A21 · Design Claude Chat Service A21 · 设计 Claude Chat 服务
Verified source经核实出处
Prompt: "Design Claude Chat Service." — interviewing.io. Credibility C/D.
Core concerns核心考察
- Session management (long-lived conversations with token limits).会话管理（带 token 限制的长对话）。
- Streaming output over SSE / WebSocket.SSE / WebSocket 流式输出。
- Token-level billing (metered at worker after tokenization).Token 级计费（tokenization 后按 worker 计）。
- Log aggregation (conversation + safety annotations for audit).日志聚合（对话 + safety 注释用于审计）。
- Safety pipeline integration (input filter, output filter, content policy).Safety 管道集成（输入过滤、输出过滤、内容策略）。
Architecture架构
flowchart LR C[Client] --> GW[Gateway] GW --> SESS[Session Service] SESS --> CONV[(Conv Store)] GW --> SAFE[Safety Input Filter] SAFE --> INF[Inference] INF --> SAFE2[Safety Output Filter] SAFE2 --> GW GW --> METER[Usage Metering]
See A11 for inference depth and A19 for prompt versioning.推理深度参考 A11，Prompt 版本化参考 A19。