv("AWS_ACCESS_KEY_ID", raising=False)
    monkeypatch.delenv("AWS_SECRET_ACCESS_KEY", raising=False)
    monkeypatch.delenv("AWS_SESSION_TOKEN", raising=False)

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeSession([], []),
    ):
        with pytest.warns(DeprecationWarning, match="api_key=.*ignored"):
            await bedrock_complete_if_cache(
                model="bedrock-model",
                prompt="hello",
                api_key="absk-should-be-ignored",
                aws_access_key_id="akid",
                aws_secret_access_key="secret",
                aws_session_token="session",
            )

    assert os.environ.get("AWS_BEARER_TOKEN_BEDROCK") == "absk-from-env"
    assert os.environ.get("AWS_ACCESS_KEY_ID") is None
    assert os.environ.get("AWS_SECRET_ACCESS_KEY") is None
    assert os.environ.get("AWS_SESSION_TOKEN") is None


@pytest.mark.offline
@pytest.mark.asyncio
async def test_bedrock_embed_forwards_sigv4_and_ignores_api_key(monkeypatch):
    monkeypatch.delenv("AWS_REGION", raising=False)
    monkeypatch.delenv("AWS_BEARER_TOKEN_BEDROCK", raising=False)
    client_kwargs_calls: list[dict] = []

    with patch(
        "lightrag.llm.bedrock.aioboto3.Session",
        return_value=_FakeEmbeddingSession([], client_kwargs_calls),
    ):
        with pytest.warns(DeprecationWarning, match="api_key=.*ignored"):
            await bedrock_embed(
                texts=["hello"],
                api_key="absk-embedding-key",
                aws_region="us-east-1",
                aws_access_key_id="akid",
                aws_secret_access_key="secret",
                aws_session_token="session",
            )

    assert client_kwargs_calls[-1] == {
        "region_name": "us-east-1",
        "aws_access_key_id": "akid",
        "aws_secret_access_key": "secret",
        "aws_session_token": "session",
    }
    assert os.environ.get("AWS_BEARER_TOKEN_BEDROCK") is None


@pytest.mark.offline
def test_bedrock_auth_docstrings_describe_generic_api_key_behavior():
    assert "AWS_BEARER_TOKEN_BEDROCK" in bedrock_complete_if_cache.__doc__
    assert "LLM_BINDING_API_KEY" in bedrock_complete_if_cache.__doc__
    assert "EMBEDDING_BINDING_API_KEY" in bedrock_embed.func.__doc__


class _FakeLightRAG:
    last_init_kwargs = None
    last_instance = None

    def __init__(self, **kwargs):
        type(self).last_init_kwargs = dict(kwargs)
        type(self).last_instance = self
        self.role_config_snapshot = {}
        for role, cfg in (kwargs.get("role_llm_configs") or {}).items():
            metadata = dict(getattr(cfg, "metadata", None) or {})
            self.role_config_snapshot[role] = {
                "binding": metadata.get("binding"),
                "model": metadata.get("model"),
                "host": metadata.get("host"),
                "is_cross_provider": metadata.get("is_cross_provider", False),
                "max_async": getattr(cfg, "max_async", None),
                "timeout": getattr(cfg, "timeout", None),
                "has_model_kwargs": getattr(cfg, "kwargs", None) is not None,
                "metadata": metadata,
            }
        self.queue_status_snapshot = {}
        self.embedding_queue_status_snapshot = {}
        self.rerank_queue_status_snapshot = {}

    def register_role_llm_builder(self, _builder) -> None:
        return None

    def set_role_llm_metadata(self, _role: str, **_metadata) -> None:
        return None

    def get_llm_role_config(self):
        return self.role_config_snapshot

    async def get_llm_queue_status(self, include_base=True):
        return self.queue_status_snapshot

    async def get_embedding_queue_status(self):
        return self.embedding_queue_status_snapshot

    async def get_rerank_queue_status(self):
        return self.rerank_queue_status_snapshot


class _FakeOllamaAPI:
    def __init__(self, *_args, **_kwargs):
        self.router = APIRouter()


def _make_args(tmp_path) -> SimpleNamespace:
    return SimpleNamespace(
        host="127.0.0.1",
        port=9621,
        log_level="INFO",
        verbose=False,
        cors_origins="*",
        whitelist_paths="/health,/api/*",
        auth_accounts="",
        token_secret=None,
        token_expire_hours=48,
        guest_token_expire_hours=24,
        jwt_algorithm="HS256",
        token_auto_renew=True,
        token_renew_threshold=0.5,
        llm_binding="bedrock",
        embedding_binding="bedrock",
        llm_binding_host="DEFAULT_BEDROCK_ENDPOINT",
        embedding_binding_host="DEFAULT_BEDROCK_ENDPOINT",
        ssl=False,
        ssl_certfile=None,
        ssl_keyfile=None,
        key=None,
        input_dir=str(tmp_path / "inputs"),
        workspace="",
        working_dir=str(tmp_path / "rag_storage"),
        llm_binding_api_key=None,
        embedding_binding_api_key="",
        aws_region="us-east-1",
        aws_access_key_id="global-akid",
        aws_secret_access_key="global-secret",
        aws_session_token="global-session",
        query_aws_region=None,
        query_aws_access_key_id=None,
        query_aws_secret_access_key=None,
        query_aws_session_token=None,
        llm_model="us.amazon.nova-lite-v1:0",
        embedding_model=None,
        embedding_dim=None,
        embedding_send_dim=False,
        embedding_token_limit=None,
        embedding_document_prefix=None,
        embedding_document_prefix_configured=False,
        embedding_query_prefix=None,
        embedding_query_prefix_configured=False,
        embedding_prefix_no_prefix_sentinel="NO_PREFIX",
        embedding_prefixes_configured=False,
        embedding_asymmetric=False,
        embedding_asymmetric_configured=False,
        max_async=4,
        summary_max_tokens=512,
        summary_context_size=4096,
        force_llm_summary_on_merge=8,
        chunk_size=1200,
        chunk_overlap_size=100,
        kv_storage="JsonKVStorage",
        graph_storage="NetworkXStorage",
        vector_storage="NanoVectorDBStorage",
        doc_status_storage="JsonDocStatusStorage",
        cosine_threshold=0.2,
        enable_llm_cache_for_extract=True,
        enable_llm_cache=True,
        vlm_process_enable=False,
        max_parallel_insert=2,
        max_graph_nodes=1000,
        simulated_model_name="lightrag",
        simulated_model_tag="latest",
        summary_language="English",
        rerank_binding="null",
        rerank_model=None,
        rerank_binding_host=None,
        rerank_binding_api_key=None,
        embedding_func_max_async=8,
        embedding_batch_num=10,
        min_rerank_score=0.0,
        related_chunk_number=5,
        top_k=10,
        llm_timeout=180,
        embedding_timeout=30,
        rerank_max_async=4,
        rerank_timeout=30,
    )


@pytest.mark.offline
@pytest.mark.asyncio
async def test_create_app_query_role_uses_bedrock_binding(tmp_path, monkeypatch):
    _reload_api_modules_if_mocked()
    monkeypatch.setattr(sys, "argv", ["pytest"])
    config = importlib.import_module("lightrag.api.config")
    config.initialize_config(_make_args(tmp_path), force=True)
    lightrag_server = importlib.import_module("lightrag.api.lightrag_server")
    monkeypatch.setattr(lightrag_server, "LightRAG", _FakeLightRAG)
    monkeypatch.setattr(lightrag_server, "check_frontend_build", lambda: (True, False))
    monkeypatch.setattr(
        lightrag_server, "create_document_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_query_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_graph_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(lightrag_server, "OllamaAPI", _FakeOllamaAPI)

    args = _make_args(tmp_path)

    with (
        patch(
            "lightrag.llm.bedrock.bedrock_complete_if_cache",
            AsyncMock(return_value="bedrock-ok"),
        ) as mocked_bedrock,
        patch(
            "lightrag.llm.openai.openai_complete_if_cache",
            AsyncMock(side_effect=AssertionError("OpenAI fallback should not be used")),
        ) as mocked_openai,
    ):
        lightrag_server.create_app(args)
        query_cfg = _FakeLightRAG.last_init_kwargs["role_llm_configs"]["query"]
        query_func = query_cfg.func
        result = await query_func("hello")

    assert query_cfg.metadata["binding"] == "bedrock"
    assert query_cfg.metadata["model"] == "us.amazon.nova-lite-v1:0"
    assert query_cfg.metadata["host"] == "DEFAULT_BEDROCK_ENDPOINT"
    assert query_cfg.metadata["api_key"] is None
    assert query_cfg.metadata["bedrock_aws_options"]["aws_region"] == "us-east-1"
    assert result == "bedrock-ok"
    assert mocked_openai.await_count == 0
    assert mocked_bedrock.await_count == 1
    assert mocked_bedrock.await_args.args[:2] == ("us.amazon.nova-lite-v1:0", "hello")
    assert "api_key" not in mocked_bedrock.await_args.kwargs
    assert (
        mocked_bedrock.await_args.kwargs["endpoint_url"] == "DEFAULT_BEDROCK_ENDPOINT"
    )
    assert mocked_bedrock.await_args.kwargs["aws_region"] == "us-east-1"
    assert mocked_bedrock.await_args.kwargs["aws_access_key_id"] == "global-akid"


@pytest.mark.offline
@pytest.mark.asyncio
async def test_create_app_bedrock_query_role_uses_role_sigv4_credentials(
    tmp_path, monkeypatch
):
    _reload_api_modules_if_mocked()
    monkeypatch.setattr(sys, "argv", ["pytest"])
    config = importlib.import_module("lightrag.api.config")
    config.initialize_config(_make_args(tmp_path), force=True)
    lightrag_server = importlib.import_module("lightrag.api.lightrag_server")
    monkeypatch.setattr(lightrag_server, "LightRAG", _FakeLightRAG)
    monkeypatch.setattr(lightrag_server, "check_frontend_build", lambda: (True, False))
    monkeypatch.setattr(
        lightrag_server, "create_document_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_query_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_graph_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(lightrag_server, "OllamaAPI", _FakeOllamaAPI)

    args = _make_args(tmp_path)
    args.query_aws_region = "us-west-2"
    args.query_aws_access_key_id = "query-akid"
    args.query_aws_secret_access_key = "query-secret"
    args.query_aws_session_token = "query-session"

    with patch(
        "lightrag.llm.bedrock.bedrock_complete_if_cache",
        AsyncMock(return_value="bedrock-ok"),
    ) as mocked_bedrock:
        lightrag_server.create_app(args)
        query_func = _FakeLightRAG.last_init_kwargs["role_llm_configs"]["query"].func
        await query_func("hello")

    assert mocked_bedrock.await_args.kwargs["aws_region"] == "us-west-2"
    assert mocked_bedrock.await_args.kwargs["aws_access_key_id"] == "query-akid"
    assert mocked_bedrock.await_args.kwargs["aws_secret_access_key"] == "query-secret"
    assert mocked_bedrock.await_args.kwargs["aws_session_token"] == "query-session"


@pytest.mark.offline
@pytest.mark.asyncio
async def test_create_app_keyword_openai_role_forwards_nested_extra_body(
    tmp_path, monkeypatch, caplog
):
    _reload_api_modules_if_mocked()
    monkeypatch.setattr(sys, "argv", ["pytest"])
    monkeypatch.setattr(logging.getLogger("lightrag"), "propagate", True)
    monkeypatch.setenv(
        "KEYWORD_OPENAI_LLM_EXTRA_BODY",
        '{"chat_template_kwargs": {"enable_thinking": false}}',
    )

    config = importlib.import_module("lightrag.api.config")
    config.initialize_config(_make_args(tmp_path), force=True)
    lightrag_server = importlib.import_module("lightrag.api.lightrag_server")
    monkeypatch.setattr(lightrag_server, "LightRAG", _FakeLightRAG)
    monkeypatch.setattr(lightrag_server, "check_frontend_build", lambda: (True, False))
    monkeypatch.setattr(
        lightrag_server, "create_document_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_query_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_graph_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(lightrag_server, "OllamaAPI", _FakeOllamaAPI)

    args = _make_args(tmp_path)
    args.keyword_llm_binding = "openai"
    args.keyword_llm_model = "xhd/Qwen3.5-35B-A3B"
    args.keyword_llm_binding_host = "https://keyword.example/v1"
    args.keyword_llm_binding_api_key = "keyword-secret"

    with (
        caplog.at_level("INFO", logger="lightrag"),
        patch(
            "lightrag.llm.openai.openai_complete_if_cache",
            AsyncMock(
                return_value='{"high_level_keywords":[],"low_level_keywords":[]}'
            ),
        ) as mocked_openai,
    ):
        lightrag_server.create_app(args)
        keyword_cfg = _FakeLightRAG.last_init_kwargs["role_llm_configs"]["keyword"]
        result = await keyword_cfg.func(
            "keyword prompt", response_format={"type": "json_object"}
        )

    assert result == '{"high_level_keywords":[],"low_level_keywords":[]}'
    assert keyword_cfg.metadata["binding"] == "openai"
    assert keyword_cfg.metadata["provider_options"]["extra_body"] == {
        "chat_template_kwargs": {"enable_thinking": False}
    }
    assert mocked_openai.await_count == 1
    assert mocked_openai.await_args.args[:2] == (
        "xhd/Qwen3.5-35B-A3B",
        "keyword prompt",
    )
    kwargs = mocked_openai.await_args.kwargs
    assert kwargs["base_url"] == "https://keyword.example/v1"
    assert kwargs["api_key"] == "keyword-secret"
    assert kwargs["response_format"] == {"type": "json_object"}
    assert kwargs["extra_body"] == {"chat_template_kwargs": {"enable_thinking": False}}

    messages = "\n".join(record.getMessage() for record in caplog.records)
    assert "Role LLM Option:" in messages
    assert " - extract: Bedrock {}" in messages
    assert " - keyword: OpenAI {'extra_body':" in messages
    assert " - query: Bedrock {}" in messages
    assert " - vlm: Bedrock {}" in messages
    assert "chat_template_kwargs" in messages
    assert "reasoning_effort" not in messages
    assert "frequency_penalty" not in messages
    assert "keyword-secret" not in messages


@pytest.mark.offline
def test_create_app_rejects_bedrock_role_api_key(tmp_path, monkeypatch):
    _reload_api_modules_if_mocked()
    monkeypatch.setattr(sys, "argv", ["pytest"])
    config = importlib.import_module("lightrag.api.config")
    config.initialize_config(_make_args(tmp_path), force=True)
    lightrag_server = importlib.import_module("lightrag.api.lightrag_server")
    monkeypatch.setattr(lightrag_server, "check_frontend_build", lambda: (True, False))

    args = _make_args(tmp_path)
    args.query_llm_binding_api_key = "absk-role"

    with pytest.raises(ValueError, match="does not support role-specific"):
        lightrag_server.create_app(args)


@pytest.mark.offline
def test_health_role_llm_config_uses_runtime_snapshot(tmp_path, monkeypatch):
    _reload_api_modules_if_mocked()
    monkeypatch.setattr(sys, "argv", ["pytest"])
    config = importlib.import_module("lightrag.api.config")
    config.initialize_config(_make_args(tmp_path), force=True)
    lightrag_server = importlib.import_module("lightrag.api.lightrag_server")
    monkeypatch.setattr(lightrag_server, "LightRAG", _FakeLightRAG)
    monkeypatch.setattr(lightrag_server, "check_frontend_build", lambda: (True, False))
    monkeypatch.setattr(
        lightrag_server, "create_document_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_query_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_graph_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(lightrag_server, "OllamaAPI", _FakeOllamaAPI)
    monkeypatch.setattr(
        lightrag_server,
        "get_namespace_data",
        AsyncMock(return_value={"busy": False}),
    )
    monkeypatch.setattr(lightrag_server, "get_default_workspace", lambda: "default")
    monkeypatch.setattr(
        lightrag_server,
        "cleanup_keyed_lock",
        lambda: {"cleanup_performed": {}, "current_status": {}},
    )

    app = lightrag_server.create_app(_make_args(tmp_path))
    _FakeLightRAG.last_instance.role_config_snapshot = {
        "query": {
            "binding": "runtime-binding",
            "model": "runtime-model",
            "host": "https://runtime.example/v1",
            "max_async": 9,
            "metadata": {"binding": "runtime-binding"},
        }
    }
    _FakeLightRAG.last_instance.queue_status_snapshot = {
        "query": {"available": True, "rejected_total": 2}
    }
    _FakeLightRAG.last_instance.embedding_queue_status_snapshot = {
        "available": True,
        "running": 1,
    }
    _FakeLightRAG.last_instance.rerank_queue_status_snapshot = {
        "available": False,
    }

    response = TestClient(app).get("/health")

    assert response.status_code == 200
    body = response.json()
    role_cfg = body["configuration"]["role_llm_config"]["query"]
    assert role_cfg["binding"] == "runtime-binding"
    assert role_cfg["model"] == "runtime-model"
    assert role_cfg["host"] == "https://runtime.example/v1"
    assert role_cfg["max_async"] == 9
    assert role_cfg["model"] != "us.amazon.nova-lite-v1:0"
    assert body["llm_queue_status"]["query"]["rejected_total"] == 2
    assert body["embedding_queue_status"]["running"] == 1
    assert body["rerank_queue_status"]["available"] is False


@pytest.mark.offline
@pytest.mark.parametrize(
    "pipeline_state, expected_active",
    [
        ({"busy": False}, False),
        ({"busy": True}, True),
        ({"busy": False, "scanning": True}, True),
        ({"busy": False, "destructive_busy": True}, True),
        ({"busy": False, "pending_enqueues": 2}, True),
        (
            {
                "busy": False,
                "scanning": False,
                "destructive_busy": False,
                "pending_enqueues": 0,
            },
            False,
        ),
    ],
)
def test_health_pipeline_active_derivation(
    tmp_path, monkeypatch, pipeline_state, expected_active
):
    _reload_api_modules_if_mocked()
    monkeypatch.setattr(sys, "argv", ["pytest"])
    config = importlib.import_module("lightrag.api.config")
    config.initialize_config(_make_args(tmp_path), force=True)
    lightrag_server = importlib.import_module("lightrag.api.lightrag_server")
    monkeypatch.setattr(lightrag_server, "LightRAG", _FakeLightRAG)
    monkeypatch.setattr(lightrag_server, "check_frontend_build", lambda: (True, False))
    monkeypatch.setattr(
        lightrag_server, "create_document_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_query_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(
        lightrag_server, "create_graph_routes", lambda *_args, **_kwargs: APIRouter()
    )
    monkeypatch.setattr(lightrag_server, "OllamaAPI", _FakeOllamaAPI)
    monkeypatch.setattr(
        lightrag_server,
        "get_namespace_data",
        AsyncMock(return_value=pipeline_state),
    )
    monkeypatch.setattr(lightrag_server, "get_default_workspace", lambda: "default")
    monkeypatch.setattr(
        lightrag_server,
        "cleanup_keyed_lock",
        lambda: {"cleanup_performed": {}, "current_status": {}},
    )

    app = lightrag_server.create_app(_make_args(tmp_path))
    response = TestClient(app).get("/health")

    assert response.status_code == 200
    body = response.json()
    assert body["pipeline_busy"] is bool(pipeline_state.get("busy", False))
    assert body["pipeline_scanning"] is bool(pipeline_state.get("scanning", False))
    assert body["pipeline_destructive_busy"] is bool(
        pipeline_state.get("destructive_busy", False)
    )
    assert body["pipeline_pending_enqueues"] == int(
        pipeline_state.get("pending_enqueues", 0)
    )
    assert body["pipeline_active"] is expected_active