266 lines
12 KiB
Python
266 lines
12 KiB
Python
import logging
|
||
from functools import lru_cache
|
||
from typing import Any, cast
|
||
|
||
from sqlalchemy.ext.asyncio import AsyncSession
|
||
|
||
from app.core.config import get_settings
|
||
from app.core.errors import RecoverableAgentError
|
||
from app.core.fund_contracts import FundQuoteQuery
|
||
from app.core.knowledge_contracts import KnowledgeSearchInput
|
||
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.vector_memory import VectorMemoryAdapter
|
||
from app.service.agent.factory import AgentFactory
|
||
from app.service.agent.governance import PlatformGovernance
|
||
from app.service.agent.implementations.customer_service import CustomerServiceAgent
|
||
from app.service.agent.implementations.fund_query_demo import FundQueryDemoAgent
|
||
from app.service.agent.implementations.risk_agent import RiskAgent
|
||
from app.service.fund_quote_service import query_fund_quote_tool
|
||
from app.service.intent_classifier import IntentClassifier
|
||
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.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。"""
|
||
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, settings.milvus_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()))
|
||
|
||
|
||
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()
|
||
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="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,
|
||
))
|
||
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"),
|
||
))
|
||
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),
|
||
)
|