大家好，我是狂师。
过去一个多月，我干了一件事，开发了一个免费开源的测试开发刷题学习网站。刷题的部分是我把测试开发的面试题梳理了一遍，整理出 10 个专题共 200 题，学习路线、实战教程、工具导航也都补充了进来。
👉 https://ai.testfather.cn/
今天就把这套「AI 测试面试题大全」和这个网站，分享给大家。
一、为什么要做这件事
起因很简单。今年九月初，我在知识星球里开了一个面试题连载，把大模型测试方向的面试题一道道拆开讲，题目全部来自真实面经。
发着发着发现两个问题。
一是散。面经帖在牛客、脉脉、各个群里到处都是，今天爬到一道 RAG 评测，明天刷到一道自动化框架，东一榔头西一棒子，攒了半年也攒不成体系。
二是浅。多数面经只有题目，顶多附一段好心人回忆的答案，背完还是不知道面试官想听什么。同样是「性能测试怎么做」，校招生答三个指标就能过，社招五年经验答三个指标直接挂，这中间的差距，面经帖里没人告诉你。
所以这 200 题，我给自己定了个整理标准，每道题都要能回答四个问题，面试官为什么问这道题，答到什么程度算过线，什么答案能拉开区分度，什么细节是加分项。
按这个标准，整理了10个专栏，一个专题 20 题。
二、面试刷题库长什么样
先看全貌。题库覆盖了测试开发求职的十个方向。
每个专题内部再按子方向分组，比如大模型测试 20 题分成认知体系、评测集、指标方法、幻觉安全、性能稳定、工程落地六组，刷完一组，这块的面试就能接得住了。
三、拆开一道题看看长啥样
光说结构虚，挑一道举例，比如：「RAG 系统的评测怎么做」。
这题现在的出镜率极高，因为知识库问答是企业落地最多的大模型应用，面试官拿它一问，你是真评过还是只看过文章。
题库里这题是这样展开的。
先一句话定调，RAG 等于检索加生成，评测必须分层做，混在一起评就没办法归因，不知道该修检索还是修生成。
然后分层展开。检索层看上下文召回率、精确率、MRR、NDCG，其中排序质量有个很实用的判断，标准片段排在第 8 位基本等于没召回。生成层看忠实度、答案相关性、引用正确性。最后端到端评整体正确率，错了做归因，没召回就调切分和索引，召回了没用上就调 prompt 和排序，生成本身错就换模型或加约束，三类归因对应三条修复路径。
真正让这道题和面经答案拉开差距的，是题库里附的考察重点。
这四个维度，就是每道题标配的「面试官视角」。刷题的时候不光知道自己答没答对，还知道这道题在面试现场是几斤几两。
200 题，每一题都是这个配置。
四、怎么刷题才有效
题库页上我写了一句话，这里再强调一遍，别光背答案。
只背下来的答案，自己没有真正理解，面试官稍加追问就BBQ了。正确的刷法是，看到题目先别看思路，自己对着空气口述一遍答案，讲不下去的地方就是你的缺口，然后再对照题库里的分层思路，看自己漏了哪一层。
每道题的思路都是按面试现场的表达结构组织的，先定调，再分层展开，最后收个尾。你口述出来的版本和它对不上的地方，要么是知识缺口，要么是表达结构问题，前者补知识，后者照着框架练几遍就顺了。
刷的顺序也有讲究。社招求职，先刷「软件测试必备」保底，再按目标岗位主攻一到两个方向，做 AI 方向就把「大模型测试」和「AI 测试提效」刷透，这两个专题眼下最值钱。校招打基础，软件测试必备加自动化测试是主力，其余专题过一遍留个印象，面试遇到至少有框架可讲。
五、网站打开就能刷，完全免费
这些内容现在全部放在网站上，不用注册，不用登录，不用下载任何东西。
👉 https://ai.testfather.cn/
打开就是学习路线首页，顶部导航点「面试题库」进题库，左侧栏十个专题随时切换，每道题独立成页，手机电脑都能看，通勤路上也能刷。
面试题库只是这个网站的一部分。整个站点是一套完整的测试开发学习体系，刷完题，接下来每一步学什么、怎么落地，站里也都铺好了。
先说学习路线。
13 个阶段，从零基础的岗位认知一路修到面试拿 offer。功能测试转测开、后端开发转测开、校招零基础，三条定制路径各自标好了重点阶段，不同背景不用从头读到尾，转行的跳过免修段直接主攻就行。每章末尾有自查清单，清单全答得上来，这一章才算过关。
9 月我写过一篇路线介绍，13 个阶段逐段拆解放在那篇里，这里不再重复展开，感兴趣的翻历史文章。
路线管学什么，教程管怎么落地。实战教程板块收录了 AI 测试实战系列全部 22 篇，围绕一个真实电商项目 shop-lab 展开，从需求拆解、用例设计，到接口自动化、UI 自动化，再把一条完整的业务链路串起来跑通。每篇对应一个 Agent Skill，接口文档解析、测试数据构造、失败自动诊断、报告生成，全是从环境搭建到跑通的真实操作，照着敲就能复现。
所有的 自研Skill 在「AI测开进化圈」星球中可以直接获取。
日常干活还差一个工具箱。资源导航板块收了 140 个测试工具，按 AI 编程、接口、UI 自动化、性能、安全五大类分好，每个工具独立一页，介绍、官网、开源地址齐全，GitHub 仓库和 star 数都实时核验过，标注了核验时间。另有 66 份推荐书单，九大分类，每份带永久网盘链接。
最近要面试的无脑先刷题库，时间充裕的从路线第一章走起，实战和工具导航留着日常慢慢用。
六、GitHub 同步开源
网站的全部内容，同步开源在 GitHub 仓库里。
👉 https://github.com/zhoujinjian/ai-testing-guide
仓库 8 月底建起来，内容用 CC BY-NC-SA 协议发布，个人学习、博客署名转载随意，注明出处就行。
每道题都是一个独立的 Markdown 文件，你觉得哪道题的思路还能补案例、哪个方向还缺专题，直接提 Issue 或 PR，我都会看。一个人整理 200 题，肯定有覆盖不到的地方，各位道友面完试回来，把被问住的新题贡献出来，这个题库就能一直长下去。
写在最后
最后说句实在的。
面试题刷多少道，从来都只是手段。200 题真正想帮你做到的，是把每个方向的考点框架装进脑子里，面试官不管怎么换着法子问，你都能顺着框架一层层往下讲。
网站免费，开源，长期更新。金九银十赶不上的，还有明年的金三银四，题库就在那里，什么时候开始刷都不晚。
觉得有用的话，去 GitHub 给个 Star ⭐，这是持续更新的最大动力。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Hello, I'm a madman.
Over the past month, I've done one thing, and I've developed a free, open-source test development learning website. The part of the puzzle is that I combed the interview questions that were developed for the test, organized a total of 10 topics totalling 200, and added to the study route, the field course, the tool navigation.
👉 https://ai.testfather.cn/
Today we'll share this "AI test interview package" and this site.
I... why did you do this?
The cause is simple. At the beginning of September, I launched a series of interviews on the Knowledge Planet, which were conducted in conjunction with the LLM test directions, all of which were based on real faces.
There are two problems.
One is scattered. The posts are all over the bulls, veins and herds, and they're crawling into a RAG assessment today, and tomorrow they're going to have an automated frame, a stick to the east, and they're not going to be systematic for six months.
Two is shallow. Most of them have only the subject, at best with an answer to a good man's memory, and they still don't know what the interviewer wants to hear. It's the same thing as "how to test performance" so that you can get through three indicators, and five years of experience and three indicators, the difference between which is not shared in the text.
So, 200 questions, I set myself a standard of four questions for each, why the interviewer asked the question, what the answer was, what the difference was, what the details were.
By this standard, 10 columns, one topic, were organized.
ii. What's the interview brush library like?
Look at the whole picture. The library covers ten directions for testing and development of jobs.
Each topic is then grouped in sub-directional groups, such as the LLM test, which is divided into cognitive systems, assessment and measurement, indicator methodology, hallucinogenic safety, performance stability, project landing, and a set of interviews can be completed.
Three, open the question and see what it looks like.
It's just a weak structure, and it's like, "What does the RAG system do?"
This is a very high-visibility issue, because the knowledge base is the largest application of LLM in business, and the interviewer asked whether you really read or only read articles.
That's how it goes.
In the first sentence, the RAG is equal to the search plus generation, and the evaluation must be done in layers, so it cannot be attributed to the combination of the evaluation, and it is not known whether the retrieving or the retrieving is generated.
Then spread out in layers. The search layer looks at context recall rate, accuracy rate, MRR, NDCG, where the sorting quality is judged in a practical way, and the standard segment is almost equal to no recall at the eighth place. Generating layers look at loyalty, relevance of the answer, correctness of reference. The final-to-end evaluation of the overall correctness rate, miscalculation, adjustment of points and indexation without recall, recall of non-useful prept and sorting, generation of the error itself into a model or constraint, and three attributions to the three repair paths.
It is the focus of the study attached to the library that really leaves the question and the face open.
These four dimensions are the "interviewer's perspective" for each theme. Not only did they know they were right, but they also knew how many pounds it was at the interview.
200 questions, each with this configuration.
iv. How does it work?
I wrote a sentence on the library page, and I'm going to repeat it here, so don't back the answer.
The only answer left behind was that he did not really understand it, and the interviewer asked a little more about BBQ. The correct way to do it is to look away from the subject, to say the answer to the air, to say what you can't say is your gap, and then to see which layer you're missing from the stratification in the library.
Each topic is organized along the lines of the expression structure of the interview site, with a set tone followed by a layering and ending. Your oral version is either a knowledge gap or an expression of structural problems, the former supplementing knowledge and the latter following the framework several times.
The order of brushes is also detailed. The job offers are backed up by the software test, followed by the main target position in one or two directions, and the AI direction is used to get through the "LLM test" and "AI test benefits", which are the most valuable topics at the moment. The basics of the school entrance are based on the need for automated testing of the software, and the remaining topics are marked over and over and over again, with interviews having at least a framework to talk about.
V. Websites can be painted, completely free.
All of this is now available on the website, without registration, without login and without downloading anything.
👉 https://ai.testfather.cn/
Opens the first page of the learning route, the top navigator's interview library goes into the library, and 10 topics in the left-hand column are switched at any time, each of them becomes a separate page, mobile computers are visible and can be painted on the commuting road.
The interview library is only part of this website. The entire site is a complete test development learning system, which completes the question and is prepared for every step that follows.
Let's start with the learning route.
13 Stages from basic job recognition to interview for offer. Functional test transfer, back-end development transfer, zero basis for recruitment. Each of the three customized paths has its own focus phase. The different contexts do not need to be read from the beginning to the end, and the shift is only required to skip the immediate primary part of the exemption. The self-checklist is available at the end of each chapter and the list can be answered in full.
In September, I wrote an introduction on the route, where 13 stages of the break-up were part of that article, where it was no longer repeated and interested in historical writing.
What's the routing, how's the curator landing? All 22 of the AI Test Fields series were recorded in the field, and started around a real electrician project, shop-lab, from demand dismantling, case design, to interface automation, UI automation, and a full business link running. Each piece corresponds to an Agent Skill, and interface document resolution, test data construction, failure autodiagnosis, report generation, all from the actual operation of setting up the environment to running through, and can be repeated by knocking.
All self-study Skill can be accessed directly on the "Ai-Speculating the Evolution Circle" planet.
There's still one toolbox missing for daily work. The resource navigation panel received 140 test tools, classified according to the five main categories of AI programming, interfaces, UI automation, performance, safety, each with a single page, a description, an official network, an open source address, a real-time verification of the GitHub repository and the number of Stars, indicating the time of verification. There are also 66 recommendation forms, nine major categories, each with permanent Internet links.
The most recent interviewer, brainless first-hand, started with sufficient time from the first chapter of the route, and the operational and tool navigation was kept slowly and daily.
vi. GitHub Sync Open Source
All content of the website, synchronous open source in GitHub repository.
👉 https://github.com/zhoujinjian/ai-testing-guide
The warehouse was built at the end of August, with content in CC BY-NC-SA protocol publish, and individual studies, blogs, and postings are available.
Every question is an independent Markdown document, and I'll look at it if you think there's a case to be taken, there's a topic to be taken, there's a point to go, there's an Issue or PR. One person can sort out 200 questions, and there must be something you can't cover, and you guys can come back and contribute to the new questions you've been asked.
At the end.
Last but not least.
How many interviews do you have? What you really want to do is to put in your head a frame of points in each direction, and the interviewer can go down one level, no matter what.
The site is free, open and updated.
If you feel useful, go to GitHub for a Star, which is the biggest incentive for continuous renewal.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.