Context Window Management

01 · 四个人共用一只行李箱01 · Four travelers, one suitcase

输入、历史、检索证据和输出都想多带一点；总量超出 4,096 token 时，应该偷偷丢掉、压成摘要，还是明确拒绝？Input, history, retrieved evidence, and output all want more room. When the total exceeds 4,096 tokens, should the system silently drop content, summarize it, or reject the request?

02拖动四块预算，超限后选择截断、摘要或拒绝并回放答案后果Resize Four Budget Blocks, Then Choose Truncate, Summarize, or Reject and Replay the Consequence

这页讲怎么管 · S3-18 讲是什么this page manages · S3-18 defines

状态 AMetric A

—

状态 BMetric B

—

状态 CMetric C

—

所有概率、命中率、得分与输出均为固定种子的教学模拟，不代表真实模型。拖回主变量即可复核结论。All probabilities, hit rates, scores, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.

03把刚才的变化拆成 4 个零件Break the Change into Four Parts

Context Window Management

INPUT

当前输入Current input

任务与本轮材料占用基础预算。The task and current-turn material consume the base budget.

HISTORY

对话历史Conversation history

旧消息需保留、淘汰或摘要。Prior messages must be retained, evicted, or summarized.

RETRIEVAL

检索证据Retrieved evidence

证据完整度与来源定位受预算影响。Budget affects evidence completeness and source localization.

OUTPUT

输出预留Output reserve

不能把全部窗口都塞给输入。The entire window cannot be consumed by input.

04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction

MECHANISM

input + history + retrieval + output ≤ window

四个滑块直接验证总和；超限策略不改变窗口上限，只改变哪些信息被保留以及系统是否继续生成。The four sliders verify the sum directly. Overflow policy does not change window capacity; it changes what survives and whether generation proceeds.

分配预算Allocate四块竞争同一窗口four blocks share one window

检测超限Detect overflow总和 > 4,096sum > 4,096

执行策略Apply policy截断 / 摘要 / 拒绝truncate / summarize / reject

05把四块拖到 5,500 token 后仍选择截断Push the Four Blocks to 5,500 Tokens and Keep Truncation

主动制造失败Create a failure

边界实验Boundary experiment

系统仍会给答案，但末尾检索证据可能消失。容量问题被处理了，证据完整性问题却暴露出来。The system still answers, but trailing retrieved evidence can disappear. Capacity is handled while evidence integrity degrades.

尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction first.

✗ S3-18 和本页讲的是同一件事。S3-18 and this page teach the same thing.→ ✓ S3-18 定义窗口是什么；本页练习怎样分配和处理超限。S3-18 defines the window; this page manages allocation and overflow.

✗ 把输出预算设为 0 能装最多证据。Set output reserve to zero to fit maximum evidence.→ ✓ 没有输出空间，系统也无法完成回答。Without output capacity, the system cannot complete an answer.

✗ 截断后能生成就说明信息够用。If truncation still generates, enough information survived.→ ✓ 答案可生成不等于关键证据仍完整。A generated answer does not prove key evidence survived.

06 · 一句话带走06 · One line to keep

上下文窗口管理是在固定容量内分配输入、历史、检索和输出；超限时必须显式选择截断、摘要或拒绝，并检查答案后果。Context window management allocates input, history, retrieval, and output within fixed capacity. Overflow requires an explicit truncate, summarize, or reject policy—and a check of the answer consequence.