从出题、审核到自适应训练，一份完整操作指南。

01 · 系统特点与快速开始

这套系统把资料出题、人工审核、题库维护、逐题自适应、信心诊断和间隔复习连在一起。AI 用于生成有来源的题目草稿；选下一题由服务器按照已经保存的答题证据计算，训练期间不依赖 AI 临时猜测学生水平。

01 · What the system does

The system connects source-backed drafts, human review, bank maintenance, question-by-question adaptation, confidence checks and spaced review. AI creates question drafts. The server selects the next question from saved performance evidence; it does not ask an LLM to guess a student's ability during practice.

02 · 怎样用 AI 生成、审核和发布题目

1. 准备原文。粘贴 150–40,000 个字符的学习资料。填写资料标题，以及页码、章节或其他定位信息。内容不足时，模型可以返回少于要求数量的题目。
2. 设定任务。选择七大考试领域之一，选择难度 1（记忆）、2（应用）或 3（较复杂推理），每次生成 1–10 道草稿。
3. 生成草稿。点击生成并等待结果。系统保存批次、原文和生成状态。同一个生成请求重试不会再调用模型、重复生成一批题。
4. 检查自动校验。系统检查四个选项是否齐全且不同、答案是否为 A–D、解析与各选项说明是否完整、引用是否确实出现在所给原文中，以及与现有题目的重合情况。
5. 逐题人工审核。打开草稿，对照原文核对题干、唯一最佳答案、每个干扰项解释、知识点、领域和难度。自动检查来源原文并不等于医学内容已经正确。
6. 修改并保存。修正后保存草稿，重新查看校验结果。不能修复的题目选择拒绝。另一位工作人员已改过的版本会要求重新加载。
7. 确认发布。勾选“已审核来源、答案和警告”的确认项，再点击发布。通过校验的题目才会写入现有题库；系统记录审核人和发布记录。
8. 检查训练结果。发布后用对应领域的智能训练检查题目。新训练会读取最新题库；已开始的训练使用生成该题时保存的版本，避免途中修改正确答案。

“生成成功”表示草稿已经生成，不表示自动进入学生题库。发布必须由有权限的工作人员明确完成。

02 · Generate, review and publish questions

1. Prepare the source.Paste 150–40,000 characters and enter the title with a page, section or other locator. Insufficient material may produce fewer questions.
2. Set the task.Select a domain, difficulty 1 (recall), 2 (application) or 3 (more involved reasoning), and 1–10 drafts.
3. Generate.The batch retains the source and status. Retrying the same generation request does not create a second model call or duplicate batch.
4. Inspect checks.Checks cover four distinct options, a valid A–D key, explanations for all options, a quotation that occurs in the source and overlap with existing questions.
5. Review each draft.Compare the question, one best answer, distractors, topic, domain and difficulty against the source. A quotation match is not proof of medical accuracy.
6. Edit and save.Save corrections and inspect validation again, or reject the draft. Stale versions require a reload.
7. Publish explicitly.Confirm that the source, key and warnings were reviewed. Only a valid, explicitly approved draft enters the existing bank; the action is audited.
8. Check in practice.New selections use the updated bank. An already-served question keeps its original snapshot so its answer key cannot change mid-session.

A successful generation produces drafts. It does not automatically publish questions to students.

03 · 怎样更新现有题目和知识点

1. 打开“现有题库管理”或“七大领域题目”，按领域和题目编号查找。先确认正在修改的是正确的记录。
2. 修改题干、A–D 四个选项、正确答案和解析。答案和选项必须一起核对；不要只改选项文字而保留错误的答案键。
3. 检查领域、难度和知识点。新题目工厂中的知识点用于知识点掌握度；已有题库的知识点通过原有知识点关联读取。
4. 保留可追溯来源。若修改了题目的事实依据，重新核对引用和解释。
5. 保存后进入学生训练验证。新选中的题目读取更新；已经发出的题目保留当时的快照。
6. 处理批量导入时，先用少量记录核对格式、领域和答案，再扩大数量；导入完成后复核，避免把未审核内容直接交给学生。

知识点分类的实际限制

自适应使用题库已有的领域和知识点标签。错误的标签会影响选题，因此老师需要持续维护分类。没有具体知识点的题目仍参与领域训练，但不会伪造一个精细知识点诊断。

03 · Update existing questions and topics

1. Use Question Bank or Seven-domain Catalog to locate the domain and question number.
2. Edit the stem, four options, key and explanation together. A changed option may require a changed key.
3. Check the domain, difficulty and topic. Factory topics and the existing bank's linked knowledge points support concept evidence.
4. Keep a traceable source and recheck explanations when factual content changes.
5. Save and inspect the item in practice. Newly selected items use the current bank; served items retain their original snapshots.
6. For bulk imports, check a small sample and field mapping before expanding the batch, then review the resulting content.

Classification matters

Adaptation uses the bank's domain and concept labels. Incorrect labels affect selection. Items without a meaningful topic can support domain practice without inventing a precise concept diagnosis.

难度分级：让升降难度有题可选

迁入的 8,145 道题原来全部标为“中等”。这时系统可以调整领域、知识点和复习优先级，也可以计算新的目标难度，但题库没有基础或进阶标签，不能声称下一题已经按难度升降。

1. 打开“难度分级审核”，选择领域，每批分析 5 或 10 道已有题目。
2. AI 提出难度及中英文理由，不修改原题和答案。
3. 老师按认知任务判断：1 是直接记忆，2 是直接应用，3 是多步骤推理或优先判断；不能只按题目长短分级。
4. 有该领域题库权限的老师选择最终等级，勾选审核确认并生效。错误建议可以拒绝。
5. 新训练立即读取确认后的难度。后台显示各等级的真实题目数，继续补足缺少的领域和难度。

题目内容改变后，旧分级建议会被标记为过期，需要重新分析。已经发出的训练题使用原快照。

Difficulty review: provide a usable range

All 8,145 migrated questions were labeled intermediate. Domain, concept and review adaptation can still work, but the bank needs reviewed level-1 and level-3 items before the trainer can actually select that full range.

1. Open Difficulty Review, choose a domain and assess 5 or 10 existing items.
2. AI proposes a level with English and Chinese reasons, preserving the original question and answer.
3. Review the cognitive task: direct recall (1), straightforward application (2), or multi-step reasoning / prioritization (3). Length alone is not difficulty.
4. A teacher with question-bank permission for the domain selects the final level, confirms the review and applies it, or rejects the proposal.
5. New selections use confirmed labels immediately. Monitor actual counts by level and fill missing domain / difficulty combinations.

A changed source question invalidates its old suggestion. Served training items retain their original snapshots.

04 · 自适应怎样真正实现

这不是固定题序，也不是每次随机抽题。每次答案保存后，服务器更新学习证据，再给所有符合条件的候选题重新评分，选择分数最高的一题。本次训练已经用过的题目被排除。

第一步：更新领域和知识点证据

每个领域从 α = 1，β = 1 开始。估计掌握度为 α ÷ (α + β)。答对增加 α；答错增加 β。信心程度改变这次回答的权重：

04 · How adaptation actually works

The sequence is neither fixed nor a fresh random draw.After each saved answer, the server updates evidence, scores every eligible candidate and selects the highest-scoring item. Questions already used in the session are excluded.

Step 1: update domain and concept evidence

Each domain starts with α = 1, β = 1. Estimated mastery is α ÷ (α + β). Correct answers increase α; errors increase β. Confidence controls the weight:

第二步：确定目标难度

领域掌握度低于 45% 时目标为难度 1；45% 到不足 75% 为难度 2；达到 75% 为难度 3。系统寻找接近目标难度的题目，而不是简单地规定“答错就下一题必然变简单”。题库中必须有对应难度，才能选到对应题目。

第三步：重新排列候选题

• 领域薄弱程度：掌握度越低，优先级越高。
• 难度匹配：越接近目标难度，得分越高。
• 到期复习：优先返回已经到复习时间的题目；“非常确定却答错”额外加权。
• 知识点薄弱程度：相同知识点的证据影响该知识点相关候选题。
• 初始覆盖：前 30 次回答优先补足七大领域的基础样本，避免一直停在一个领域。
• 考前重点：保存的考试日期在未来 14 天内时，已识别的弱领域额外加权。
• 防止重复：本次训练用过的题目不再选；刚做过而尚未到期的题目降低优先级。

第四步：显示真实选择依据

学生答题后展开“这次回答怎样改变后续训练”，可查看该领域掌握度前后变化、增加的正向／错误证据、实际下一题领域与难度，以及各项选题分值。这些内容来自刚刚保存的训练结果。

一个具体例子

初始 α=1、β=1，掌握度 50%，目标难度 2。如果“非常确定”却答错，β 变为 2.5，掌握度降为 28.57%，该领域目标难度变为 1，并安排 6 小时后复习。下一题仍要与其他领域、到期复习和本次排除题综合比较，所以可能先补其他领域的基础样本。

这是带权重的掌握度与复习选题系统，不是经过官方校准的 IRT/CAT 考试，也不提供“通过概率”。题目难度和知识点来自已有标签，需要老师审核维护。

Step 2: determine a target difficulty

Mastery below 45% targets level 1; 45% to below 75% targets level 2; 75% and above targets level 3. The engine favors available questions near that target. A wrong answer does not force the very next question to be easier regardless of other needs.

Step 3: rank candidates again

• Domain gap:weaker domains receive more priority.
• Difficulty fit:closer matches to the target receive a higher score.
• Due review:due questions gain priority, with extra weight for confident errors.
• Concept gap:evidence for a concept affects related candidates.
• Baseline coverage:the first 30 answers encourage coverage across the seven domains.
• Exam focus:weak domains gain additional weight during the 14 days before the saved exam date.
• Exposure:items used in this session are excluded; recent, not-yet-due items are downranked.

Step 4: show the decision

After answering, expand “How this answer changes your training” to see the before/after mastery evidence, actual next domain and difficulty, and the selection factors. These values come from the saved live run.

A concrete example

Starting at α=1 and β=1 gives 50% mastery and target difficulty 2. A very-confident error changes β to 2.5, producing 28.57% mastery and target difficulty 1 for that domain, with review in 6 hours. Other domains, due reviews and exclusions can still determine which item is next.

This is a weighted mastery and review selection system, not an officially calibrated IRT/CAT exam or a pass-probability model. Teachers maintain the source difficulty and concept labels.

06 · AI 设置、失败处理和更新

1. 超级管理员打开“AI 接口设置”。在受保护表单中保存密钥，选择可用模型并验证连接。接口密钥不应写在题目、备注或对话里。
2. 打开题目工厂，看接口配置状态。状态可用后，用少量、有明确来源的学习资料试生成。
3. 若显示接口不可达、HTTP 错误或返回格式不正确，检查接口启用状态、密钥、账户额度和模型可用性。失败不会自动发布题目。
4. 若草稿生成了但校验未通过，先看具体错误：引用是否来自原文、解析是否完整、选项是否重复，再编辑保存。
5. 系统支持连接已保存的模型接口；Dify 工作流也有适配入口，但必须另行配置满足题目输出格式的工作流，不能把“装了 Dify”当作“所有出题流程已经配置完”。

怎样理解“训练 AI”

这里的日常“训练”指学生训练和题库质量改进，不是自动微调大模型。老师通过优质来源、审核、知识点分类和题目更新提高题库质量；系统通过学生真实答题不断更新学习证据。

06 · AI setup, failures and updates

1. The owner opens AI connection settings, saves a key in the protected form, selects an available model and verifies it.
2. Check the provider status in Question Factory and try a small, clearly sourced passage.
3. For connection, HTTP or format errors, inspect enablement, key, available credit and model availability. A failed request publishes nothing.
4. For invalid drafts, inspect the exact quotation, explanation and option errors; edit and save again.
5. The saved model provider is supported. A Dify workflow adapter is also available, but a compatible workflow must be configured separately.

What “training AI” means here

Daily training means student practice and question-bank improvement, not automatic model fine-tuning. Teachers improve source quality, review and classification; the system updates learning evidence from actual answers.

07 · 学生怎样练习、复习和查看计划

1. 使用已有学生账户登录，完成所需注册资料。进入“智能训练”。会员权限和每日免费次数与现有系统一致。
2. 在学习计划中填写每天可用分钟数和预计考试日期。首次先完成几组短训练，让七大领域逐步建立样本。
3. 每题选择一个答案，再选择“猜测／比较确定／非常确定”，提交后查看解析和后续训练依据。
4. “专项训练”限定领域；“到期复习”只练已到时间的题；“节奏冲刺”提供逐题计时提示。
5. 答错通常 12 小时后复习；非常确定却答错为 6 小时；猜对为 12 小时。稳定答对后的复习间隔依次为 1、3、7、14、30、60 天。
6. 查看真实领域表现、待复习题、知识点优先级和历史训练。练习指标在至少 20 次近期回答、且至少 3 个领域各有 3 次回答后显示。

练习指标由掌握度 35%、领域覆盖 25%、近期正确率 25%、答题节奏 15% 组成，只说明本系统的练习情况，不能当作考试通过概率。

07 · Student practice, review and plans

1. Sign in with the existing student account, complete required registration and open Smart Training. Existing membership and daily limits apply.
2. Save daily study minutes and an optional exam date. Start with short sessions to build evidence across domains.
3. Select one answer and a confidence level, then inspect the explanation and adaptation details.
4. Focused practice limits the domain; due review uses due items; a pacing sprint adds timing prompts.
5. Errors usually return in 12 hours, confident errors in 6 hours and guessed-correct answers in 12 hours. Secure correct recall advances through 1, 3, 7, 14, 30 and 60 days.
6. Inspect domain evidence, review priorities, concepts and history. The practice index needs at least 20 recent answers and three domains with three answers each.

The index combines mastery (35%), coverage (25%), recent accuracy (25%) and pacing (15%). It is a practice signal, not a pass prediction.

08 · 角色、菜单和每日工作流程

超级管理员：配置 AI、分配员工权限、管理会员和系统。出题老师：需要“AI 教学工具”权限才能生成、审核和发布。题库老师：按授权领域维护题目。学生：训练和查看自己的记录。

后台所有管理页使用统一菜单、右上角中文／English 切换和“本页指南”。语言选择保存在当前浏览器；更换页面后继续使用同一种界面语言。切换语言不修改题目原文或学生数据。

老师每日建议流程

1. 查看训练控制中心，确认题库、接口和待审核草稿状态。
2. 审核新草稿，修复来源、答案和分类问题后发布。
3. 维护旧题中的错误或不明确解析。
4. 通过有权限的学习报表查看领域薄弱点，安排有针对性的课程复习。
5. 需要理解选题逻辑时，查看学生答题反馈中的实际依据和本指南的算法演示。

08 · Roles, navigation and daily operations

Owner:configure AI, staff permissions, memberships and systems. Question author:needs AI tools permission to generate, review and publish. Question editor:maintains authorized domains. Student:trains and views personal records.

Administration pages share navigation, a Chinese / English switch and a page guide. The browser retains the language across pages. Interface switching does not rewrite question text or student data.

A teacher's daily workflow

1. Inspect bank, provider and draft status in the training control center.
2. Review drafts and resolve source, key and classification problems before publication.
3. Maintain unclear or incorrect existing items.
4. Use authorized learning reports to guide domain-specific review teaching.
5. Inspect live answer feedback and the algorithm demonstration when checking selection behavior.