Decision patterns · 决策模式

Every decision pattern the catalogue files public Jev examples under, as plain pages in English and Chinese. The searchable catalogue has the same rows with filters.

目录收录的 Jev 公开示例按决策模式归类，这里是每个模式的静态页面，有中英两版。可搜索的目录里是同样的条目，另有筛选。

• Tool selection · 工具选择

  239 rows

  Which tool or action the agent should call next.

  智能体下一步该调用哪个工具或动作。

• Intent routing · 意图路由

  39 rows

  Classify what the user wants and send the request down the right branch.

  判断用户意图，把请求分流到正确的分支。

• Context compaction · 上下文压缩

  38 rows

  Decide which tool calls and results still matter so stale context can be dropped.

  判断哪些工具调用和结果仍然相关，从而丢弃过期上下文。

• Safety gating · 安全闸门

  144 rows

  Decide whether an action is safe to run. Defence in depth, never a security boundary.

  在执行前判断一个动作是否安全。属纵深防御，绝不是安全边界。

• Output validation · 输出校验

  138 rows

  Check a model's output against a rubric before it reaches a user.

  在输出到达用户前，按评分标准检查模型产出。

• Retry control · 重试控制

  8 rows

  Decide whether a failed step is worth retrying.

  判断失败的步骤是否值得重试。

• Human escalation · 人工升级

  73 rows

  Use calibrated confidence to decide what a person must see.

  用校准置信度决定哪些情况必须由人来看。

• Model routing · 模型路由

  49 rows

  Pick which downstream model or tier should handle a request.

  选择由哪个下游模型或档位处理请求。

• Speculative fan-out · 并行扇出

  35 rows

  Pack many questions — including speculative ones — into one request and let code pick what mattered.

  把大量问题（包括推测性的）打包进一次请求，再由代码挑出真正用得上的答案。

• Search & ranking · 检索与排序

  73 rows

  Score or re-rank candidates from a cheaper retrieval step.

  对来自廉价检索步骤的候选做打分或重排。

• Structured extraction · 结构化抽取

  18 rows

  Pull typed fields out of messy text by choosing among candidates rather than generating them.

  从杂乱文本中取出类型化字段 —— 靠在候选中选择，而不是生成。

• Classification · 分类

  127 rows

  Put an item into a taxonomy, including deep hierarchies walked with probabilities.

  把条目归入分类体系，包括用概率遍历的深层层级。

• ML feature extraction · 机器学习特征抽取

  8 rows

  Turn free text into numeric features for a classical downstream model.

  把自由文本转成数值特征，喂给下游的传统模型。

• Document triage · 文档分拣

  22 rows

  Classify and route incoming documents, invoices and forms.

  对进来的文档、发票、表单做分类和路由。

• Support triage · 工单分拣

  9 rows

  Route support tickets and conversations by intent and urgency.

  按意图和紧急度路由支持工单与会话。

• Content scoring · 内容评分

  166 rows

  Score quality, risk or relevance on an ordered scale.

  在有序量表上给质量、风险或相关性打分。

• Recommendation · 实时推荐

  2 rows

  Choose what to surface next, fast enough for a live conversation.

  选择下一步呈现什么，快到能用在实时会话里。

• Overview · 总览

  481 rows

  Surveys the model or the space rather than one pattern.

  介绍模型或整个领域，而非单一模式。