用户从不新开会话致成本超均值 8 倍，作者用新话题弹窗解决
Lopsided_Feedback_81 · reddit · 2026-11-02
An indie developer running a data-analyst AI assistant for small e-commerce stores (3-10 person teams) shares a real lesson: assume your users don't know how to use LLMs.
Analyzing MLflow traces, they found some users cost over 8x the average per conversation. Not because of big data pulls — these users simply never start a new conversation, piling unrelated questions (refund rates, FB campaign performance) into one monolithic, days-long thread where old context can't even hit cache.
- Context compression/truncation was tried before and clearly hurt performance; near-limit warnings alone weren't enough
- Fix: before sending each new message to the agent, a quick check decides if it's related to the conversation; if not, a slightly annoying modal asks "This seems like a new topic. Start a new chat?"
- Result: 90% "Yes" rate, and snowballing conversations dropped sharply
Takeaway: products must guard-rail non-technical users rather than expect best practices.
「编程与Agent」频道最新
- 播客实录：管理 10 人+50 个 Agent、砍掉 97% Claude 成本、扛过月 55 万美元 API 账单 — Ror_Fly · 2026-11-03
- ICSE 2027 LLM4Code 工作坊开放投稿,11月13日截止 — LingmingZhang · 2026-11-03
- Claude Code 启动上下文占用从 63K 降到 22K,省 65% — dSebastien · 2026-11-03
- 「我沉迷点这些按钮」：编码 agent 权限弹窗成圈内梗 — ThePeterMick · 2026-11-02
- 用户抗议 OpenAI 拟合并 ChatGPT 与 Codex 用量额度 — TinfoilTricorn · 2026-11-02
- 离谱想法：带一支 agent 团队能否把人练成更好的工程经理 — HankYeomans · 2026-11-02