A19 · Prompt Playground / Experiment Platform A19 · Prompt 实验平台
Verified source经核实出处
Prompt: "Design a prompt playground…compare outputs…collaborate…prompt versioning…experiment management…evaluation…context too large" — PracHub, Onsite 2026-02. Credibility B.
Treat it as an experiment platform (not a UI wrapper)当作实验平台（而非 UI 包装）
flowchart LR UI[Web UI] --> API[Playground API] API --> V[(Prompt Versions)] API --> RUN[Run Orchestrator] RUN --> LLM[Model API] RUN --> E[(Eval Results)] API --> COL[Collaboration / ACL]
Four product-workflow questions you must answer必须回答的 4 个产品工作流问题
- Reproducibility per experiment: pin model version, params, context, tool versions, dataset version.每次实验可复现：固定模型版本、参数、上下文、工具版本、数据集版本。
- Comparison: side-by-side view + automated metrics (accuracy, preference, toxicity).对比：并排视图 + 自动指标（准确率、偏好、毒性）。
- Permissions: workspace, share links, audit log.权限：workspace、分享链接、审计日志。
- Oversize context: reference external docs (RAG), chunk prompt, or use prompt caching for repeated prefixes.超大上下文：引用外部文档（RAG）、分块 prompt、重复前缀用 prompt cache。
Prompt caching gotchaPrompt caching 陷阱
- Anthropic's prompt caching has 5-min default TTL (extensible to 1 hour at cost).Anthropic 的 prompt cache 默认 5 分钟 TTL（付费可扩展到 1 小时）。
- In a playground, cache the system prompt + background material prefix to cut cost; cache key MUST include workspace / tenant to prevent cross-tenant leaks.playground 中缓存 system prompt + 背景材料前缀以降本；cache key 必须包含 workspace/tenant 防止跨租户泄漏。