跳到主要内容Skip to main content

动手开发者路线The Builder's Route

约 60 分钟about 60 minutes

按搭建顺序读接口、检索、工具调用、评测与成本。阅读用于理解链路;实际接入还需按当前版本检查并验证。Study interfaces, retrieval, tools, evaluation and cost in build order. This explains the pipeline; actual integration still needs version checks and validation.

学完自测Checkpoint
  1. 说清一次 API 调用的费用由哪几段 token 构成
  2. 解释检索链路里分块与嵌入各自决定什么
  3. 指出接入工具调用时最先要设的两道防线
  1. Break down an API call's cost into its token segments
  2. Explain what chunking and embedding each determine in a retrieval pipeline
  3. Name the first two defenses to put in place when adding tool calling

先试着回答,再展开参考答案。答不上来的地方,可从上方步骤返回对应词条。参考回答是核对要点,不要求逐字复述。Try answering before opening the reference. Revisit the relevant step above if needed. These answers highlight key ideas; you do not need to repeat the wording.

第 1 题 · 参考回答Question 1 · Reference answer

输入包括指令、历史、检索材料和工具返回;输出是模型生成的内容。按各自用量乘相应单价,再核对缓存、工具与重试是否另计,不用一个总 Token 数乘同一单价。Input includes instructions, history, retrieved material and tool results; output is generated content. Apply the relevant rates and check cache, tool and retry charges separately. Do not assume one rate for every token.

回看这一题的概念Revisit this concept

第 2 题 · 参考回答Question 2 · Reference answer

分块决定一次取回的信息单元;嵌入决定文本的向量表示。先查答案所在块是否进入候选,再检查排序与生成,避免把检索问题误当成生成问题。Chunking defines retrieval units; embedding represents text as vectors. Check whether the answer-bearing chunk reaches candidates before investigating ranking and generation.

回看这一题的概念Revisit this concept

第 3 题 · 参考回答Question 3 · Reference answer

执行端校验工具名称、参数与授权范围;把检索内容和工具结果作为不可信数据处理。涉及写入或外部影响的动作,按业务风险设置审批与可撤回方式。Validate tool names, arguments and authorization at execution. Treat retrieved text and tool results as untrusted data. Add approval and recovery according to the risk of writes or external effects.

回看这一题的概念Revisit this concept

用在一个具体任务里Apply this to a concrete task

按一个小问题练习复现与验收Practice reproduction and acceptance on a small issue →