微信小程序 · 开发中 · 敬请期待 WeChat Mini Program · In Development · Coming Soon
ExamFlow 是一款正在开发中的微信考培刷题小程序：试卷自动切片入库，AI 变造命题、双盲校验后进入题库；再配合错题本与「一键举一反三」，把刷过的每一题都吃透。 ExamFlow is a WeChat mini program for exam practice, now in development: past papers are sliced into a structured bank automatically, AI-derived questions pass a double-blind check before entering, and a wrong-question book plus one-tap similar-question drills help you truly master every question you answer.
产品形态示意（纯 CSS 绘制）· 实际界面以正式发布版本为准 Stylized UI sketch (pure CSS) · the released app may differ
小程序端共三个页面：首页考种导航、卡片式刷题、错题本。围绕它们，是这些已经实现的能力（在本地联调中全部可跑通）。 The mini program has three pages — exam-type home, card-based quiz, wrong-question book — and these capabilities, all working in local integration today.
按考种随机抽 10–20 题组卷；卡片式作答，单选 / 多选 / 判断实时判分（多选须集合完全一致），交卷即得得分汇总与正确率。 Pull 10–20 random questions per exam type and answer card by card. Single, multiple and true/false questions are graded instantly — multiple-choice answers must match exactly as a set — with a score summary when you finish.
输入知识点关键词，跨考种检索同考点题目集中突破；每道题都带考点标签，薄弱点逐个击破。 Enter a knowledge-point keyword to pull matching questions across every exam type. Each question carries its knowledge-point tag, so weak spots get picked off one by one.
答错自动进本（复习次数 +1），答对自动置为已掌握；「待攻克 / 已掌握 / 全部」三态过滤，支持去重练与一键清零重练。 Wrong answers file themselves in (review count +1) and correct ones mark questions as mastered. Filter by pending, mastered or all, then de-duplicate and reset the whole book for a fresh round.
对任意一题，AI 提取核心考点、更换背景场景、加入强干扰项，实时变造出同类新题；无库存时后台异步生成（HTTP 202 + 前端每 2 秒轮询），做完原题立刻能练同类。 For any question, the AI extracts the core knowledge point, swaps the scenario and adds strong distractors to mint fresh variants. If none are in stock, the backend generates them asynchronously (HTTP 202 + 2-second client polling), so the follow-up drill is ready right after you finish the original.
核心亮点The highlight把当前考种随机抽 20 题生成可打印 PDF：题目在前、答案与解析在后，中文字体内嵌，可在微信里下载、预览与转发。 Export 20 random questions of the current exam type as a printable PDF — questions first, answers and analyses at the back, embedded Chinese font — downloadable, previewable and shareable inside WeChat.
可打印 · 可转发Printable & shareable题干与详解完整渲染 Markdown 排版；LaTeX 数学命令先离线转换为 Unicode 数学符号（√x、½、R₁ × R₂…），零外网依赖，真机与审核都稳，异常时自动回退纯文本。 Stems and analyses render full Markdown, and LaTeX math is converted offline into Unicode symbols (√x, ½, R₁ × R₂…) before rendering — zero network dependency, safe on real devices and in review, with automatic fallback to plain text.
从试卷文件到可刷的题，中间没有人工搬运。FastAPI 后端用三个 Agent 串起切片、命题与盲测校验。 From raw paper files to drillable questions with no manual copywork in between. A FastAPI backend chains three agents for slicing, deriving and blind verification.
上传 .pdf / .txt / .md 试卷，自动清洗页眉页脚与干扰字符；LLM 强制输出严格 JSON（json-repair 降级 + 最多重试 3 次），经 Pydantic 校验后入库为 original 真题。 Upload a paper in .pdf / .txt / .md; headers, footers and noise are scrubbed automatically. The LLM is forced to emit strict JSON (json-repair fallback, up to 3 retries), validated by Pydantic, then stored as an original question.
POST /api/pipeline/extract命题 Agent 提取原题核心考点，更换背景场景、保留考点、加入强干扰项，生成 3–5 道「同考点、不同皮」的高仿衍生题。 The deriver agent extracts the original's core knowledge point, swaps the scenario, keeps the point and adds strong distractors to create 3–5 look-alike questions.
POST /api/pipeline/derive盲测 Agent 不看标准答案、独立作答；与标准答案一致才允许入库（source_type = derived 并记 parent_id），不一致则弃用重来，整体最多 3 轮。 A separate blind-tester agent answers without ever seeing the key. Only questions whose answer agrees with the official one enter the bank as derived (with parent_id recorded); the rest are discarded and retried, up to 3 rounds.
blind_tester agent随机组卷接口按考种 / 知识点抽题，交卷判分并同步错题本；衍生题无库存时返回 202 后台生成，前端轮询取题，练完即闭环。 The random-paper API draws by exam type or knowledge point, grading on submit and syncing the wrong book; missing derived questions trigger a 202 background job the client polls — closing the loop right after practice.
GET /api/exams/randomMock 模式（LLM_MOCK=true）不需要任何 API Key，就能完整跑通切片、衍生命题与盲测闭环，适合演示与自动化测试；真实模式接入任意 OpenAI 兼容端点（DeepSeek / Qwen / Moonshot 等），带 60 秒超时与指数退避重试，未配置 Key 时返回清晰的中文错误而非崩溃。 Mock mode (LLM_MOCK=true) runs the entire pipeline — slicing, deriving, blind testing — with zero API keys, ideal for demos and automated tests. Real mode plugs into any OpenAI-compatible endpoint (DeepSeek / Qwen / Moonshot…) with a 60-second timeout, exponential-backoff retries, and a clear Chinese error message instead of a crash when the key is missing.
下面这道单选题及其解析，逐字来自当前题库（GAOKAO-Bench 高考数学真题，题目 ID 1355，经切片流水线入库）。LaTeX 源文本已按小程序同款规则离线转换为 Unicode 数学符号——你看到的就是小程序里的样子。 The question and analysis below come verbatim from the current bank (GAOKAO-Bench real gaokao math, question ID 1355, imported via the slicing pipeline). The LaTeX source has been converted offline to Unicode math symbols exactly as the mini program renders it.
（5 分）曲线 y = x / (x + 2) 在点 (−1, −1) 处的切线方程为（ ） (5 pts) The tangent line to the curve y = x / (x + 2) at the point (−1, −1) is ( )
在小程序里，这道题答错会自动进入错题本，并能一键举一反三生成「同考点、换场景」的新题。 In the mini program, a wrong answer files this question into the wrong book, and one tap derives fresh questions with the same knowledge point in a new scenario.
以下数字直接读取自本地题库（2026-10-24 实测），其中 3 道为已通过双盲校验的 AI 衍生题；题库随导入持续增长。 Figures below are read directly from the local bank (measured 2026-10-24); 3 of them are AI-derived questions that passed the double-blind check, and the bank keeps growing with every import.
公务员行测逻辑推理真题数据集（约 8700 题），已写好转换器接通并导入 1000+ 题。 Open dataset of civil-service logical-reasoning questions (~8,700); a converter is wired up and 1,000+ questions already imported.
2010–2022 高考真题集；已导入高考数学真题 50 题、高考物理真题 50 题，均带真题解析。 2010–2022 gaokao papers; 50 real math questions and 50 real physics questions imported, each with its official-style analysis.
已映射 6 个学科：高等数学、大学物理、计算机综合、机器学习、考研政治，logical 并入行测逻辑真题（按题干去重）。 Six subjects mapped in: calculus, college physics, CS fundamentals, machine learning and postgrad politics, with logical merged into the civil-service logic type (deduplicated by stem).
考研、四六级、行测、高等数学、数据结构、大学生竞赛、软考架构师等科目的自建模拟卷与演示种子题。 Hand-built mock papers and demo seed questions across postgrad entrance, CET, civil service, calculus, data structures, student contests and the software-architect exam.
以下是仓库提交记录（2026 年 10 月）能印证的真实进度，以及接下来的计划。 Real progress as evidenced by the repository's commit history (October 2026), plus what comes next.
FastAPI 后端流水线 + 小程序三个页面，端到端冒烟自检通过。 FastAPI pipeline backend + three mini-program pages, with the end-to-end smoke test passing.
接入 towxml 渲染题干与详解，LaTeX 离线转 Unicode，Markdown 完整排版。 towxml integrated for stems and analyses, LaTeX converted offline to Unicode, full Markdown layout.
新增 8 考种 84 题样例卷，首页按分类关键词自动归类的考种导航上线。 Sample papers for 8 more subjects (84 questions) and keyword-based category navigation on the home page.
接入 LogiQA 数据集转换器，并沉淀了一份真题获取渠道指南（官方 / 开源数据集 / 社区）。 LogiQA dataset converter connected, plus a written sourcing guide (official / open datasets / community channels).
三页设计系统 token 化重做（配色 / 卡片 / 选项 / 导航），新增 PDF 试卷导出（中文字体内嵌）。 All three pages restyled on a tokenized design system (colors, cards, options, nav), and PDF paper export added with an embedded Chinese font.
题库已积累 16 考种 1452 题；继续打磨体验、扩充题库，为上线做准备。 The bank has grown to 16 exam types and 1,452 questions; polishing experience and expanding content while preparing for release.
引入 SM-2 / FSRS 算法（参考 open-spaced-repetition 开源实现），按遗忘曲线自动安排每日错题复习队列。 Adopt SM-2 / FSRS (following the open-spaced-repetition implementations) to schedule a daily wrong-question review queue along the forgetting curve.
按渠道指南陆续接入 LogiQA 2.0（中英双语 ~15k 题）、AGIEval、RACE、C-Eval 等数据集。 Following the sourcing guide: LogiQA 2.0 (bilingual, ~15k), AGIEval, RACE, C-Eval and more.
真实 LLM 模式的服务化部署、域名与 HTTPS，以及小程序提审发布。 Production deployment for real-LLM mode, domain and HTTPS, then submitting the mini program for review and release.