把任意材料变成试卷
Material into exam
上传 PDF / Word 或粘贴一段材料，自动切分知识点，生成单选、判断、填空和简答。
Upload a PDF or Word file, or paste text: get topics, multiple-choice, true/false, cloze and short-answer questions.
点睛
把自己的学习材料变成一套可以自动判分、并且能解释为什么得这个分的考试。
在线使用（Web 版）
Use it online (web app)
上传自己的材料 → 自动出题 → 逐得分点判定 → 错题本与薄弱点。自带模型 Key（支持 DeepSeek、智谱 GLM、通义千问、Kimi、OpenAI、Claude 等），材料与密钥都只在你自己的账号下。
Upload your material, generate an exam, get point-by-point grading, then track mistakes and weak topics. Bring your own model key (DeepSeek, GLM, Qwen, Kimi, OpenAI, Claude and more); your material and keys stay in your own account.
目录
Contents
为什么做Why
解决什么问题Problems
可以用来做什么Use cases
错题会自己排期Spaced review
上传 PDF 与追问Upload and ask
可迁移、可纠错、可删除Portable, correctable, deletable
它怎么工作How it works
三个可核对的保证Three guarantees
关键设计Design decisions
实测与边界Evidence and limits
怎么用起来Get started
为什么做
Why this exists
备考的瓶颈通常不是「没有资料」，而是「不知道自己哪里没掌握」。读完一份材料，感觉都懂了；真正被问到时，才发现有些要点根本没说出来。而通用题库和你的材料往往对不上：题目是别人出的，考点是别人的重点。
The bottleneck in exam prep is rarely the material. It is not knowing which parts you actually failed to learn. After reading, everything feels understood; when questioned, some points turn out to be missing. Generic question banks do not fix this either, because the questions and the key points belong to someone else.
这个项目的目标很具体：把你自己的材料变成一场考试，并且让每一分都能追溯到「哪一个要点没说到」。
The goal is narrow: turn your own material into an exam where every point traces back to a specific idea you did or did not express.
解决什么问题
What it solves
- 01
学完没有反馈
No feedback loop
材料是单向输入的。没有问答环节，就很难知道要点是否真的记住了。
Material is one-way input. Without a questioning step, there is no evidence that a point was actually learned.
- 02
对错之外的空白
Beyond right and wrong
简答题只给一个总分，看不出是漏了定义、漏了条件，还是把结论说反了。
A single score for a written answer hides whether a definition, a condition or the conclusion itself was missing.
- 03
通用题库与材料脱节
Generic banks do not match
题库考的是别人的重点，答案也无法回到你手上这份材料的原文。
Question banks test someone else's priorities, and their answers cannot be traced back to your source document.
- 04
进度只靠感觉
Progress by feeling
没有错题本和掌握度，复习只能从头再来一遍。
Without a mistake log or mastery signal, revision restarts from zero every time.
可以用来做什么
What it can do
看不懂就追问
Ask when stuck
就这份材料提问，回答逐句带引注，能点回原文（PDF 还能给到页码）；材料里没有的会明说。
Ask about the material and get an answer cited sentence by sentence, linking back to the source (with page numbers for PDFs); anything the material does not cover is stated plainly.
逐得分点判分
Point-level grading
简答题不是给一个 0–100 的黑盒分数，而是逐条判定「这个要点说到了没有」。
A written answer is graded point by point instead of receiving one opaque 0-100 score.
把不确定标出来
Uncertainty made visible
模型没把握的题目标记为「待复核」，给出分数区间，并且不计入掌握度。
Low-confidence judgments are flagged with a score range and excluded from mastery tracking.
错题本与薄弱点
Mistake log and weak spots
自动收集错题，按知识点聚合掌握度，一键重考错题；失分的题目还会按 FSRS 自动进入复习排期。
Mistakes are collected, mastery is aggregated per topic, retakes reuse the same questions, and lost points are scheduled for spaced review with FSRS.
错题不是终点：交卷之后它会自己排期
The exam does not end at submission: wrong answers schedule themselves
只有错题本是不够的——记下错题不等于记住它。交卷之后，每一道失分的题目会立刻进入复习队列（当天即可重来），之后按 FSRS-5 安排下一次出现的时间。
A mistake log alone is not enough: recording a mistake is not the same as learning it. After submission, every question you lost points on enters the review queue immediately (due the same day), and FSRS-5 schedules when it comes back.
- 复习时先用同一条判定链路给分，再把分数映射成 FSRS 评分（1–4），推进 stability 与 difficulty，算出下一次到期时间：答得稳，间隔越来越长；答错，当天就来（约 10 分钟后）。
- Each review is graded by the same pipeline, then the score maps to an FSRS rating (1–4) that updates stability and difficulty and produces the next due date: answer confidently and intervals stretch out, miss it and it comes back the same day.
- 待复核的判定不参与排期——不确定的结果不应该决定你的复习节奏。手写作答等无法自动判定的场景可以自评推进。
- Judgments flagged for review never affect scheduling: an uncertain result should not set your study rhythm. Handwritten answers that cannot be auto-judged can be self-rated.
- 仪表盘与结果页都会显示「今天有几张卡到期」，复习页按到期顺序出示卡片，并标出逾期天数与上次得分。
- The dashboard and the result page show how many cards are due today; the review page serves cards in due order with overdue days and the last score.
直接上传 PDF / Word，以及随时追问
Upload a PDF or Word file, then ask follow-up questions
不用先把资料转成纯文本。上传 PDF 会逐页解析并记录每一页在正文里的位置，Word（.docx）取正文；解析结果先回填到表单让你确认，确认后才保存——解析错的段落不会悄悄进材料。扫描件与图片型 PDF 提取不到文字时会直接说清楚，而不是给你一份空材料。
You no longer have to convert your notes to plain text first. PDFs are parsed page by page and each page's position inside the body is recorded, so a citation can point at a page; Word files give up their .docx text. Extraction only fills the form — nothing is saved until you confirm it, so a bad extraction never lands in your material silently. Scanned or image-only PDFs are reported as having no text instead of producing an empty material.
- 解析在服务端完成（pdfjs 与 mammoth），单个文件上限 4 MB；旧版
.doc会提示先另存为.docx，加密 PDF 也会给出对应的中文提示。 - Parsing runs on the server (pdfjs and mammoth), capped at 4 MB per file. Legacy
.docfiles are asked to be re-saved as.docx, and encrypted PDFs get their own Chinese error message. - 同一套解析器也做进了命令行：
npx eyedot extract --file 笔记.pdf --out material.md，Agent 用同一份正文出题。 - The same parser ships as a CLI command:
npx eyedot extract --file notes.pdf --out material.md, so an agent can quiz you on the exact same text.
第二个能力是「就这份材料提问」。做法不是把整份材料塞给模型，而是先在材料内检索出最相关的几句，编号后交给模型，要求它每句事实性陈述都带上引注；回答回来之后再逐条校验。
The second capability is asking questions about the material. Instead of stuffing the whole document into a prompt, the app retrieves the most relevant sentences, numbers them, and requires the model to cite a number on every factual sentence — then verifies the citations after the answer comes back.
引注必须指得回去
Citations must resolve
模型只能引用给定编号；编造出 [9] 这类越界编号会被代码直接删掉，并作为一条核验提示列出来，而不是留在正文里假装有出处。
The model may only cite the numbers it was given. An invented marker like [9] is deleted by code and reported as a verification issue instead of being left in the text looking like a source.
没出处的句子会被点名
Uncited sentences are named
没有带引注、又有实质内容的句子会被单独列成清单；材料之外的知识必须另起一句、以「【模型补充】」开头。
Substantive sentences without a citation are listed separately, and anything outside the material must start a new sentence tagged "model-added".
检索不到就不生成
Nothing found, nothing invented
材料里找不到相关句子时，直接回答「材料里没有直接说明」，不调用模型、不消耗额度。没有配置对话模型时降级成「离线摘录」：只摘原文，不做生成。
When no relevant sentence exists, the app answers "the material does not say" without calling a model or spending quota. With no chat model configured it degrades to an offline extract: raw sentences only, no generation.
数据是你的，判错有出口，账号能删掉
Your data, a way to disagree, and a delete button
学习工具最容易被抱怨的两件事：数据被锁住、判错了还没处说。这两条都做了具体实现，而不是写在文案里。
Two complaints kill study tools: data you cannot get out, and a wrong grade with nowhere to report it. Both are implemented, not just promised in copy.
三种格式都能带走
Three export formats
Markdown（材料正文 + 题目与评分点 + 学习状态）、Anki CSV（复习卡片，按 RFC 4180 转义）、完整备份 JSON（含材料、作答、判定、掌握度、复习排期）。都不压缩、不含你的 API Key。
Markdown (material, questions and rubric points, learning state), Anki CSV (review cards, RFC 4180 escaped) and a full JSON backup covering material, answers, judgments, mastery and review schedule. Uncompressed, and never containing your API key.
判错就上报，而且冻结现场
Report a wrong grade, with the scene frozen
结果页每题都有「这题判错了？」。提交时会把当时的题目、你的作答、逐得分点的概率和引擎版本一起存下来——判定逻辑会迭代，只记一个题目 ID，三个月后回看已经无法复核。
Every question has a "graded this wrong?" button. Submitting freezes the question, your answer, the per-point probabilities and the engine version — grading logic evolves, so a bare question ID would be unreviewable three months later.
上报会长成评测集
Reports grow the benchmark
npm run feedback:golden 把上报导出成金标准候选，人工确认标签后再并入评测集。这是把标注集从 12 题做到 60–100 题最省力的来源，而不是靠自己编题。
npm run feedback:golden exports reports as benchmark candidates; after a human confirms the labels they join the golden set. That is the cheapest path from 12 items to the 60–100 target, instead of inventing questions ourselves.
账号可以真的删掉
The account can really be deleted
设置页输入邮箱二次确认后删除账号与全部派生数据，会话立即失效。另有独立的服务条款与隐私说明，写清数据处理、版权责任与生成内容免责。
Settings deletes the account and every derived record after an email confirmation, and the session dies immediately. Separate terms and privacy pages spell out data handling, copyright responsibility and the limits of generated content.
它怎么工作
How it works
这里有一个容易混淆的地方：判定层不是「一个会打分的聊天模型」。它只回答一个很窄的问题——这个得分点说到了没有——然后给出类型化概率（命题为真的概率、选项的概率分布、有序 rubric 上的分数），不生成任何解释文字。同一个接口下也可以换成 Jev 这类决策模型（System One），或者干脆不用模型。所以整条链路上有三种明确的角色分工。
One distinction matters: the grader is not "a chat model that hands out scores". It answers one narrow question — was this rubric point covered — and returns typed probabilities rather than prose. The same interface can be backed by a decision model such as Jev (System One), or by no model at all. That splits the pipeline into three explicit roles.
01出题与拆点Generate普通 LLM：切分知识点、出题、拆评分点LLM splits topics, writes questions and rubric points
02客观题判分Deterministic代码比对：只有填空的语义等价才问模型Code compares answers; only semantic cloze asks a model
03主观题判分Point by point每个得分点一条判定，代码加权合成One judgment per rubric point, combined in code
- 出题必须由生成式模型完成。判定侧只回答「说到了没有」，一个字都不生成，所以材料理解、题干、评分点都得由通用大模型来写，并且每道题都要能在原文中定位到出处。
- Question writing needs a generative model. The grading side only answers "was it covered" and produces no text at all, so topics, stems and rubric points come from a general LLM, and every question must anchor back to a verbatim span.
- 客观题由代码判分：选择、判断、填空先做归一化比对（全半角、大小写、标点、可接受写法），只有字面不一致的填空才会调用一次「语义等价吗」的判定。
- Objective questions are graded in code. Choices, true/false and cloze are normalised first; only a cloze answer that does not match literally triggers one semantic-equivalence question.
- 简答题拆成一批原子问题：每个得分点一条 noul 问题，问题里带着该得分点的陈述与原文证据，最后按权重合成，再减去「与材料矛盾」和「编造材料外事实」这两项扣分。
- A written answer is decomposed into atomic questions: one per rubric point, each carrying its own statement and evidence span. The score is a weighted sum minus explicit penalties for contradicting the material or inventing facts.
三个可核对的保证
Three guarantees you can check
「凭什么相信这个分数」不应该靠信任，而应该能逐条核对。项目把三件事写成了契约，并且做了校验：
"Why should I trust this score" should not require trust. Three things are written as contracts and verified in code:
材料事实必须可定位
Material facts are locatable
题目的出处与每个得分点的依据，必须逐字出现在材料中；定位失败会被记为违规并在报告里列出。报告里所有字段都带标签：材料原文 / 模型补充。
A question's source and every rubric point's evidence must appear verbatim in the material. Failures are reported as violations, and every field is tagged material or model-added.
覆盖率不假装完整
Coverage is not faked
材料会被切成要点单位，没有被任何题目覆盖的部分会被如实列出来（例如「覆盖 11/14」），而不是暗示这份卷子覆盖了全部内容。
The material is split into units, and anything no question covers is listed honestly (for example "11/14 covered") instead of implying full coverage.
不确定就标出来
Uncertainty is surfaced
判定强度不足的得分点会让整题变成「待复核」，给出分数区间，并且不计入知识点掌握度。
A weakly judged key point turns the question into "needs review" with a score range, excluded from mastery tracking.
关键设计
Design decisions
- 每个得分点的问题必须自带它所判的那一点。判定请求里的多个问题是互相独立的，如果所有问题共用同一段指令、只靠问题名字区分，模型根本不知道自己在判哪一点——判定质量会在没有报错的情况下悄悄崩掉。这是实现时踩到的第一个真问题。
- Each rubric question must carry the point it judges. The questions inside one request are independent; if every question shares one instruction and only the key name differs, the model cannot know which point it is deciding. Quality collapses silently, with no error to notice.
- 不确定就别装懂。判定强度低于阈值的关键点会让整题变成「待复核」，给出分数区间，并且不计入知识点掌握度。
- Never fake certainty. A key point judged too weakly turns the whole question into a review item with a score range, excluded from mastery.
- 扣分规则在代码里，而不在模型里。「矛盾」和「编造」两项概率由模型给出，但怎么扣、扣多少由代码决定。
- Penalties live in code, not in the model. The model provides probabilities for contradiction and fabrication; the code decides how much to subtract.
- 引擎可替换。判定层只有一个接口（
DecisionEngine）：默认是通用大模型判定，TypeSafe Jev 与词面基线是另外两个实现。供应商涨价、模型下线、或者出现更好的判定方式时，换实现不需要动业务代码。 - The engine is swappable. Grading depends on one interface (
DecisionEngine): a general-LLM judge by default, with TypeSafe Jev and a lexical baseline as the other two implementations. A price change, a model sunset or a better judging method replaces one file, not the product. - 成本与准入可控。每人每天限量，自带密钥（BYOK）不占平台额度；一道主观题的全部得分点合并进一次判定请求，而不是每个点开一轮对话，所以「逐点判定」在成本上站得住。
- Cost and access stay bounded. Daily quotas apply, BYOK calls bypass platform quota, and all the rubric points of one written answer go into a single request instead of a chat turn per point — which is what makes point-by-point grading affordable.
- 数据层不是「能跑就行」。Postgres 表之间带
ON DELETE CASCADE外键，出题、删材料、删账号等关键多步写入在单事务里提交；每日额度用条件更新原子占用，避免并发请求同时穿过闸门。数据库当前按 Neon Free Tier 起步，备份、监控与升级触发写进了docs/operating.md。 - The data layer is not "good enough to run". Postgres tables carry
ON DELETE CASCADEforeign keys; critical multi-step writes such as creating an exam or deleting material/account commit in one transaction; daily quotas are reserved atomically with conditional updates. The database starts on Neon Free Tier, with backups, monitoring and upgrade triggers documented indocs/operating.md.
实测与边界
Evidence and limits
项目自带一个判定评测脚本和一份小规模金标准集（12 道主观题、42 个得分点，逐点人工标注）。它输出逐点准确率、Brier 分数、校准分桶与自一致性，并设了门槛：逐点准确率不低于 90%，且校准分桶单调。
The project ships an evaluation harness with a small golden set (12 written questions, 42 labelled rubric points). It reports per-point accuracy, Brier score, calibration buckets and self-consistency, with a gate of at least 90% accuracy and monotonic calibration.
有意思的是，把「离线演示引擎」放进去跑：逐点准确率 71.4%、Brier 0.216，校准分桶单调（0.6–0.8 档 60%、0.8–1.0 档 80%），但 12 道题里有 9 道因为判定强度不足被标成「待复核」。这正是要把「判定强度」做成一等公民的理由：词面重合能蒙对一些点，却给不出可用的把握程度，而这件事必须有数字，不能靠感觉。
Run the offline demo engine through it and you get 71.4% per-point accuracy with a Brier score of 0.216 and monotonic calibration — but 9 of the 12 questions are flagged for review because their judgment strength is too low. Word overlap can guess some points right while being unable to say how sure it is, and that is why the gate is a number rather than an opinion.
| 适合 | Good fit | 不适合 | Poor fit |
|---|---|---|---|
| 有明确要点的背诵型材料（面试八股、法条、术语、流程） | Memorisation-heavy material with explicit points | 需要执行或符号验证的题（复杂计算、代码正确性） | Tasks needing execution or symbolic checking |
| 要点式简答：判断「说到了没有」 | Point-based written answers | 创新写作、开放论述的“好坏”评价 | Judging the quality of open-ended writing |
| 量大、需要成本的批量判定 | High-volume, cost-sensitive grading | 需要模型给出解释理由的场景（Jev 只给概率） | Scenarios demanding a written justification |
为什么叫「点睛」
Why it is called Eyedot
它原来叫「Jev 备考」——拿判定引擎的名字当产品名。这其实是个陷阱：引擎在架构上是可替换的（DecisionEngine 有三个实现），名字却焊死在一家供应商上；而且中文用户念不出 Jev，搜 Jev 搜到的也是那个模型，不是这个产品。
It used to be called "Jev Exam" — the product named after its grading engine. That is a trap: the engine is swappable by design (three DecisionEngine implementations), while the name would have been welded to one vendor, and nobody searching for the product would have found it behind the model's own results.
改叫「点睛」，是因为画龙点睛说清了它真正在做的事：重要的不是总分，是缺的那一点。这个名字也直接画进了图标——左边三条长短不一的横线是材料里的要点（长度不一，因为要点本来就不等长），右边三个圆是一次判定的三种结果：命中、未命中、待复核；其中命中的那一点是琥珀金，那是整张图里唯一被点亮的颜色，也是吉祥物点头顶悬着的那一点。把「待复核」留成第三只圆，则是提醒自己：不确定必须是第一公民，不能做成一个假装确定的分数。
The new name says what the product actually does: the score is not the point, the missing point is. It is drawn into the icon too — three bars of unequal length are the points in your material, and the three circles are the three outcomes of one judgment: hit, miss, needs review. The hit is the only warm colour in the whole mark, the same amber as the dot floating above the mascot. Keeping "needs review" as a third circle is a reminder that uncertainty is a first-class citizen here, not an exception hidden behind a confident-looking number.
怎么用起来
Getting started
有两种用法。Web 应用是完整闭环：克隆仓库、装依赖、填一个模型密钥（平台统一出题与判定，也可以自带密钥），本地跑起来即可。没有密钥也能运行，只是会退化成离线演示模式，界面上会明确标注。
There are two ways to use it. The web app is the full loop: clone the repository, install dependencies, provide one model key (the platform can grade and generate for you, or bring your own), and run it locally. It also runs without any key, degrading to a clearly labelled offline demo mode.
另一种是 Agent Skill：把仓库克隆到 Codex / Claude Code 的 skills 目录，然后在对话里说「用这份材料考我」。出题由 Agent 完成，校验、判定与报告渲染由命令行完成，不需要服务器和数据库。
The other is an Agent Skill: clone the repository into the skills directory of Codex or Claude Code and ask it to quiz you on a material. The agent writes the questions; verification, grading and report rendering run from the command line, with no server or database.
判定能力还做成了命令行与 MCP：npx eyedot verify / grade / render，或把 MCP server 挂进 Agent（四个工具：校验、出模板、判分、渲染）。这样同一套判定实现既能用在网页里，也能被 Codex、Claude Code、Cursor 直接调用——不用自己拼 shell 命令。
The same grading capability also ships as a CLI and an MCP server: npx eyedot verify / grade / render, or register the MCP server so an agent can call four tools (verify, template, grade, render) directly instead of assembling shell commands.
npx eyedot verify --material m.md --exam exam.json --strict
npx eyedot verify --material m.md --exam exam.json --strict
npx eyedot grade --answers answers.json --engine auto
npx eyedot grade --answers answers.json --engine auto
claude mcp add eyedot -- npx -y eyedot@latest mcp
claude mcp add eyedot -- npx -y eyedot@latest mcp
