[{"data":1,"prerenderedAt":1448},["ShallowReactive",2],{"blog-list-zh":3},[4],{"id":5,"title":6,"body":7,"config":1432,"date":1433,"description":1434,"draft":1435,"extension":1436,"image":1432,"meta":1437,"navigation":1438,"path":1439,"seo":1440,"stem":1441,"tags":1442,"toolbar":1432,"translationKey":1446,"updated":1433,"__hash__":1447},"blog/zh/blog/zero-hallucination-qa.md","我是如何实现阅读器「零幻觉」问答的",{"type":8,"value":9,"toc":1396},"minimark",[10,18,33,36,41,48,53,58,76,81,94,99,133,136,140,158,165,169,184,189,226,233,237,250,275,280,397,415,422,424,428,435,450,457,477,483,485,489,492,498,500,504,527,537,598,601,612,622,629,631,635,642,648,655,659,666,674,681,685,695,738,749,755,757,761,775,783,790,793,835,845,853,859,866,870,879,885,896,898,902,908,912,919,923,946,956,958,962,968,1026,1032,1034,1038,1045,1068,1072,1092,1103,1105,1109,1120,1123,1146,1157,1163,1165,1169,1188,1199,1201,1205,1227,1238,1240,1244,1263,1269,1271,1275,1362,1369,1380],[11,12,13],"p",{},[14,15],"img",{"alt":16,"src":17},"封面：零幻觉问答","https://cdn.linghuxiong.com/resources/snapshots/ai-chat-cover.png",[19,20,21],"blockquote",{},[11,22,23,24,28,29,32],{},"本文分享 AI 阅读器 ",[25,26,27],"strong",{},"零幻觉问答"," 的工程实现：回答严格基于当前书籍原文，关键论述可 ",[25,30,31],{},"一键溯源"," 到具体段落。如果你也在做 AI 阅读、文档 QA 或 RAG 类应用，希望三次迭代的经验与最终架构能有所参考。",[34,35],"hr",{},[37,38,40],"h2",{"id":39},"一实践历程三个阶段的演进","一、实践历程：三个阶段的演进",[11,42,43,44,47],{},"零幻觉问答并非一开始就设计完备，而是在 ",[25,45,46],{},"成本、延迟和准确率"," 的拉扯中逐步演进的。下面按时间顺序回顾三个阶段，便于理解当前架构为何长成这样。",[49,50],"mermaid",{":config":51,"code":52},"config","flowchart%20LR%0A%20%20%20%20P1%5B%E9%98%B6%E6%AE%B5%E4%B8%80%EF%BC%9A%E5%85%A8%E6%96%87%E7%9B%B4%E5%A1%9E%5D%20--%3E%20P2%5B%E9%98%B6%E6%AE%B5%E4%BA%8C%EF%BC%9ALLM%20%E6%8F%90%E5%8F%96%E5%85%B3%E9%94%AE%E5%8F%A5%5D%0A%20%20%20%20P2%20--%3E%20P3%5B%E9%98%B6%E6%AE%B5%E4%B8%89%EF%BC%9A%E7%89%87%E6%AE%B5%E7%B4%A2%E5%BC%95%20%2B%20Tool%20%E6%A3%80%E7%B4%A2%5D%0A%20%20%20%20P1%20-.-%3E%7C%E6%85%A2%E3%80%81%E8%B4%B5%E3%80%81%E9%95%BF%E4%B9%A6%E4%B8%8D%E5%87%86%7C%20X1%5B%E6%B7%98%E6%B1%B0%5D%0A%20%20%20%20P2%20-.-%3E%7C%E4%B8%A2%E7%BB%86%E8%8A%82%E3%80%81%E4%BB%8D%E5%81%8F%E6%85%A2%7C%20X2%5B%E6%B7%98%E6%B1%B0%5D%0A%20%20%20%20P3%20--%3E%7C%E5%BD%93%E5%89%8D%E6%96%B9%E6%A1%88%7C%20OK%5B%E9%9B%B6%E5%B9%BB%E8%A7%89%20%2B%20%E5%8F%AF%E6%BA%AF%E6%BA%90%5D",[54,55,57],"h3",{"id":56},"阶段一全文直塞-context最简单也最先暴露问题","阶段一：全文直塞 Context（最简单，也最先暴露问题）",[11,59,60,63,64,67,68,71,72,75],{},[25,61,62],{},"做法："," 用户打开一本书提问时，将提取出的 ",[25,65,66],{},"全部正文"," 放进 System Prompt 或 User 消息，交给对话模型作答。若全书超过约 ",[25,69,70],{},"40 万字符","，则 ",[25,73,74],{},"硬截断","——只保留前面一段，后续章节对模型不可见。",[11,77,78],{},[25,79,80],{},"优点：",[82,83,84,88,91],"ul",{},[85,86,87],"li",{},"实现成本极低，几乎不需要预处理；",[85,89,90],{},"短书、结构简单的文档效果尚可——模型确实「看到了整本书」；",[85,92,93],{},"交互简单：问就能答，没有「请先等待分析」的等待态。",[11,95,96],{},[25,97,98],{},"缺点（很快变得不可接受）：",[82,100,101,107,113,123],{},[85,102,103,106],{},[25,104,105],{},"响应慢","：每次提问都要把海量文本送进模型，首 Token 延迟和总耗时随书长线性恶化；",[85,108,109,112],{},[25,110,111],{},"Token 成本高","：同一本书每问一次就重复付一遍全文的输入费用；",[85,114,115,118,119,122],{},[25,116,117],{},"长书严重失真","：超过 40 万字符后被截断，后半本、附录、结论章节等于不存在，且 UI 往往 ",[25,120,121],{},"没有明确告知"," 已截断；",[85,124,125,128,129,132],{},[25,126,127],{},"检索粒度为零","：模型要在几十万字里「大海捞针」，容易漏细节，也更容易产生 ",[25,130,131],{},"看似合理、实则无据"," 的概括——阅读场景最忌讳这类幻觉。",[11,134,135],{},"阶段一适合验证 MVP，不适合作为产品级方案。",[54,137,139],{"id":138},"阶段二用轻量-llm-提取关键句压缩-context但压缩得太狠","阶段二：用轻量 LLM 提取关键句（压缩 Context，但压缩得太狠）",[11,141,142,144,145,148,149,152,153,157],{},[25,143,62],{}," 在提问前（或首次打开书时），用 ",[25,146,147],{},"成本更低的模型"," 对正文做一轮预处理：按 Spine 分章（或整书分段），抽取 ",[25,150,151],{},"关键句","，输出时保留 ",[154,155,156],"code",{},"[f文件-起始-结束]"," 形式的位置标记，再将摘录拼成较短文本，作为后续问答的 Context。",[11,159,160,161,164],{},"典型链路是 ",[25,162,163],{},"Extract → Cache → Chat","：先离线或按需跑一遍提取并落库，之后每次提问复用同一份「关键句合集」。这与很多文档 QA 原型里「先压缩文档、再拿压缩结果做 QA」的思路相同，也是我们在阶段二实际采用过的路线。",[11,166,167],{},[25,168,80],{},[82,170,171,178,181],{},[85,172,173,174,177],{},"每次提问送入模型的文本 ",[25,175,176],{},"明显缩短","，单次 Token 消耗较阶段一显著下降；",[85,179,180],{},"预处理结果可缓存，同一本书不必每次提问都重新提取；",[85,182,183],{},"已引入位置标记，为后续溯源打下基础。",[11,185,186],{},[25,187,188],{},"缺点（长书场景下依然扛不住）：",[82,190,191,197,207,216],{},[85,192,193,196],{},[25,194,195],{},"细节大量丢失","：「关键句」由模型主观筛选，论证链上的限定条件、反例等容易被丢掉，答案容易「正确但片面」；",[85,198,199,202,203,206],{},[25,200,201],{},"长书 Context 仍然偏大","：大部头作品即便只留关键句，拼接后的输入依然可观，",[25,204,205],{},"延迟和成本只是缓解，没有根治","；",[85,208,209,212,213,206],{},[25,210,211],{},"双重 LLM 误差","：提取阶段可能漏选，问答阶段又可能误读摘录，错误会 ",[25,214,215],{},"叠加",[85,217,218,221,222,225],{},[25,219,220],{},"静态 Context","：无论用户问的是某一章细节还是全书结构，送进模型的都是 ",[25,223,224],{},"同一份预提取文本","，无法按问题动态收窄范围。",[11,227,228,229,232],{},"这一阶段的教训很明确：",[25,230,231],{},"问题不在「有没有压缩」，而在「压缩是否按需、以及能否回到原文」","。",[54,234,236],{"id":235},"阶段三片段索引-tool-按需检索-原文回传当前方案","阶段三：片段索引 + Tool 按需检索 + 原文回传（当前方案）",[11,238,239,241,242,249],{},[25,240,62],{}," 基本思路参考了 ",[243,244,248],"a",{"href":245,"rel":246},"https://github.com/VectifyAI/PageIndex",[247],"nofollow","PageIndex","，相对阶段二，核心变化有三点：",[251,252,253,259,269],"ol",{},[85,254,255,258],{},[25,256,257],{},"预处理产物是结构化索引","（目录级摘要 + 精确字符 span），而不是把摘录直接当作问答 Context；",[85,260,261,264,265,268],{},[25,262,263],{},"每次提问由模型通过 Tool Calling 按需检索","，再 ",[25,266,267],{},"拉取带位置标记的原文"," 作答；",[85,270,271,274],{},[25,272,273],{},"System Prompt 与前端联动","，约束引用格式，并支持点击角标跳转、高亮原文。",[11,276,277],{},[25,278,279],{},"三阶段对比：",[281,282,283,302],"table",{},[284,285,286],"thead",{},[287,288,289,293,296,299],"tr",{},[290,291,292],"th",{},"维度",[290,294,295],{},"阶段一（全文直塞）",[290,297,298],{},"阶段二（关键句提取）",[290,300,301],{},"阶段三（当前）",[303,304,305,324,338,352,366,383],"tbody",{},[287,306,307,311,314,317],{},[308,309,310],"td",{},"单次提问 Context",[308,312,313],{},"全书（或截断后的前半本）",[308,315,316],{},"预提取关键句合集",[308,318,319,320,323],{},"仅与问题相关的少量 ",[25,321,322],{},"原文"," 片段",[287,325,326,329,332,335],{},[308,327,328],{},"长书准确性",[308,330,331],{},"超 40 万字符后严重下降",[308,333,334],{},"依赖提取质量，易丢细节",[308,336,337],{},"按目录/span 检索，不受全书长度硬截断",[287,339,340,343,346,349],{},[308,341,342],{},"响应速度",[308,344,345],{},"慢",[308,347,348],{},"略好，长书仍慢",[308,350,351],{},"检索 + 短 Context，明显更快",[287,353,354,357,360,363],{},[308,355,356],{},"Token 成本",[308,358,359],{},"极高",[308,361,362],{},"中等偏高",[308,364,365],{},"预处理摊销 + 按需付费",[287,367,368,371,374,377],{},[308,369,370],{},"溯源能力",[308,372,373],{},"弱（难标注出处）",[308,375,376],{},"有位置标记，但内容已是二次筛选",[308,378,379,380],{},"角标对应 ",[25,381,382],{},"真实原文 span",[287,384,385,388,391,394],{},[308,386,387],{},"工程复杂度",[308,389,390],{},"低",[308,392,393],{},"中",[308,395,396],{},"高",[11,398,399,402,403,406,407,410,411,414],{},[25,400,401],{},"为何停在阶段三："," 阅读场景的零幻觉，关键不是「让模型看过尽量多的字」，而是 ",[25,404,405],{},"「作答前必须拿到与问题相关的原文证据」","。阶段一、二都在 Context ",[25,408,409],{},"体积"," 上做文章；阶段三把链路拆成 ",[25,412,413],{},"「索引（预处理）→ 检索（Tool）→ 取证（原文）→ 作答（约束生成）」","，才同时兼顾准确率、成本与可溯源性。",[11,416,417,418,421],{},"下文展开 ",[25,419,420],{},"阶段三"," 的实现细节。",[34,423],{},[37,425,427],{"id":426},"二问题定义阅读场景下幻觉比普通-chat-更致命","二、问题定义：阅读场景下，幻觉比普通 Chat 更致命",[11,429,430,431,434],{},"普通 ChatBot 偶发错误，用户往往可以容忍。但在 ",[25,432,433],{},"书籍 QA"," 里，幻觉的代价更高：",[82,436,437,444,447],{},[85,438,439,440,443],{},"用户问的是 ",[25,441,442],{},"这本书"," 说了什么，不是问模型的 parametric memory；",[85,445,446],{},"一句似是而非的「书中观点」，可能误导笔记、引用甚至二次传播；",[85,448,449],{},"没有出处，用户无法核实，产品信任很难建立。",[11,451,452,453,456],{},"因此，「零幻觉」在工程上落地为三条 ",[25,454,455],{},"可执行"," 的规则：",[251,458,459,465,471],{},[85,460,461,464],{},[25,462,463],{},"书内问题必须先查书","：凡可能与当前书籍相关的问题，模型必须先走检索（Tool），再组织答案；",[85,466,467,470],{},[25,468,469],{},"答案必须可溯源","：关键结论附带原文位置标记，前端可解析并跳转高亮；",[85,472,473,476],{},[25,474,475],{},"查不到就说查不到","：书中没有的内容应明确告知，而不是用通用知识冒充「书中观点」。",[11,478,479,480,482],{},"下文按 ",[25,481,420],{}," 的数据流，说明上述规则如何落地。",[34,484],{},[37,486,488],{"id":487},"三整体架构预处理-工具检索-约束生成-可点击溯源","三、整体架构：预处理 → 工具检索 → 约束生成 → 可点击溯源",[49,490],{":config":51,"code":491},"flowchart%20TB%0A%20%20%20%20subgraph%20prep%20%5B%E7%A6%BB%E7%BA%BF%2F%E9%A6%96%E6%AC%A1%E9%A2%84%E5%A4%84%E7%90%86%5D%0A%20%20%20%20%20%20%20%20A%5B%E6%8C%89%E7%9B%AE%E5%BD%95%E6%88%96%E9%95%BF%E5%BA%A6%E5%88%87%E5%88%86%E5%85%A8%E4%B9%A6%5D%20--%3E%20B%5BLLM%20%E7%94%9F%E6%88%90%E7%89%87%E6%AE%B5%E6%91%98%E8%A6%81%5D%0A%20%20%20%20%20%20%20%20B%20--%3E%20C%5B%E6%9C%AC%E5%9C%B0%E6%8C%81%E4%B9%85%E5%8C%96%20Segment%20%E7%BC%93%E5%AD%98%5D%0A%20%20%20%20end%0A%0A%20%20%20%20subgraph%20ask%20%5B%E7%94%A8%E6%88%B7%E6%8F%90%E9%97%AE%5D%0A%20%20%20%20%20%20%20%20D%5B%E7%94%A8%E6%88%B7%E8%BE%93%E5%85%A5%E9%97%AE%E9%A2%98%5D%20--%3E%20E%7B%E5%B7%B2%E6%9C%89%20Segment%20%E7%BC%93%E5%AD%98%3F%7D%0A%20%20%20%20%20%20%20%20E%20--%3E%7C%E5%90%A6%7C%20F%5B%E6%8F%90%E5%8F%96%E5%85%A8%E6%96%87%20%2F%20%E8%AF%A2%E9%97%AE%E6%98%AF%E5%90%A6%E9%A2%84%E5%A4%84%E7%90%86%5D%0A%20%20%20%20%20%20%20%20F%20--%3E%20prep%0A%20%20%20%20%20%20%20%20E%20--%3E%7C%E6%98%AF%7C%20G%5B%E6%B3%A8%E5%86%8C%20Tool%20Calling%5D%0A%20%20%20%20end%0A%0A%20%20%20%20subgraph%20retrieve%20%5B%E5%B7%A5%E5%85%B7%E6%A3%80%E7%B4%A2%5D%0A%20%20%20%20%20%20%20%20G%20--%3E%20H%7B%E9%97%AE%E9%A2%98%E7%B1%BB%E5%9E%8B%7D%0A%20%20%20%20%20%20%20%20H%20--%3E%7C%E5%85%A8%E4%B9%A6%E6%A6%82%E8%A7%88%2F%E4%B9%A6%E8%AF%84%7C%20I%5Bget_full_book_segment_summaries%5D%0A%20%20%20%20%20%20%20%20H%20--%3E%7C%E5%85%B7%E4%BD%93%E4%BA%8B%E5%AE%9E%2F%E4%BA%BA%E7%89%A9%2F%E7%AB%A0%E8%8A%82%7C%20J%5Bget_related_segment_summaries%5D%0A%20%20%20%20%20%20%20%20J%20--%3E%20K%5BLLM%20%E4%BB%8E%E6%91%98%E8%A6%81%E7%9B%AE%E5%BD%95%E4%B8%AD%E9%80%89%E7%9B%B8%E5%85%B3%E7%89%87%E6%AE%B5%20ID%5D%0A%20%20%20%20%20%20%20%20K%20--%3E%20L%5B%E6%8C%89%20span%20%E6%8B%89%E5%8F%96%E5%8E%9F%E6%96%87%20%2B%20%E4%BD%8D%E7%BD%AE%E6%A0%87%E8%AE%B0%5D%0A%20%20%20%20%20%20%20%20I%20--%3E%20M%5B%E6%8B%BC%E6%8E%A5%E5%85%A8%E4%B9%A6%E7%89%87%E6%AE%B5%E6%91%98%E8%A6%81%5D%0A%20%20%20%20end%0A%0A%20%20%20%20subgraph%20answer%20%5B%E7%94%9F%E6%88%90%E4%B8%8E%E5%B1%95%E7%A4%BA%5D%0A%20%20%20%20%20%20%20%20L%20--%3E%20N%5BTool%20%E7%BB%93%E6%9E%9C%E5%9B%9E%E4%BC%A0%E6%A8%A1%E5%9E%8B%5D%0A%20%20%20%20%20%20%20%20M%20--%3E%20N%0A%20%20%20%20%20%20%20%20N%20--%3E%20O%5BSystem%20Prompt%20%E7%BA%A6%E6%9D%9F%E5%BC%95%E7%94%A8%E6%A0%BC%E5%BC%8F%5D%0A%20%20%20%20%20%20%20%20O%20--%3E%20P%5B%E6%B5%81%E5%BC%8F%E8%BE%93%E5%87%BA%E7%AD%94%E6%A1%88%20%2B%20%E4%BD%8D%E7%BD%AE%E8%A7%92%E6%A0%87%5D%0A%20%20%20%20%20%20%20%20P%20--%3E%20Q%5B%E6%B8%B2%E6%9F%93%E5%8F%AF%E7%82%B9%E5%87%BB%E5%BC%95%E7%94%A8%E8%A7%92%E6%A0%87%5D%0A%20%20%20%20%20%20%20%20Q%20--%3E%20R%5B%E7%82%B9%E5%87%BB%20%E2%86%92%20%E9%A2%84%E8%A7%88%E5%8E%9F%E6%96%87%20%E2%86%92%20%E8%B7%B3%E8%BD%AC%E9%AB%98%E4%BA%AE%5D%0A%20%20%20%20end",[11,493,494,495,232],{},"核心思路可以概括为：",[25,496,497],{},"不让模型「凭记忆答题」，而是让它「先取证、再作答、并标注出处」",[34,499],{},[37,501,503],{"id":502},"四预处理把整本书变成可检索的片段索引","四、预处理：把整本书变成可检索的「片段索引」",[11,505,506,507,510,511,514,515,518,519,522,523,526],{},"若每次提问仍采用 ",[25,508,509],{},"阶段一"," 的全文 Context，长书必然爆 Token，检索粒度也过粗。阶段三的解法是：用户首次对某本书发起 AI 对话时，后台异步跑 ",[25,512,513],{},"片段摘要任务","，按 ",[25,516,517],{},"目录结构"," 或 ",[25,520,521],{},"文本长度"," 将全书切成若干 ",[154,524,525],{},"Segment","，为每个片段生成摘要，并持久化到本地 IndexedDB。",[11,528,529,530,532,533,536],{},"每个 ",[154,531,525],{}," 在数据结构上包含摘要与 ",[25,534,535],{},"正文物理位置","：",[281,538,539,549],{},[284,540,541],{},[287,542,543,546],{},[290,544,545],{},"字段",[290,547,548],{},"含义",[303,550,551,565,578,588],{},[287,552,553,562],{},[308,554,555,558,559],{},[154,556,557],{},"startFileIndex"," / ",[154,560,561],{},"endFileIndex",[308,563,564],{},"Spine 文件索引（PDF 则每页一个文件）",[287,566,567,575],{},[308,568,569,558,572],{},[154,570,571],{},"startOffset",[154,573,574],{},"endOffset",[308,576,577],{},"字符级起止偏移",[287,579,580,585],{},[308,581,582],{},[154,583,584],{},"sequence",[308,586,587],{},"线性阅读顺序",[287,589,590,595],{},[308,591,592],{},[154,593,594],{},"title",[308,596,597],{},"对应目录标题",[11,599,600],{},"切分策略兼顾精度与成本：单目录正文不超过约 20KB 时只总结该节点；同级目录会合并成批（15KB～20KB）再调用 LLM；无目录的大块正文则按 3～4 万字符区间切段。",[11,602,603,604,607,608,611],{},"摘要生成时的 System Prompt 会要求 ",[25,605,606],{},"保留原文位置标记","（格式 ",[154,609,610],{},"[f数字-数字-数字]","），以便后续 Tool 回传原文时，位置信息与 spine 字符偏移一致。核心约束如下：",[613,614,620],"pre",{"className":615,"code":617,"language":618,"meta":619},[616],"language-text","如果总结内容与原文某段相关，须保留段末位置信息，格式 [f数字-数字-数字]（如 [f1-90-109]）。\n位置标记是整体，禁止修改、合并或省略其中的任何字符或数值。\n","text","",[154,621,617],{"__ignoreMap":619},[11,623,624,625,628],{},"预处理完成后，问答不再依赖「整书 Context」，而是依赖 ",[25,626,627],{},"结构化片段索引","——这是长书场景下零幻觉的工程前提。",[34,630],{},[37,632,634],{"id":633},"五位置标记体系把出处编码进文本","五、位置标记体系：把「出处」编码进文本",[11,636,637,638,641],{},"零幻觉不仅要求内容来自原文，还要求 ",[25,639,640],{},"出处可机器解析、可在 UI 中跳转","。我们采用内联位置标记：",[613,643,646],{"className":644,"code":645,"language":618},[616],"[f{fileIndex}-{startChar}-{endChar}]\n",[154,647,645],{"__ignoreMap":619},[11,649,650,651,654],{},"例如 ",[154,652,653],{},"[f5-123-165]"," 表示：第 5 个 Spine 文件（从 0 起算）中，字符偏移 123～165 的文本区间。",[54,656,658],{"id":657},"_51-标记如何写入正文","5.1 标记如何写入正文",[11,660,661,662,665],{},"正文提取层在输出片段时，为每个小段在段末写入 ",[154,663,664],{},"[f{fileIndex}-{start}-{end}]","。示意：",[613,667,672],{"className":668,"code":670,"language":671,"meta":619},[669],"language-typescript","const position = `[f${fileIndex}-${absOffset}-${absOffset + segment.length}]`;\nfileLines.push(segment.text.trim() + position);\n","typescript",[154,673,670],{"__ignoreMap":619},[11,675,676,677,680],{},"无论是预处理摘要还是 Tool 回传的原文摘录，位置信息都与 ",[25,678,679],{},"Spine 字符偏移"," 对齐，而不是让模型「估算页码」。",[54,682,684],{"id":683},"_52-对模型输出的约束","5.2 对模型输出的约束",[11,686,687,688,694],{},"在组装 System Prompt 时，我们单独约定了 ",[25,689,690],{},[691,692,693],"span",{},"Position Citation Rules","，核心五条：",[251,696,697,707,717,723,732],{},[85,698,699,702,703,706],{},[25,700,701],{},"标准格式","：必须使用 ",[154,704,705],{},"[f_fileIndex-startChar-endChar]","，三段数字缺一不可；",[85,708,709,712,713,716],{},[25,710,711],{},"只引用当前来源","：角标须 ",[25,714,715],{},"原样复制"," 自本轮 System/User 消息或 Tool 返回文本中的标记；",[85,718,719,722],{},[25,720,721],{},"禁止伪造","：不得自行计算、修改或编造位置；",[85,724,725,728,729,206],{},[25,726,727],{},"宁缺毋滥","：当前上下文没有合法标记时，正常作答即可，",[25,730,731],{},"不要输出任何位置标记",[85,733,734,737],{},[25,735,736],{},"紧跟论述","：标记须紧跟相关句段，禁止在文末堆砌引用清单。",[11,739,740,741,744,745,748],{},"前端展示前还会过滤模型偶发输出的 ",[25,742,743],{},"两段位"," 非法标记（如 ",[154,746,747],{},"[f1-293]","），避免无效角标进入 UI。",[11,750,751],{},[14,752],{"alt":753,"src":754},"引用溯源弹窗","https://cdn.linghuxiong.com/resources/snapshots/ai-chat.png",[34,756],{},[37,758,760],{"id":759},"六tool-calling先检索再回答","六、Tool Calling：先检索，再回答",[11,762,763,764,767,768,771,772,232],{},"当对话绑定某本书（存在 ",[154,765,766],{},"resourceId","，且 ",[154,769,770],{},"chatType === 'chat'","）时，每次生成前会向模型注册两个 Tool，并挂载对应的 executor。整体遵循 OpenAI 兼容的 ",[25,773,774],{},"function calling 循环",[54,776,778,779,782],{"id":777},"_61-get_related_segment_summaries-针对具体问题查片段","6.1 ",[154,780,781],{},"get_related_segment_summaries"," —— 针对具体问题查片段",[11,784,785,786,789],{},"适用于：概念、人物、情节、章节细节等 ",[25,787,788],{},"有明确检索意图"," 的问题。",[11,791,792],{},"流程简述：",[251,794,795,802,808,815,829],{},[85,796,797,798,801],{},"模型将用户口语 ",[25,799,800],{},"改写为书中可能出现的术语","（System Prompt 中的「Optimize Search Queries」）；",[85,803,804,805,206],{},"调用 Tool，传入 ",[154,806,807],{},"question",[85,809,810,811,814],{},"将所有片段摘要按 Token 预算 ",[25,812,813],{},"分批","（单批约 3 万 Token，最多 5 批）；",[85,816,817,818,821,822,825,826,206],{},"每批发起一次 ",[25,819,820],{},"独立的 LLM 请求","，从 ",[154,823,824],{},"{ id, title, summary }"," 列表中选出相关片段 ID（最多 5 个），返回 JSON，形如 ",[154,827,828],{},"{\"Thinking\":\"...\",\"answer\":[\"1\",\"3\"]}",[85,830,831,832,834],{},"根据选中 Segment 的 span，从 Spine ",[25,833,267],{},"（不是摘要），作为 Tool 结果回传。",[11,836,837,840,841,844],{},[25,838,839],{},"关键设计：Tool 回传原文，而非摘要。"," 模型作答时看到的是真实段落 + 内联 ",[154,842,843],{},"[f…]","，避免「摘要 → 再概括」带来的漂移。",[54,846,848,849,852],{"id":847},"_62-get_full_book_segment_summaries-全书概览类问题","6.2 ",[154,850,851],{},"get_full_book_segment_summaries"," —— 全书概览类问题",[11,854,855,856,789],{},"适用于：「总结全书」「点评这本书」「整体结构/主题」等 ",[25,857,858],{},"需要全局视野",[11,860,861,862,865],{},"按阅读顺序拼接所有片段的 ",[154,863,864],{},"summary"," 回传，避免逐段相关度筛选遗漏关键章节。",[54,867,869],{"id":868},"_63-system-prompt书优先工具优先","6.3 System Prompt：书优先、工具优先",[11,871,872,873,878],{},"绑定书籍时，System Prompt 注入 ",[25,874,875],{},[691,876,877],{},"Core Principles for Reading Assistant","，核心三条：",[613,880,883],{"className":881,"code":882,"language":618},[616],"1. Book First, Tool First\n   - 任何可能与书籍相关的问题，必须先调用工具检索；\n   - 答案必须主要依据检索结果，禁止不检索就编造「书中内容」。\n\n2. General Knowledge as Fallback Only\n   - 仅当：纯闲聊 / 用户明确要求不用书 / 工具无结果时，才可使用通用知识；\n   - 若书中没有，必须先声明「书中未提及此内容」，再补充通用知识。\n\n3. Direct Style\n   - 直入主题，禁止「根据提供的材料…」「综上所述…」等套话。\n",[154,884,882],{"__ignoreMap":619},[11,886,887,888,891,892,895],{},"生成层实现标准 Tool 循环：",[154,889,890],{},"tool_calls"," → 执行 executor → 追加 ",[154,893,894],{},"role: tool"," → 继续请求，直到输出最终文本。启用 tools 时关闭 thinking 通道，避免与 function call 协议冲突。",[34,897],{},[37,899,901],{"id":900},"七前端溯源从角标到原文高亮","七、前端溯源：从角标到原文高亮",[11,903,904,905,907],{},"模型输出的 ",[154,906,653],{}," 不会直接展示，在渲染层转为可点击引用。",[54,909,911],{"id":910},"_71-角标渲染","7.1 角标渲染",[11,913,914,915,918],{},"展示前将位置标记规范化为 Markdown 链接，例如 ",[154,916,917],{},"[1]([f5-123-165])","，再渲染为序号角标；同一位置多次出现时可去重，避免 UI 堆叠。",[54,920,922],{"id":921},"_72-点击交互","7.2 点击交互",[251,924,925,934,940],{},[85,926,927,930,931,933],{},[25,928,929],{},"首次点击","：解析 ",[154,932,843],{}," → 取 fileIndex 与字符偏移 → 从 Spine 原文提取文本 → 弹出预览（可带目录标题）；",[85,935,936,939],{},[25,937,938],{},"再次点击同一角标","：关闭弹窗；",[85,941,942,945],{},[25,943,944],{},"确认跳转","：打开阅读视图，按字符区间高亮。",[11,947,948,949,952,953,232],{},"从模型复制的标记到用户看到的原文，中间 ",[25,950,951],{},"不经 LLM 二次加工","，溯源链路全程 ",[25,954,955],{},"确定、可复现",[34,957],{},[37,959,961],{"id":960},"八边界情况与诚实降级","八、边界情况与诚实降级",[11,963,964,965,536],{},"零幻觉不等于「永远有答案」，而是 ",[25,966,967],{},"没有证据时不瞎编",[281,969,970,980],{},[284,971,972],{},[287,973,974,977],{},[290,975,976],{},"场景",[290,978,979],{},"行为",[303,981,982,990,1002,1010,1018],{},[287,983,984,987],{},[308,985,986],{},"片段摘要尚未生成",[308,988,989],{},"先提取全文做摘要",[287,991,992,995],{},[308,993,994],{},"Tool 检索无结果",[308,996,997,998,1001],{},"返回 ",[154,999,1000],{},"(No relevant segment excerpts found…)","，模型应声明书中未提及",[287,1003,1004,1007],{},[308,1005,1006],{},"模型输出了非法两段位标记",[308,1008,1009],{},"前端过滤，不展示无效角标",[287,1011,1012,1015],{},[308,1013,1014],{},"用户纯闲聊",[308,1016,1017],{},"System Prompt 允许脱离书籍，用通用知识回答",[287,1019,1020,1023],{},[308,1021,1022],{},"导出对话",[308,1024,1025],{},"可将角标转为阅读器深链接，便于分享或归档",[11,1027,1028],{},[14,1029],{"alt":1030,"src":1031},"对话导出","https://cdn.linghuxiong.com/resources/snapshots/ai-chat-export.png",[34,1033],{},[37,1035,1037],{"id":1036},"九设计取舍为什么不用向量-rag","九、设计取舍：为什么不用「向量 RAG」？",[11,1039,1040,1041,1044],{},"做文档 QA 的同行常会问：既然要做检索增强，为什么不走 ",[25,1042,1043],{},"Embedding + 向量库 Top-K"," 这条标准路线？",[11,1046,1047,1048,1051,1052,1055,1056,1059,1060,1063,1064,1067],{},"实际上 ",[25,1049,1050],{},"我们也在做 RAG","——每次回答前都会先查书、再生成。差别在于：社区语境里的 RAG 往往默认包含 ",[25,1053,1054],{},"向量化与相似度检索","；当前方案是 ",[25,1057,1058],{},"「片段索引 + Tool 按需拉原文」","（阶段三），",[25,1061,1062],{},"刻意不引入向量层","。下面从 ",[25,1065,1066],{},"架构约束"," 说明取舍，并非否定向量 RAG 的价值。",[54,1069,1071],{"id":1070},"界定范围不是不用检索而是不用向量检索","界定范围：不是不用检索，而是不用「向量检索」",[82,1073,1074,1083],{},[85,1075,1076,1079,1080,232],{},[25,1077,1078],{},"广义 RAG","：检索相关材料 → 再生成 → ",[25,1081,1082],{},"我们在做",[85,1084,1085,1088,1089,232],{},[25,1086,1087],{},"向量 RAG","：召回依赖 Embedding 相似度 → ",[25,1090,1091],{},"当前版本不做",[11,1093,1094,1095,1098,1099,1102],{},"全书预处理为 ",[25,1096,1097],{},"片段摘要索引","；提问时模型通过 Tool 选段，再 ",[25,1100,1101],{},"回传原文","。检索增强存在，但不依赖单独的 embedding 模型与向量索引维护。",[34,1104],{},[54,1106,1108],{"id":1107},"原因一支持自定义-llm-provider配置链路要尽量短","原因一：支持自定义 LLM Provider，配置链路要尽量短",[11,1110,1111,1112,1115,1116,1119],{},"产品允许用户自由接入 ",[25,1113,1114],{},"自有 API Key","、自定义 Base URL，或使用 ",[25,1117,1118],{},"本地 Ollama","——对话模型由用户自选，成本和数据路径可控。这对很多自托管、多模型对比的场景是硬需求。",[11,1121,1122],{},"叠加典型向量 RAG 后，集成面会明显变宽：",[82,1124,1125,1136,1139],{},[85,1126,1127,1128,1131,1132,1135],{},"除 ",[25,1129,1130],{},"Chat 模型"," 外，通常还需 ",[25,1133,1134],{},"Embedding 模型","（另一 model name，有时还是另一个 endpoint）；",[85,1137,1138],{},"Ollama 等本地部署还要单独拉 embedding 模型，并处理维度、接口兼容；",[85,1140,1141,1142,1145],{},"故障域变复杂：Chat 正常但 ",[25,1143,1144],{},"检索为空"," 时，可能是 embedding、索引或维度不一致，排查成本高于「单 Provider 全链路」。",[11,1147,1148,1149,1152,1153,1156],{},"当前方案里，",[25,1150,1151],{},"选段与作答共用同一套 Provider 配置","，避免「Chat 用 A、建索引用 B」。若你在做 ",[25,1154,1155],{},"可插拔 LLM"," 的应用，这往往比多几个点的召回率更重要。",[11,1158,1159],{},[14,1160],{"alt":1161,"src":1162},"自定义 AI 服务商","https://cdn.linghuxiong.com/resources/snapshots/ai-customize-providers.png",[34,1164],{},[54,1166,1168],{"id":1167},"原因二embedding-与索引强绑定切换-provider-成本高","原因二：Embedding 与索引强绑定，切换 Provider 成本高",[11,1170,1171,1172,1175,1176,1179,1180,1183,1184,1187],{},"向量 RAG 里常被低估的一点：",[25,1173,1174],{},"向量不是通用中间格式，而是某个 embedding 模型下的坐标。"," 建库用模型 A、查询用模型 B 时，相似度通常 ",[25,1177,1178],{},"不可比","——换模型往往意味着 ",[25,1181,1182],{},"全书重新向量化","，且不同模型的 ",[25,1185,1186],{},"向量维度","（768 / 1024 / 1536 …）会绑死存储 schema。",[11,1189,1190,1191,1194,1195,1198],{},"阶段三持久化的是 ",[25,1192,1193],{},"结构化摘要 + 字符 span","，不存向量；切换 Chat 模型时 ",[25,1196,1197],{},"无需重建索引","，证据链（原文位置）不变。这与「用户随时对比不同 LLM」的目标更一致。",[34,1200],{},[54,1202,1204],{"id":1203},"原因三有目录的长文档结构化路由往往已够用","原因三：有目录的长文档，结构化路由往往已够用",[11,1206,1207,1208,1211,1212,1215,1216,1219,1220,1226],{},"电子书、PDF 通常有 ",[25,1209,1210],{},"章节结构","；预处理已产出 ",[25,1213,1214],{},"段标题 + 摘要","。对「某一章讲了什么」「书中如何定义某概念」类问题，在摘要目录上选段再 ",[25,1217,1218],{},"拉回原文","，实践中效果稳定；且 Tool 回传的是 ",[25,1221,1222,1223,1225],{},"带 ",[154,1224,843],{}," 的原文","，零幻觉仍锚定在字符 span 上。",[11,1228,1229,1230,1233,1234,1237],{},"向量检索在语义模糊、跨语言、长段落字面匹配等场景仍有优势；在 ",[25,1231,1232],{},"有 TOC、可预处理、要强溯源"," 的阅读器里，优先把复杂度放在 ",[25,1235,1236],{},"Tool + 原文回传 + 引用约束"," 上，ROI 通常更高。",[34,1239],{},[54,1241,1243],{"id":1242},"后续方向混合召回而非推倒重来","后续方向：混合召回，而非推倒重来",[11,1245,1246,1247,1250,1251,1254,1255,1258,1259,1262],{},"不排除将来增加 ",[25,1248,1249],{},"向量粗召回","（例如 embedding 只筛 Top-N 候选章节），最终仍走 ",[25,1252,1253],{},"选段 → 原文回传 → 可点击溯源","，零幻觉规则不变。若引入，会尽量满足：Embedding ",[25,1256,1257],{},"可选","、换模型时 ",[25,1260,1261],{},"显式提示重建索引","，避免 silent wrong retrieval。",[11,1264,1265,1266,232],{},"在此之前，优先保证：",[25,1267,1268],{},"任意 OpenAI 兼容 Chat API 即可工作，换 Chat 模型不必重建本地索引",[34,1270],{},[37,1272,1274],{"id":1273},"十小结","十、小结",[281,1276,1277,1290],{},[284,1278,1279],{},[287,1280,1281,1284,1287],{},[290,1282,1283],{},"环节",[290,1285,1286],{},"手段",[290,1288,1289],{},"作用",[303,1291,1292,1303,1316,1329,1340,1351],{},[287,1293,1294,1297,1300],{},[308,1295,1296],{},"预处理",[308,1298,1299],{},"按目录/长度切分 + 片段摘要缓存",[308,1301,1302],{},"长书可检索、可定位",[287,1304,1305,1308,1313],{},[308,1306,1307],{},"位置标记",[308,1309,1310,1312],{},[154,1311,156],{}," 写入原文",[308,1314,1315],{},"出处可机器解析",[287,1317,1318,1321,1326],{},[308,1319,1320],{},"Tool 检索",[308,1322,1323,1324],{},"按问题查片段/全书摘要，回传 ",[25,1325,322],{},[308,1327,1328],{},"作答前强制取证",[287,1330,1331,1334,1337],{},[308,1332,1333],{},"System Prompt",[308,1335,1336],{},"书优先、禁止伪造角标、查不到要说",[308,1338,1339],{},"约束生成行为",[287,1341,1342,1345,1348],{},[308,1343,1344],{},"前端溯源",[308,1346,1347],{},"角标 → 预览 → 跳转高亮",[308,1349,1350],{},"用户可核验证据",[287,1352,1353,1356,1359],{},[308,1354,1355],{},"不用向量检索",[308,1357,1358],{},"单 Provider、换 Chat 模型无需重建索引",[308,1360,1361],{},"降低集成与迁移成本",[11,1363,1364,1365,1368],{},"「零幻觉」不是指望模型从不犯错，而是 ",[25,1366,1367],{},"用工程结构把输出锁在证据链上","：没有检索结果就不应冒充书中内容；有检索结果则应给出可核验的原文位置。",[11,1370,1371,1372,1375,1376,1379],{},"若你也在做 AI 阅读或文档 QA，希望 ",[25,1373,1374],{},"全文直塞 → 关键句提取 → Tool-First 按需检索"," 这条演进路径，以及 ",[25,1377,1378],{},"内联位置标记 + 原文回传"," 的做法，能作为可参考的一种实现。",[19,1381,1382],{},[11,1383,1384,1385,1390,1391,1395],{},"以上是我们在开发 ",[243,1386,1389],{"href":1387,"rel":1388},"https://reader.linghuxiong.com",[247],"令狐兄","（Foxycape）AI 阅读器实践心得，仅供参考。文末可前往 ",[243,1392,1394],{"href":1393},"/zh#download","下载页面"," 体验阅读器。",{"title":619,"searchDepth":1397,"depth":1397,"links":1398},2,[1399,1405,1406,1407,1408,1412,1419,1423,1424,1431],{"id":39,"depth":1397,"text":40,"children":1400},[1401,1403,1404],{"id":56,"depth":1402,"text":57},3,{"id":138,"depth":1402,"text":139},{"id":235,"depth":1402,"text":236},{"id":426,"depth":1397,"text":427},{"id":487,"depth":1397,"text":488},{"id":502,"depth":1397,"text":503},{"id":633,"depth":1397,"text":634,"children":1409},[1410,1411],{"id":657,"depth":1402,"text":658},{"id":683,"depth":1402,"text":684},{"id":759,"depth":1397,"text":760,"children":1413},[1414,1416,1418],{"id":777,"depth":1402,"text":1415},"6.1 get_related_segment_summaries —— 针对具体问题查片段",{"id":847,"depth":1402,"text":1417},"6.2 get_full_book_segment_summaries —— 全书概览类问题",{"id":868,"depth":1402,"text":869},{"id":900,"depth":1397,"text":901,"children":1420},[1421,1422],{"id":910,"depth":1402,"text":911},{"id":921,"depth":1402,"text":922},{"id":960,"depth":1397,"text":961},{"id":1036,"depth":1397,"text":1037,"children":1425},[1426,1427,1428,1429,1430],{"id":1070,"depth":1402,"text":1071},{"id":1107,"depth":1402,"text":1108},{"id":1167,"depth":1402,"text":1168},{"id":1203,"depth":1402,"text":1204},{"id":1242,"depth":1402,"text":1243},{"id":1273,"depth":1397,"text":1274},null,"2026-06-03","分享 AI 阅读器零幻觉问答的工程实现：回答严格基于当前书籍原文，关键论述可一键溯源到具体段落。",false,"md",{},true,"/zh/blog/zero-hallucination-qa",{"title":6,"description":1434},"zh/blog/zero-hallucination-qa",[1443,1444,1445],"阅读器","AI","技术","zero-hallucination-qa","eTMSHCpDay5ePBNbB3ZRWHkahVC0sdaI3zbi8opYj9E",1780489852786]