WisdomAI Should Chunk Meaning, Not Characters
WisdomAI 应该按义理切分，而不是按字符切分
How I am turning Liao-Fan’s Four Lessons into a traceable knowledge system without flattening a classical text into arbitrary fragments.
我如何把《了凡四训》转化为可追溯的知识系统，同时避免把经典机械地切成失去义理的文本碎片。
A fixed-length text splitter can preserve tokens while destroying thought.
That matters for any knowledge system, but it becomes especially important in WisdomAI. The Buddhist MVP is not intended to retrieve a sentence and improvise advice around it. It should identify a person’s problem, locate the relevant principle, show the textual basis, and only then offer a bounded practice.
Start from the book’s own structure
Liao-Fan’s Four Lessons already has four distinct chapters: changing destiny, reforming faults, cultivating virtue, and the effects of humility. Their themes and forms of argument differ. The first boundary should therefore come from the text—not from a character counter.
固定长度切分可以保留 Token，却可能破坏思想。
这对任何知识系统都重要，但对 WisdomAI 尤其关键。佛教 MVP 的目标不是检索到一句话，再围绕它即兴生成建议；它应该先识别人的问题，找到相关义理，展示文本依据，然后才给出有边界的实践建议。
先尊重经典自己的结构
《了凡四训》本身就分为立命之学、改过之法、积善之方和谦德之效四篇。它们的主题与论证方式不同，所以第一层边界应该来自文本本身，而不是字符计数器。
Book
identity and provenance
Chapter
theme and summary
Semantic Section
one complete principle
Passage Chunk
precise retrieval unit
整本书
身份与来源
篇章
主题与摘要
义理单元
一个完整原理
检索 Chunk
精确命中的小单元
Semantic sections are the center of the design
Within “Reforming Faults,” sections might include why change is necessary, the three minds required for reform, changing through action, changing through principle, changing from the mind, and signs that change is taking effect. Each section should answer a real question, such as: “Why does reform require courage?”
Only after that semantic boundary is established should length become a constraint. Classical Chinese may need 200–500 characters; a translation may use 500–900; a commentary can be longer. A complete story or classification should remain intact even when it exceeds a default limit.
not: ENTIRE BOOK → EVERY 800 CHARACTERS
Small chunks retrieve; parent sections answer
WisdomAI can use Parent–Child chunking: small child chunks improve recall and precision, while their parent section restores the complete reasoning context before generation.
义理单元才是设计中心
以“改过之法”为例，可以进一步拆成：为什么必须改过、改过须发三心、从事上改、从理上改、从心上改，以及改过后的征验。每个 Section 都应该回答一个真实问题，例如：“为什么改过需要勇心？”
只有确定义理边界之后，长度才成为约束。文言原文可以从 200–500 字开始，白话译文约 500–900 字，讲解可以更长；完整故事与完整分类即使略超限制，也不应该被机械截断。
而不是：整本书 → 每 800 字切一刀
小块负责检索，Parent 负责回答
WisdomAI 可以采用 Parent–Child Chunking：较小的 Child Chunk 提升检索准确度，命中后再回到完整 Parent Section，让生成模型获得足够的推理上下文。
Authority must remain explicit
The original text, a modern translation, a teacher’s commentary, and an AI-generated application do not carry the same authority. Combining them into one large string would make retrieval convenient while making provenance invisible.
权威层级必须保持显式
原文、现代译文、老师或法师的讲解，以及 AI 生成的应用建议，并不具有相同权威。把它们拼成一个大字符串，虽然方便 Embedding，却会让出处和责任边界消失。
Each layer can have its own embedding, but retrieval should know which one it found. When the user asks what the text says, original passages rank first. When the user asks what it means, translations and reviewed commentaries can be raised. When WisdomAI suggests a practice, it must still cite the original evidence behind the reasoning.
Metadata is part of reasoning—not housekeeping
每一层都可以生成独立 Embedding，但检索系统必须知道自己命中了什么。当用户问“原文怎么说”，优先返回原文；当用户问“这句话是什么意思”，提升译文与经审核讲解；当 WisdomAI 提出实践建议时，仍然必须回溯原文依据。
Metadata 是推理的一部分，不只是资料管理
principles: ["勇于改过", "知过即改"]
practice: ["明确停止的行为", "今天执行第一步", "持续观察"]
authority: original · human_reviewed
canonical_status: non_buddhist_canon
risk_flags: ["不得把不幸简单归因于德行不足", "不得承诺行善必得世俗回报"]
A story is not a formula
A story about virtue and good fortune illustrates a principle. It should not become a mechanical claim that one action guarantees one worldly reward.
A moral text is not a sutra
Liao-Fan’s Four Lessons carries Confucian, Buddhist, and Daoist influences, but it is not Buddhist scripture. WisdomAI must name that boundary.
故事不是因果公式
行善得福的故事用于说明义理，不应该被转化成“做一件善事，就必然得到某种世俗回报”的机械承诺。
劝善书不等于佛经
《了凡四训》兼有儒释道影响，是传统修身与劝善文本，并不是佛教经典。WisdomAI 必须明确标注这一边界。
The principle is simple: chunk by meaning, separate by authority, retrieve small, answer with context, and preserve the reasoning path.
For WisdomAI, better retrieval is not only a relevance problem. It is a responsibility problem.
最终原则很简单：按义理切分，按权威分层；用小块检索，用大块回答；用图关系保留理论链条。
对 WisdomAI 来说，更好的检索不仅是相关性问题，也是责任问题。