跳到主要内容Skip to main content

数据与 RAG 路线The Data & RAG Route

约 50 分钟about 50 minutes

数据从哪来、能不能用、怎么切、怎么测。读完能搭一条最小可评测的 RAG 语料,而不是把文档一股脑丢进向量库。Where data comes from, whether it is usable, how to chunk it and how to test it. Finish with a minimum evaluable RAG corpus, not a pile of PDFs in a vector store.

学完自测Checkpoint
  1. 拿到一份数据集,能列出至少三项必须核对的授权信息
  2. 说清「按固定字数切」和「按语义结构切」各自会在什么问题上失效
  3. 用 10 个真实问题跑一次检索,记录命中率、引用准确率和拒答率
  1. Given a dataset, list at least three licence facts that must be checked
  2. Explain when fixed-length chunking and semantic chunking each fail
  3. Run 10 real questions through retrieval and record hit rate, citation accuracy and refusal rate

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

许可证类型、商用/再分发/衍生作品条款、地域与期限、是否含个人数据、原始来源能否追溯。Licence type, commercial/redistribution/derivative terms, territory and term, whether personal data is included, and whether the original source is traceable.

回看这一题的概念Revisit this concept

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

固定字数会把一个条件切成两半,答案缺上下文;语义切分遇到格式混乱、无标题的文档会切得太碎或太大。Fixed length splits a condition in half and removes context; semantic chunking over-splits or over-merges messy documents without headings.

回看这一题的概念Revisit this concept

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

命中率看正确文档是否进前 k;引用准确率看答案是否引对原文;拒答率看无依据时是否说不知道。Hit rate: does the right document enter top-k; citation accuracy: does the answer cite the right passage; refusal rate: does it say “I don’t know” without evidence.

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

挑一个数据集,按数据卡模板补全来源、授权与偏差三节Pick a dataset and fill source, licence and bias sections using the dataset-card template →