按业务方确定的取向实现:金融场景确定性优先,能溯源到公司资料的才答,答不了就 引导客户拨打客服热线,绝不用模型猜答案。端到端验收 8/8 通过。 新增: - app/service/knowledge_search_service.py:知识检索。未复用记忆的 VectorMemoryAdapter 是因为它只返回 (memory_uuid, score),会丢掉知识块的标题与正文,而客服回答必须能把 原文与出处一起交付。检索失败一律返回 degraded 而不抛异常,由 Agent 走兜底。 - app/service/knowledge_tool.py + app/core/knowledge_contracts.py:只读工具 search_knowledge。 走 ToolExecutor 而不是让 Agent 直接持有检索服务,是为了让白名单、权限、审计、超时 都归基座统一管理;工具只读也符合 ToolRegistry 的硬约束。复用既有权限码 knowledge:reference:read(customer 角色已具备),不新增权限点。 - app/service/agent/implementations/customer_service.py:Agent 本体,刻意保持薄—— 意图分发 + 四条出口(faq/产品/政策直返、闲聊走模型、其余与异常引导人工)。 直接返回知识原文而不经模型改写,答案的字面内容全部来自公司已发布资料。 - tools/publish_customer_service_config.py:发布意图工具白名单。 - tools/customer_service_check.py:端到端验收(8 个用例,含越界请求与知识库外问题)。 装配: - bootstrap 新增 get_knowledge_search_service 工厂,注册 search_knowledge 工具与 customer_service Agent。 - runtime_config_service 新增 load_active_prompt:提示词绑定 release_id,按当前生效 版本读取,未发布时回落代码默认值。闲聊话术因此可审核、可回滚,不必改代码发版。 过程中发现并处理的三个问题: 1. 自造 source_references 被基座合规闸门拒绝。governance.review_output 只接受 「本次召回的记忆」与「本次成功调用的工具」两类引用(用于防止伪造来源), knowledge 类型会被判非法并使整个 run 失败。处理方式是**不放开那道校验**, 而把知识出处(文件标题与内部编号)写进正文,source_references 交给基座自动附加。 2. 发布配置是整版本替换语义:新版本会清空旧版本的全部配置项。若只发客服白名单, 示例 Agent 的 fund_query_demo:fund_quote 会被静默清空。故发布脚本先读取当前生效 版本的全部配置项并原样继承,再追加新增项。 3. 验收脚本自身两处自伤:打印 emoji 触发 GBK UnicodeEncodeError、以及读错结果字段 (RunQueryService 返回的答案键是 content 不是 text)。 已知缺口(未修,已记录): - CoreResult.transfer_required 未持久化:conversation_message 不存该标记, API 读不到"本次是否引导了人工"。当前靠正文里的固定话术判断。 - 知识块引用(source_type=knowledge)尚未启用,需先让 ToolExecutor 把工具返回的 doc_id 登记为本次可引用来源。 验证:ruff 通过、mypy 107 文件无错、unit+contract 447 passed; tools/customer_service_check.py 8/8 通过(含越界请求、投诉、知识库外问题三类 必须引导人工的场景,以及 7 个零容忍负面词零命中)。
186 lines
7.4 KiB
Python
186 lines
7.4 KiB
Python
"""发布客服 Agent 的运行期配置:意图工具白名单(走发布状态机)。
|
||
|
||
两个必须讲清的点:
|
||
|
||
1. **为什么必须发这一步**:工具白名单是失败关闭的——`ToolExecutor` 拿发布配置里
|
||
`agent_tools` / `customer_service:<intent>` 的 `allowed_tools` 与代码声明的
|
||
`AgentDefinition.allowed_tools` 取交集,缺配置时交集为空、任何工具调用都被拒。
|
||
「Agent 写好了但没发配置」的表现是"客服什么都答不了、一直在引导人工"。
|
||
|
||
2. **为什么必须继承现有配置项**:`config_release` 是**整版本替换**语义——激活新版本后,
|
||
旧版本的所有配置项都不再生效。若只发布客服自己的白名单,示例 Agent 的
|
||
`fund_query_demo:fund_quote` 会被静默清空。所以发布前先把当前 effective 版本里的
|
||
配置项原样搬进新版本,再追加本次新增项。
|
||
|
||
用法:python tools/publish_customer_service_config.py
|
||
"""
|
||
|
||
import asyncio
|
||
import datetime as dt
|
||
import json
|
||
import sys
|
||
import uuid
|
||
from pathlib import Path
|
||
from typing import Any
|
||
|
||
import asyncmy
|
||
import httpx
|
||
import jwt
|
||
|
||
from app.core.config import get_settings
|
||
from app.main import create_app
|
||
|
||
ADMIN = "9003"
|
||
AGENT_TYPE = "customer_service"
|
||
TOOL_NAME = "search_knowledge"
|
||
# 只有会调用工具的意图才需要白名单;chitchat(模型生成)与 transfer_human(引导人工)
|
||
# 都不查知识库。给它们配空白名单反而会掩盖"配置漏配",因此不发布这两条。
|
||
INTENTS = ("faq", "product_inquiry", "policy_explain")
|
||
|
||
|
||
def token(subject: str) -> str:
|
||
settings = get_settings()
|
||
private_key = Path(settings.jwt_private_key_path).read_text(encoding="utf-8")
|
||
now = dt.datetime.now(dt.UTC)
|
||
return jwt.encode(
|
||
{
|
||
"sub": subject, "iss": settings.jwt_issuer, "aud": settings.jwt_audience,
|
||
"exp": now + dt.timedelta(minutes=30), "nbf": now - dt.timedelta(seconds=5),
|
||
"jti": str(uuid.uuid4()),
|
||
},
|
||
private_key,
|
||
algorithm="RS256",
|
||
)
|
||
|
||
|
||
async def active_config_items() -> list[dict[str, Any]]:
|
||
"""读取当前生效版本的全部配置项,用于在新版本里原样继承。"""
|
||
settings = get_settings()
|
||
# MYSQL_DSN 形如 mysql+asyncmy://user:pass@host:port/db
|
||
dsn = settings.mysql_dsn.split("://", 1)[1]
|
||
credentials, location = dsn.split("@", 1)
|
||
user, password = credentials.split(":", 1)
|
||
host_port, database = location.split("/", 1)
|
||
host, _, port = host_port.partition(":")
|
||
connection = await asyncmy.connect(
|
||
host=host, port=int(port or 3306), user=user, password=password, db=database
|
||
)
|
||
try:
|
||
cursor = connection.cursor()
|
||
await cursor.execute(
|
||
"""
|
||
SELECT i.namespace, i.config_key, i.value_json, i.schema_version
|
||
FROM platform_config_item i
|
||
JOIN config_release r ON r.id = i.release_id
|
||
WHERE r.status = 'active'
|
||
"""
|
||
)
|
||
rows = await cursor.fetchall()
|
||
finally:
|
||
connection.close()
|
||
items: list[dict[str, Any]] = []
|
||
for namespace, config_key, value_json, schema_version in rows:
|
||
value = json.loads(value_json) if isinstance(value_json, str) else value_json
|
||
items.append({
|
||
"namespace": namespace,
|
||
"item_key": config_key,
|
||
"value_json": value,
|
||
"schema_version": schema_version,
|
||
})
|
||
return items
|
||
|
||
|
||
async def post(
|
||
client: httpx.AsyncClient, path: str, *, auth: dict[str, str],
|
||
payload: dict[str, object] | None = None, if_match: str | None = None,
|
||
) -> httpx.Response:
|
||
headers = {**auth, "Idempotency-Key": uuid.uuid4().hex}
|
||
if if_match:
|
||
headers["If-Match"] = if_match
|
||
return await client.post(path, json=payload, headers=headers)
|
||
|
||
|
||
async def etag_of(client: httpx.AsyncClient, path: str, auth: dict[str, str]) -> str | None:
|
||
return (await client.get(path, headers=auth)).headers.get("ETag")
|
||
|
||
|
||
async def main() -> int:
|
||
app = create_app()
|
||
auth = {"Authorization": f"Bearer {token(ADMIN)}"}
|
||
async with httpx.AsyncClient(
|
||
transport=httpx.ASGITransport(app=app), base_url="http://test", timeout=60
|
||
) as client:
|
||
inherited = await active_config_items()
|
||
print(f"当前生效版本的配置项:{len(inherited)} 条(将原样继承)")
|
||
for item in inherited:
|
||
print(f" · {item['namespace']} / {item['item_key']}")
|
||
|
||
new_items = [
|
||
{
|
||
"namespace": "agent_tools",
|
||
"item_key": f"{AGENT_TYPE}:{intent}",
|
||
"value_json": {"allowed_tools": [TOOL_NAME]},
|
||
"schema_version": "1",
|
||
}
|
||
for intent in INTENTS
|
||
]
|
||
inherited_keys = {(str(i["namespace"]), str(i["item_key"])) for i in inherited}
|
||
pending = [
|
||
item for item in new_items
|
||
if (str(item["namespace"]), str(item["item_key"])) not in inherited_keys
|
||
]
|
||
if not pending:
|
||
print("客服白名单已存在于当前生效版本,无需发布")
|
||
return 0
|
||
|
||
created = await post(client, "/api/v1/admin/config-releases", auth=auth, payload={
|
||
"release_no": f"cs-tools-{uuid.uuid4().hex[:12]}",
|
||
"title": "客服 Agent 意图工具白名单",
|
||
"change_summary": "新增 faq/product_inquiry/policy_explain 的知识检索白名单,并继承既有配置项",
|
||
})
|
||
if created.status_code != 201:
|
||
print(f"创建发布版本失败:{created.status_code} {created.text[:200]}")
|
||
return 1
|
||
release_id = int(created.json()["data"]["id"])
|
||
print(f"\n发布版本 id={release_id}")
|
||
|
||
base = f"/api/v1/admin/config-releases/{release_id}/platform-config-items"
|
||
for item in [*inherited, *pending]:
|
||
response = await post(client, base, auth=auth, payload=item)
|
||
mark = "继承" if item in inherited else "新增"
|
||
print(f" [{mark}] {item['namespace']}/{item['item_key']} → {response.status_code}")
|
||
if response.status_code != 201:
|
||
print(f" 失败:{response.text[:200]}")
|
||
return 1
|
||
|
||
release_base = f"/api/v1/admin/config-releases/{release_id}"
|
||
submitted = await post(
|
||
client, f"{release_base}/validations", auth=auth, payload={},
|
||
if_match=await etag_of(client, release_base, auth),
|
||
)
|
||
print(f"\n提交复核:{submitted.status_code}")
|
||
reviewed = await post(
|
||
client, f"{release_base}/reviews", auth=auth,
|
||
payload={"decision": "approved", "comment": "客服工具白名单"},
|
||
if_match=await etag_of(client, release_base, auth),
|
||
)
|
||
print(f"审核:{reviewed.status_code}")
|
||
activated = await post(
|
||
client, f"{release_base}/activations", auth=auth, payload={},
|
||
if_match=await etag_of(client, release_base, auth),
|
||
)
|
||
print(f"激活:{activated.status_code}")
|
||
if activated.status_code not in (200, 201):
|
||
print(f" 失败:{activated.text[:200]}")
|
||
return 1
|
||
print(f"最终状态:{activated.json()['data']['status']}")
|
||
|
||
remaining = await active_config_items()
|
||
print(f"\n激活后生效版本配置项:{len(remaining)} 条")
|
||
for item in remaining:
|
||
print(f" · {item['namespace']} / {item['item_key']} = {item['value_json']}")
|
||
return 0
|
||
|
||
|
||
sys.exit(asyncio.run(main()))
|