249 lines
14 KiB
Python
249 lines
14 KiB
Python
import logging
|
||
import re
|
||
from collections.abc import Callable
|
||
from typing import Protocol
|
||
|
||
from sqlalchemy import select, text
|
||
from sqlalchemy.ext.asyncio import AsyncSession
|
||
|
||
from app.core.compliance_context import first_violation as _first_violation
|
||
from app.core.contracts import (
|
||
AgentDefinition,
|
||
AgentResult,
|
||
RecalledMemory,
|
||
RequestContext,
|
||
ResolvedAgentConfig,
|
||
)
|
||
from app.core.errors import ForbiddenAgentError, RecoverableAgentError
|
||
from app.infrastructure.db import SessionFactory
|
||
from app.model.configuration import ConfigRelease
|
||
from app.service.memory_recall_service import MemoryRecallService
|
||
from app.service.runtime_config_service import RuntimeConfigService
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
# 组装层注入的召回服务工厂：允许带 Redis 缓存与向量库，缺省退化为纯结构化召回。
|
||
RecallServiceFactory = Callable[[AsyncSession], MemoryRecallService]
|
||
|
||
#: 固定免责声明在 `agent_reply_template` 里的 `template_code`（Task 1 已种入并审核通过）。
|
||
#: 按 `template_code` 解析而不是按 `scene`：唯一键 `uk_reply_template_active_one` 保证的是
|
||
#: "同一 template_code 只有一条 active"，`scene` 上没有唯一性——按 scene 取会在同场景多模板时
|
||
#: 取到哪条不确定，而免责声明是必须逐字固定的合规话术。
|
||
DISCLAIMER_TEMPLATE_CODE = "TPL_DISCLAIMER"
|
||
|
||
#: 平台对外客服热线：**公开业务号码**，必须原样出现在面向客户的回复里
|
||
#: （spec B §5 要求固定联系方式）。脱敏规则专门放行它，
|
||
#: 否则 `TPL_TRANSFER_HUMAN` 与安全路由的 P2 话术会被打成 "[手机号已脱敏]"，
|
||
#: 客户拿不到联系方式（已实测复现）。与 `app/core/customer_service_rules.CONTACT_PHONE`
|
||
#: 是同一个号码，改一处必须同步另一处。
|
||
CUSTOMER_SERVICE_HOTLINE = "15936583816"
|
||
|
||
#: 代码兜底免责声明，逐字等于 `TPL_DISCLAIMER` 的已审文案。
|
||
#: 免责声明是安全关键路径（上线门禁 F5：面向客户输出 100% 附固定话术），所以**不允许**
|
||
#: "库里查不到就没有话术"这条路径——取数失败或话术缺失时用本常量，宁可多一次失败的查询，
|
||
#: 也不能少一句话术。文案本身不得命中任何禁用字面（`agent_negative_word` 规则与下方
|
||
#: `hard_patterns`）：治理层是朴素子串匹配、没有否定式豁免，写过「非保本」的合规话术会被
|
||
#: 自己的规则拦下（Task 1 已修复过一次的"自绊"）。
|
||
FALLBACK_DISCLAIMER = (
|
||
"本内容仅为投资分析参考，不构成任何直接投资建议，不构成对任何产品的收益承诺，"
|
||
"据此操作风险自负，请谨慎对待。"
|
||
)
|
||
|
||
|
||
class AgentGovernance(Protocol):
|
||
async def resolve(
|
||
self, definition: AgentDefinition, context: RequestContext
|
||
) -> ResolvedAgentConfig: ...
|
||
|
||
async def recall(self, context: RequestContext) -> tuple[RecalledMemory, ...]: ...
|
||
|
||
async def review(
|
||
self, result: AgentResult, context: RequestContext, config: ResolvedAgentConfig,
|
||
memories: tuple[RecalledMemory, ...],
|
||
) -> AgentResult: ...
|
||
|
||
|
||
class PlatformGovernance:
|
||
"""Request-scoped snapshots; no AsyncSession or customer state retained on the singleton."""
|
||
|
||
def __init__(self, recall_factory: RecallServiceFactory | None = None) -> None:
|
||
# 组装层可注入带缓存/向量库的召回服务；缺省退化为纯结构化召回，
|
||
# 依赖缺失或故障只降级，不让记忆功能整体不可用。
|
||
self._recall_factory = recall_factory
|
||
|
||
async def resolve(
|
||
self, definition: AgentDefinition, context: RequestContext
|
||
) -> ResolvedAgentConfig:
|
||
async with SessionFactory() as session:
|
||
release = await session.scalar(select(ConfigRelease).where(
|
||
ConfigRelease.status == "active"))
|
||
tools: dict[str, tuple[str, ...]] = {}
|
||
if release is not None:
|
||
service = RuntimeConfigService(session)
|
||
for intent in definition.supported_intents:
|
||
configured = await service.allowed_tools(
|
||
release.id, definition.agent_type, intent)
|
||
tools[intent] = tuple(sorted(set(configured) & set(definition.allowed_tools)))
|
||
rows = (await session.execute(text("""
|
||
SELECT match_type, word_pattern, applicable_agents FROM agent_negative_word
|
||
WHERE status='active' AND reviewer_id IS NOT NULL AND reviewed_at IS NOT NULL
|
||
"""))).mappings().all()
|
||
rules: list[tuple[str, str]] = []
|
||
import json
|
||
|
||
for row in rows:
|
||
agents = row["applicable_agents"]
|
||
if isinstance(agents, str):
|
||
agents = json.loads(agents)
|
||
if not agents or definition.agent_type not in agents:
|
||
# 失败关闭：`applicable_agents` 为空数组/NULL 表示"不适用任何 Agent"，
|
||
# 跳过该规则。
|
||
# 旧判据 `if agents and ...` 在空取值时整条 `continue` 不执行，规则因此**溢出到
|
||
# 所有 Agent**（实测 advisor 也加载到 11 条）：配置为空不是失效，而是扩大适用
|
||
# 范围，与项目一贯的失败关闭原则相反。
|
||
continue
|
||
if row["match_type"] not in {"exact", "contains"}:
|
||
# Python re has no execution timeout: never run unchecked admin regex.
|
||
raise RecoverableAgentError("禁止表达正则规则需要受限匹配器")
|
||
rules.append((str(row["match_type"]), str(row["word_pattern"])))
|
||
return ResolvedAgentConfig(
|
||
config_version=release.release_no if release else f"code:{definition.version}",
|
||
release_id=release.id if release else None,
|
||
prompt_version="none", model_endpoint="", allowed_tools=(),
|
||
allowed_tools_by_intent=tools, negative_rules=tuple(rules),
|
||
)
|
||
|
||
async def recall(self, context: RequestContext) -> tuple[RecalledMemory, ...]:
|
||
async with SessionFactory() as session:
|
||
service = (
|
||
self._recall_factory(session)
|
||
if self._recall_factory is not None
|
||
else MemoryRecallService(session)
|
||
)
|
||
customer_id = int(context.user_id)
|
||
result = await service.recall(customer_id)
|
||
if result.degraded:
|
||
logger.warning("memory recall degraded customer_id=%s reasons=%s",
|
||
customer_id, ",".join(result.degraded_reasons))
|
||
return tuple(
|
||
RecalledMemory(memory_uuid=item.memory_uuid, customer_id=str(customer_id),
|
||
content=item.content)
|
||
for item in result.items
|
||
)
|
||
|
||
async def review(
|
||
self, result: AgentResult, context: RequestContext, config: ResolvedAgentConfig,
|
||
memories: tuple[RecalledMemory, ...],
|
||
) -> AgentResult:
|
||
# 裁定 1：读库是异步的，由本方法（异步层）做；`review_output` 保持同步、不接触数据库，
|
||
# 只把拿到的文本追加到输出末尾。
|
||
disclaimer = await _load_template_text(DISCLAIMER_TEMPLATE_CODE)
|
||
return review_output(result, context, config, memories, disclaimer=disclaimer)
|
||
|
||
|
||
async def _load_template_text(template_code: str) -> str | None:
|
||
"""按 `template_code` 读取生效话术；查不到或取数失败返回 None，由调用方退回代码兜底。
|
||
|
||
取数条件与 `resolve()` 的规则查询同口径（`status='active'` 且审核字段非空），
|
||
避免"库里改了状态但代码照用"的偏差。异常一律吞掉并记日志：免责声明必须 100% 注入，
|
||
数据库故障时"少一句话术"比"整条回复失败"更糟，退回 `FALLBACK_DISCLAIMER` 即可满足门禁。
|
||
"""
|
||
try:
|
||
async with SessionFactory() as session:
|
||
content: str | None = await session.scalar(
|
||
text("""
|
||
SELECT content_text FROM agent_reply_template
|
||
WHERE template_code = :code AND status='active'
|
||
AND reviewer_id IS NOT NULL AND reviewed_at IS NOT NULL
|
||
ORDER BY version DESC LIMIT 1
|
||
"""),
|
||
{"code": template_code},
|
||
)
|
||
except Exception:
|
||
logger.warning("免责声明话术查询失败，退回代码兜底 template_code=%s",
|
||
template_code, exc_info=True)
|
||
return None
|
||
if content is None or not str(content).strip():
|
||
logger.warning("免责声明话术不存在或为空，退回代码兜底 template_code=%s", template_code)
|
||
return None
|
||
return str(content)
|
||
|
||
|
||
def review_output(
|
||
result: AgentResult, context: RequestContext, config: ResolvedAgentConfig,
|
||
memories: tuple[RecalledMemory, ...],
|
||
*,
|
||
disclaimer: str | None = None,
|
||
) -> AgentResult:
|
||
"""同步治理：引用校验 → 负面词判定/替换 → 脱敏 → 追加固定免责声明。
|
||
|
||
`disclaimer` 由异步层（`PlatformGovernance.review`）注入库内话术；本函数是同步的、
|
||
不接触数据库。默认 `None` 表示"调用方未提供"，此时用 `FALLBACK_DISCLAIMER` 兜底——
|
||
因此既有调用点不必改签名也能拿到固定话术（门禁 F5 的 100% 覆盖）。
|
||
"""
|
||
content = result.result
|
||
issued_tools = {f"{context.trace_id}:{record.tool_name}" for record in content.tool_calls
|
||
if record.status == "succeeded"}
|
||
known = {memory.memory_uuid for memory in memories if memory.customer_id == context.user_id}
|
||
for reference in content.source_references:
|
||
valid = ((reference.source_type == "memory" and reference.source_id in known)
|
||
or (reference.source_type == "tool" and reference.source_id in issued_tools))
|
||
if not valid:
|
||
raise ForbiddenAgentError("引用未来自本次已授权召回结果")
|
||
output = content.text
|
||
hard_patterns = ("保证收益", "稳赚不赔", "保本保收益", "已为您下单", "已替您交易")
|
||
# 语境豁免（详见 `app/core/compliance_context.py`）：零容忍规则是朴素子串匹配，
|
||
# 无法区分"作出承诺"与"禁止承诺/谈论该表述"。政策类问答会被误伤 —— 实测
|
||
# 「基金销售有哪些合规要求」的答案因含"严禁承诺保本保收益""禁止使用『保证收益』"
|
||
# 被整条替换，答复从 1097 字掉到 83 字，合规问答反而答不出来。
|
||
# 判定：命中处附近出现否定/禁止/引用线索才豁免，否则照旧拦（宁可误拦）。
|
||
blocked = _first_violation(output, hard_patterns) is not None
|
||
for kind, pattern in config.negative_rules:
|
||
if kind == "exact":
|
||
blocked |= output == pattern
|
||
else:
|
||
blocked |= _first_violation(output, (pattern,)) is not None
|
||
if blocked:
|
||
content = content.model_copy(update={
|
||
"text": "该内容需要人工核实。基金投资存在风险，本系统不代客交易。",
|
||
"source_references": (), "transfer_required": True,
|
||
"transfer_reason": "compliance_review_required",
|
||
})
|
||
else:
|
||
# Apply to text and citation titles, not only to the displayed answer.
|
||
def redact(value: str) -> str:
|
||
def _mask_mobile(match: re.Match[str]) -> str:
|
||
# 平台对外的客服热线是**公开业务号码**，不是客户个人手机号：
|
||
# 它必须原样出现在面向客户的回复里（spec B §5 要求固定联系方式），
|
||
# 否则 `TPL_TRANSFER_HUMAN` / 安全路由的 P2 话术会被脱敏成
|
||
# "[手机号已脱敏]"，客户拿不到联系方式（实测过）。
|
||
if match.group(0) == CUSTOMER_SERVICE_HOTLINE:
|
||
return match.group(0)
|
||
return "[手机号已脱敏]"
|
||
|
||
value = re.sub(r"(?<!\d)1[3-9]\d{9}(?!\d)", _mask_mobile, value)
|
||
return re.sub(r"(?<!\d)\d{15,19}[Xx]?(?!\d)", "[敏感号码已脱敏]", value)
|
||
|
||
content = content.model_copy(update={
|
||
"text": redact(output),
|
||
"source_references": tuple(ref.model_copy(update={
|
||
"title": redact(ref.title) if ref.title else None,
|
||
}) for ref in content.source_references),
|
||
})
|
||
# 门禁 F5：面向客户输出 100% 附固定话术，所以**拦截分支与正常分支都要走到这里**——
|
||
# 被负面词替换掉的回复同样是面向客户的输出。位置固定在最后：文案不参与上面的负面词
|
||
# 判定，也不会被替换逻辑吃掉。
|
||
# 幂等：只认**我们自己追加过的那个形状**（"\n\n" + 话术），而不是"文本以话术结尾"。
|
||
# 比 `endswith(disclaimer_text)` 更紧的原因：后者在**退化文案**下仍会漏加——管理员把文案
|
||
# 发布成"。"、答案恰好是"您好。"时，`endswith("。")` 为真 → 不追加，客户一条话术都拿不到；
|
||
# 认形状后这种文本必然被追加。失败形态因此从"合规漏加"变成"极端情况下多追加一次"，
|
||
# 这正是门禁 F5（面向客户输出 100% 附固定话术）要的方向。正文中间出现同样文字也不会被
|
||
# 误判成"已追加"——只有**末尾这个形状**才算。
|
||
# 空串不算话术：显式传空一律退回代码兜底，避免"传了但传空"成为绕过强制注入的后门。
|
||
disclaimer_text = (disclaimer or "").strip() or FALLBACK_DISCLAIMER
|
||
# 追加形状只在这里定义一次：判据与追加共用同一份，避免两处各写一个分隔符而漂移。
|
||
appended_shape = f"\n\n{disclaimer_text}"
|
||
if not content.text.endswith(appended_shape):
|
||
content = content.model_copy(update={"text": f"{content.text}{appended_shape}"})
|
||
return result.model_copy(update={"result": content})
|