def weekly_digest(tenant_id):

1. 取一周日志，按事件类型聚合

events = query_logs(tenant_id, days=7)

2. 规则预处理：归并重复、压缩字段

compact = []

for type, group in group_by(events, "type"):

if is_repetitive(type): # 如失败登录

compact.append(f"{type}: {len(group)}次, 涉及{uniq_users(group)}个账号")

else:

compact += [compact_format(e) for e in group] # 只留who/what/when/target

3. 小模型先打风险标（便宜，可大量调用）

tagged = llm_small("对以下事件分类并标记可疑项", compact)

4. 中档模型生成摘要（只对压缩后内容）

digest = llm_mid(f"""为租户{tenant_id}生成周审计摘要：

• 高风险操作及涉及账号
• 异常模式（时间、频率）
• 每条结论附事件ID引用

事件列表：{tagged}""")

return digest # 平均延迟 < 20s，成本 < 0.05元/租户

Suggested rollout pace:

1. **Week 1**: Pick one tenant's historical logs for an offline pilot; validate summary quality and manually check accuracy;
2. **Week 2**: Connect to the unified gateway and the log pipeline; start with just the "trace-back scenario" (fastest time to value);
3. **Week 3**: Launch weekly report automation with a manual spot-check mechanism (randomly review 3 reports per week);
4. **Week 4**: Tune prompts and compression rules based on spot-check feedback; once stable, gradually roll out to all tenants.

## Results Reconciliation

Actual numbers three months after launch:

| Scenario | Before | After |
|------|------|------|
| Weekly audit summary | 6 hours per report | 15 minutes reviewing AI output |
| Single problem trace-back | ~40 minutes | ~3 minutes |
| Monthly total investment | ~104 person-hours | ~8 person-hours + under 200 RMB in token costs |

The time saved isn't the whole point—more importantly, **trace-back went from "a developer's burden" to something support staff can do themselves**, removing a link from the issue-response chain.

## Final Thoughts

For small teams, the barrier to shipping AI applications is no longer model capability, but engineering minutiae: integration, switching, billing, fallbacks. Hand these off to a unified AI API gateway, and you can focus your energy on the actual business summarization logic.

If you also want to quickly get your first AI scenario running, you can start by registering a gateway account and getting the summary feature up with a single base_url:

👉 [https://api.thistoken.ai/register](https://api.thistoken.ai/register)

---

Ready to try it yourself? Sign up at https://api.thistoken.ai/register to get your API key and start building.