group_fqcd_jr/app/service/agent/bootstrap.py

import logging
from functools import lru_cache
from typing import Any, cast

from sqlalchemy.ext.asyncio import AsyncSession

from app.core.advisor_allocation_contracts import AssetAllocationQuery
from app.core.config import get_settings
from app.core.errors import RecoverableAgentError
from app.core.fund_contracts import FundQuoteQuery
from app.core.investment_goal_contracts import InvestmentGoalQuery
from app.core.knowledge_contracts import KnowledgeSearchInput
from app.core.nl2sql_contracts import FinancialNL2SQLInput
from app.core.portfolio_analysis_contracts import PortfolioAnalysisQuery
from app.core.product_comparison_contracts import ProductComparisonQuery
from app.core.product_recommendation_contracts import ProductRecommendationQuery
from app.core.risk_contracts import RiskAlertEvidenceQuery, RiskAlertQuery
from app.infrastructure.fund_quote_cache import FundQuoteCache
from app.infrastructure.graph import build_graph_driver
from app.infrastructure.memory_cache import MemoryCacheAdapter
from app.infrastructure.milvus_knowledge_writer import MilvusKnowledgeWriter
from app.infrastructure.milvus_profile_projection import PROFILE_COLLECTION
from app.infrastructure.milvus_profile_vector_client import MilvusProfileVectorClient
from app.infrastructure.vector_memory import VectorMemoryAdapter
from app.service.agent.factory import AgentFactory
from app.service.agent.governance import PlatformGovernance
from app.service.agent.implementations.advisor import AdvisorAgent
from app.service.agent.implementations.customer_service import CustomerServiceAgent
from app.service.agent.implementations.financial_nl2sql import FinancialNL2SQLAgent
from app.service.agent.implementations.fund_query_demo import FundQueryDemoAgent
from app.service.agent.implementations.platform_probe import (
    PROBE_ALT_TOOL,
    PROBE_PERMISSION,
    PROBE_TOOL,
    PlatformProbeAgent,
    ProbeEchoArgs,
    probe_alt_tool,
    probe_echo_tool,
)
from app.service.agent.implementations.risk_agent import RiskAgent
from app.service.agent.offsite_fund_agent import OffsiteFundAgent
from app.service.agent.promotion_material_agent import PromotionMaterialAgent
from app.service.asset_allocation_service import asset_allocation_tool
from app.service.customer_profile_service import (
    CustomerProfileQuery,
    query_customer_profile_tool,
)
from app.service.financial_nl2sql_service import query_financial_data_tool
from app.service.fund_quote_service import query_fund_quote_tool
from app.service.intent_classifier import IntentClassifier
from app.service.investment_goal_service import investment_goal_query_tool
from app.service.knowledge_search_service import KnowledgeSearchService
from app.service.knowledge_tool import knowledge_search_tool
from app.service.memory_recall_service import MemoryRecallService
from app.service.model_gateway import (
    DatabaseModelEndpointResolver,
    DatabaseModelGateway,
    ModelDispatchService,
    ModelEmbeddingService,
    ModelGenerationService,
)
from app.service.portfolio_analysis_service import portfolio_analysis_tool
from app.service.product_comparison_service import product_comparison_tool
from app.service.product_recommendation_service import product_recommendation_tool
from app.service.relationship_service import RelationshipService
from app.service.risk_tools import (
    get_alert_evidence_tool,
    get_risk_overview_tool,
    search_risk_alerts_tool,
)
from app.service.runtime_config_service import load_active_intent_configs
from app.service.suitability_service import SuitabilityToolInput, suitability_tool_handler
from app.service.tool_executor import ToolDefinition, ToolExecutor, ToolRegistry

logger = logging.getLogger(__name__)


@lru_cache(maxsize=1)
def get_model_service() -> ModelGenerationService:
    """生产模型装配的唯一入口；业务 Agent 与 Worker 记忆抽取共用同一实例。"""
    return ModelGenerationService(ModelDispatchService(DatabaseModelGateway()))


@lru_cache(maxsize=1)
def get_memory_cache_adapter() -> MemoryCacheAdapter | None:
    """Redis 记忆缓存适配器；构造失败返回 None——缓存只是优化层，不得阻塞召回。"""
    try:
        from redis.asyncio import Redis

        settings = get_settings()
        client = Redis.from_url(
            settings.redis_url,
            socket_connect_timeout=settings.redis_connect_timeout_seconds,
            socket_timeout=settings.redis_connect_timeout_seconds,
            decode_responses=True,
        )
        return MemoryCacheAdapter(client)
    except Exception:
        logger.warning("memory cache adapter unavailable; recall runs without cache",
                       exc_info=True)
        return None


@lru_cache(maxsize=1)
def get_fund_quote_cache() -> FundQuoteCache | None:
    """行情短缓存：与记忆召回共用同一个 Redis 适配器。

    Redis 不可用时这里仍会返回适配器，但读写异常由 `MemoryCacheAdapter` 内部
    吞掉并以 `degraded` 语义返回，行情查询会退化为直连外部数据源而不会阻塞；
    适配器构造失败（例如缺少 redis 依赖）则返回 None，效果相同。缓存永远只是
    优化层，不得让行情查询因缓存故障失败。
    """
    adapter = get_memory_cache_adapter()
    if adapter is None:
        return None
    return FundQuoteCache(adapter)


@lru_cache(maxsize=1)
def get_vector_memory_adapter() -> VectorMemoryAdapter | None:
    """Milvus 适配器；构造失败返回 None（语义通道关闭），不影响结构化召回。

    集合名取 `PROFILE_COLLECTION`，与投影写入路径（`MilvusProfileProjection`）
    **同一个常量**。此前这里读的是 `settings.milvus_collection`（`jr_memory`，
    该集合从未创建，现已删除该配置项），于是适配器构造成功、每次 `search` 都抛异常
    并被降级成 `milvus_unavailable` —— 表现为"语义召回永远没数据"，
    而投影那条链路看起来一切正常。
    """
    try:
        from pymilvus import MilvusClient  # type: ignore[import-untyped]

        settings = get_settings()
        client = MilvusClient(uri=settings.milvus_uri, token=settings.milvus_token or None)
        return VectorMemoryAdapter(client, PROFILE_COLLECTION)
    except Exception:
        logger.warning("vector memory adapter unavailable; semantic recall disabled",
                       exc_info=True)
        return None


@lru_cache(maxsize=1)
def get_memory_embedding_service() -> ModelEmbeddingService:
    """文本向量化入口：与文本生成共用同一套端点解析与受控降级。"""
    return ModelEmbeddingService(ModelDispatchService(DatabaseModelGateway()))


@lru_cache(maxsize=1)
def get_milvus_knowledge_writer() -> MilvusKnowledgeWriter | None:
    """知识向量**写**适配器（装配入口）；`milvus_uri` 缺失时返回 None。

    与召回侧的 `get_vector_memory_adapter()` 分离：写路径不与检索进程共用客户端
    （读写物理隔离，向量库故障不能从写路径传染到问答主链路）。构造是惰性的
    （`MilvusKnowledgeWriter.__init__` 不连 Milvus），所以这里返回实例不代表连接可用；
    真连不上时在首次写入抛 `RecoverableAgentError`，由 `OutboxWorker` 退避重试/判死信。

    返回 None 的语义是**显式降级**：`WorkerRuntime` 会因此不注册
    `knowledge.vector_sync_requested` / `knowledge.vector_delete_requested` 两个 handler，
    事件在库里保持 pending（可观测、可重放），并在启动路径留一条 warning —— 绝不静默，
    也绝不伪造同步成功。
    """
    settings = get_settings()
    uri = (settings.milvus_uri or "").strip()
    if not uri:
        logger.warning(
            "milvus_uri not configured; knowledge vector writer disabled and "
            "knowledge.vector_sync_requested events will stay pending"
        )
        return None
    return MilvusKnowledgeWriter(uri, settings.milvus_token or "")


@lru_cache(maxsize=1)
def get_milvus_profile_vector_client() -> MilvusProfileVectorClient | None:
    """长期记忆画像向量**写**适配器（装配入口）；`milvus_uri` 缺失时返回 None。

    与 `get_milvus_knowledge_writer` 同一取向：读写隔离、构造惰性、缺配置显式降级。
    返回 None 的语义是**显式降级**——`WorkerRuntime` 会因此不注册画像投影 handler，
    `memory_sync_outbox` 的事件在库里保持 pending（可观测、可重放），并留一条 warning，
    绝不静默，也绝不伪造同步成功。
    """
    settings = get_settings()
    uri = (settings.milvus_uri or "").strip()
    if not uri:
        logging.getLogger(__name__).warning(
            "milvus_uri not configured; profile vector projection disabled and "
            "memory_sync_outbox events will stay pending"
        )
        return None
    return MilvusProfileVectorClient(uri, settings.milvus_token or "")


async def _embed_text(text: str) -> list[float]:
    """把文本向量化；端点来自发布配置（task_type=embedding），无端点时失败关闭。"""
    endpoints = await DatabaseModelEndpointResolver().resolve(
        agent_type="memory_recall", task_type="embedding"
    )
    if not endpoints:
        raise RecoverableAgentError("没有可用的 embedding 端点")
    execution = await get_memory_embedding_service().embed(endpoints, text)
    return execution.vector


@lru_cache(maxsize=1)
def get_knowledge_search_service() -> KnowledgeSearchService:
    """客服知识检索装配：Milvus 客户端 + 向量化入口。

    与记忆的语义通道同一取向：Milvus 不可达或缺少 embedding 端点时**不抛异常**，
    而是返回 `available=False` 的实例，检索结果标记 `degraded`，由客服 Agent 走
    「引导客户致电人工客服」。基础设施故障不该表现成客户可见的错误。
    """
    client = None
    try:
        from pymilvus import MilvusClient

        settings = get_settings()
        client = MilvusClient(uri=settings.milvus_uri, token=settings.milvus_token or None)
    except Exception:
        logger.warning("knowledge vector client unavailable; search degrades", exc_info=True)
    return KnowledgeSearchService(client, _embed_text)


@lru_cache(maxsize=1)
def get_relationship_service() -> RelationshipService | None:
    """图关系读服务：客户 → 产品／标签／事件 的多跳查询入口。

    驱动构造失败时返回 None（图能力关闭），由调用方降级——图库不可用不该阻塞主链路，
    与 Milvus 侧"语义通道缺失不影响结构化召回"是同一取向。

    注意：本服务**只读**，且关系类型受 `RelationshipService.ALLOWED_RELATIONSHIPS` 白名单约束；
    写入走 `GraphProjectionWorker`（由领域事件驱动），这里不提供任意写接口。
    """
    driver = build_graph_driver()
    if driver is None:
        return None
    return RelationshipService(driver)


def build_memory_recall_service(session: AsyncSession) -> MemoryRecallService:
    """记忆召回组装：结构化召回始终可用，Redis 缓存与语义通道可用时叠加。

    语义通道需要**两件事同时具备**：可达的 Milvus 与可用的 embedding 端点。缺少
    embedding 端点时向量化按设计失败关闭，结果标记 `embedding_failed` 并保留结构化
    召回——这是配置缺口而非功能缺失，配置端点后语义召回无需改代码即可生效。
    """
    vector = get_vector_memory_adapter()
    return MemoryRecallService(
        session,
        vector=vector,
        cache=get_memory_cache_adapter(),
        embed=_embed_text if vector is not None else None,
    )


@lru_cache(maxsize=1)
def get_agent_factory() -> AgentFactory:
    """HTTP 与 Worker 共用的唯一底座依赖组装入口。"""
    registry = ToolRegistry()
    # 基座验证探针的工具：只回显参数，用于端到端触发 ToolExecutor 的四种拒绝分支。
    # 它不查库、不写状态，注册在这里也不会被任何业务 Agent 的白名单引用。
    registry.register(ToolDefinition(
        name=PROBE_TOOL,
        input_model=ProbeEchoArgs,
        handler=cast(Any, probe_echo_tool),
        required_permission=PROBE_PERMISSION,
        allowed_roles=("admin",),
    ))
    # 第二个探针工具，两个用途：
    # ① 让"有白名单但不含该工具"的分支可被构造（见 platform_probe 的说明）；
    # ② 让**工具层**的"角色不能使用工具"分支可被构造 —— 它的角色集合**比 Agent 的更窄**
    #    （Agent 允许 admin，本工具只允许 risk_operator）。必须更窄才行：Agent 层的
    #    validate_access 会先按 AgentDefinition.allowed_roles 拦截，两者一致时永远进不到
    #    工具层的角色校验。
    registry.register(ToolDefinition(
        name=PROBE_ALT_TOOL,
        input_model=ProbeEchoArgs,
        handler=cast(Any, probe_alt_tool),
        required_permission=PROBE_PERMISSION,
        allowed_roles=("risk_operator",),
    ))
    registry.register(ToolDefinition(
        name="check_suitability",
        input_model=SuitabilityToolInput,
        handler=cast(Any, suitability_tool_handler),
        required_permission="suitability:read",
        allowed_roles=("customer", "advisor", "operator", "admin"),
    ))
    registry.register(ToolDefinition(
        name="query_fund_quote",
        input_model=FundQuoteQuery,
        handler=cast(Any, query_fund_quote_tool),
        required_permission="fund:quote:read",
        allowed_roles=("customer", "advisor", "operator", "risk_operator", "admin"),
        # C3 数值依据：EastmoneyAdapterFactory 的单代码最坏预算为
        # 4.0s(单次) + 0.2s(退避) + 4.0s(重试) = 8.2s，且 FundQuoteService
        # 已按代码并发，总耗时与代码数量无关。15s ≈ 8.2s × 1.8，余量覆盖
        # DNS/TLS 建连与事件循环调度开销。原值 5s 小于适配器默认单次超时
        # （12s）与重试预算，多代码查询必然先撞工具超时。
        timeout_seconds=15,
    ))
    registry.register(ToolDefinition(
        name="query_customer_profile",
        input_model=CustomerProfileQuery,
        handler=cast(Any, query_customer_profile_tool),
        # 复用既有权限码：`memory:read:self` 表示"只读本人画像"，客服 Agent 用它在
        # 确定性识别出画像问题后取权威字段；读他人画像需要 `memory:read:customer`，
        # 由服务内部按数据范围二次校验（`self` / `own_customers` / `all`）。
        required_permission="memory:read:self",
        allowed_roles=("customer", "advisor", "operator", "risk_operator", "admin"),
    ))
    registry.register(ToolDefinition(
        name="query_financial_data",
        input_model=FinancialNL2SQLInput,
        handler=cast(Any, query_financial_data_tool),
        required_permission="financial:nl2sql:read",
        allowed_roles=("advisor", "operator", "admin", "super_admin"),
        timeout_seconds=10,
    ))
    registry.register(ToolDefinition(
        name="search_knowledge",
        input_model=KnowledgeSearchInput,
        handler=cast(Any, knowledge_search_tool),
        # 复用既有权限码（customer 角色已具备），不新增权限点
        required_permission="knowledge:reference:read",
        allowed_roles=("customer", "advisor", "operator", "admin"),
        # 检索含一次 embedding 调用 + 三次集合检索；embedding 实测 0.42s，
        # 10s 覆盖冷启动与 Milvus 抖动，又不至于让客户等太久
        timeout_seconds=10,
    ))
    # 兼容一期发布配置和旧客户端的工具名；实现仍复用同一个只读检索处理器。
    registry.register(ToolDefinition(
        name="query_knowledge",
        input_model=KnowledgeSearchInput,
        handler=cast(Any, knowledge_search_tool),
        required_permission="knowledge:query",
        allowed_roles=("visitor", "customer"),
        timeout_seconds=10,
    ))
    registry.register(ToolDefinition(
        name="search_risk_alerts",
        input_model=RiskAlertQuery,
        handler=cast(Any, search_risk_alerts_tool),
        required_permission="risk:alert:read",
        allowed_roles=("risk_operator", "admin"),
    ))
    registry.register(ToolDefinition(
        name="get_risk_overview",
        input_model=RiskAlertQuery,
        handler=cast(Any, get_risk_overview_tool),
        required_permission="risk:alert:read",
        allowed_roles=("risk_operator", "admin"),
    ))
    registry.register(ToolDefinition(
        name="get_alert_evidence",
        input_model=RiskAlertEvidenceQuery,
        handler=cast(Any, get_alert_evidence_tool),
        required_permission="risk:alert:read",
        allowed_roles=("risk_operator", "admin"),
    ))
    registry.register(ToolDefinition(
        name="query_investment_goal",
        input_model=InvestmentGoalQuery,
        handler=cast(Any, investment_goal_query_tool),
        required_permission="investment-goal:read:self",
        allowed_roles=("customer", "advisor", "operator", "admin"),
    ))
    registry.register(ToolDefinition(
        name="analyze_portfolio",
        input_model=PortfolioAnalysisQuery,
        handler=cast(Any, portfolio_analysis_tool),
        required_permission="portfolio-analysis:read:self",
        allowed_roles=("customer", "advisor", "operator", "admin"),
        timeout_seconds=10,
    ))
    registry.register(ToolDefinition(
        name="generate_asset_allocation",
        input_model=AssetAllocationQuery,
        handler=cast(Any, asset_allocation_tool),
        required_permission="asset-allocation:generate:self",
        allowed_roles=("customer", "advisor", "operator", "admin"),
        timeout_seconds=15,
    ))
    registry.register(ToolDefinition(
        name="recommend_products",
        input_model=ProductRecommendationQuery,
        handler=cast(Any, product_recommendation_tool),
        required_permission="product-recommendation:generate:self",
        allowed_roles=("customer", "advisor", "operator", "admin"),
        timeout_seconds=15,
    ))
    registry.register(ToolDefinition(
        name="compare_products",
        input_model=ProductComparisonQuery,
        handler=cast(Any, product_comparison_tool),
        required_permission="product-comparison:read:self",
        allowed_roles=("customer", "advisor", "operator", "admin"),
        timeout_seconds=10,
    ))

    model_service = get_model_service()
    endpoint_resolver = DatabaseModelEndpointResolver()
    factory = AgentFactory(
        # 记忆召回接入统一治理链：Agent 的 recall_memory 走组合召回服务，
        # 而不是每个 Agent 自行决定召回方式。
        governance=PlatformGovernance(recall_factory=build_memory_recall_service),
        model_service=model_service,
        tool_executor=ToolExecutor(registry),
        intent_classifier=IntentClassifier(
            model_service, config_loader=load_active_intent_configs
        ),
        intent_endpoint_resolver=endpoint_resolver,
    )
    register_business_agents(factory)
    return factory


def register_business_agents(factory: AgentFactory) -> None:
    """业务 Agent 的统一注册入口：组员在这里登记自己的一行 `factory.register(...)`。

    HTTP 服务与 Worker 共用 `get_agent_factory()` 返回的同一个工厂，因此这里注册
    一次即可在两个入口生效。注册只声明"代码允许什么"：真正能调用哪些工具，还要
    看当前 active 的 `config_release` 里为该 `agent_type:intent` 发布的工具白名单
    （两者取交集，缺发布配置时白名单为空、工具失败关闭）。
    """
    factory.register(
        FundQueryDemoAgent.definition,
        lambda _context: FundQueryDemoAgent(FundQueryDemoAgent.definition),
    )
    # 客服 Agent：只回答能溯源到公司资料的问题，答不了引导客户致电人工客服。
    # 它声明了 5 个意图，但真正能调用 search_knowledge 的范围由发布配置逐意图收窄
    # （`agent_tools` 里 `customer_service:<intent>`），未发布的意图工具失败关闭。
    factory.register(
        CustomerServiceAgent.definition,
        lambda _context: CustomerServiceAgent(CustomerServiceAgent.definition),
    )
    factory.register(
        RiskAgent.definition,
        lambda _context: RiskAgent(RiskAgent.definition),
    )
    factory.register(
        OffsiteFundAgent.definition,
        lambda _context: OffsiteFundAgent(OffsiteFundAgent.definition),
    )
    factory.register(
        PromotionMaterialAgent.definition,
        lambda _context: PromotionMaterialAgent(PromotionMaterialAgent.definition),
    )
    # 基座验证探针：只读、无副作用，用于端到端验证工具链路的拒绝行为。
    # 角色限定 admin，业务上不对外暴露用途。
    factory.register(
        PlatformProbeAgent.definition,
        lambda _context: PlatformProbeAgent(PlatformProbeAgent.definition),
    )
    factory.register(
        AdvisorAgent.definition,
        lambda _context: AdvisorAgent(AdvisorAgent.definition),
    )
    factory.register(
        FinancialNL2SQLAgent.definition,
        lambda _context: FinancialNL2SQLAgent(FinancialNL2SQLAgent.definition),
    )