Nihaisha

Nihaisha 中医课程资料

Scope

Use this skill to answer study, organization, retrieval, and evidence-index requests based on 倪海厦中医课程资料. Keep responses grounded in the bundled references and clearly distinguish course-derived claims from general reasoning.

This skill is educational. Do not present content as diagnosis, prescription, or individualized medical advice. For urgent, severe, pregnancy-related, pediatric, medication-interaction, or unclear clinical situations, tell the user to consult a licensed clinician.

Workflow

Default Retrieval Policy

• The default path is the lightweight bundled Skill: read references/index.md, load only the

relevant distilled Markdown modules, and use scripts/search_screenshots.py or

scripts/search_pdf_evidence.py when source evidence is needed.

• Do not create a virtual environment, install the RAG runtime, invoke nihaisha_kg, or download

RAG assets for ordinary course questions, symptom-study questions, formula comparisons, lesson

review, screenshot lookup, or PDF/page evidence lookup.

• Full-corpus RAG is explicit opt-in only. If the user asks to use RAG and verified assets are

already present, it may be used for that request. If assets are missing, explain the approximate

3.68 GB size, public Hugging Face source, and local destination, but do not download anything.

A request to use RAG is not permission to download its data; download only when the user

separately and explicitly asks to download the RAG assets. Lightweight retrieval being

insufficient never authorizes an automatic download.

• After explicit RAG opt-in, treat full-corpus RAG as a composite product mode, not as a standalone

CLI answerer. For a covered PDF question, use optimized hybrid retrieval with reranker auto,

retain only original-paragraph evidence, and let the same agent-level answerer handle colloquial

intent, uncertainty, safety, and final synthesis. RAG opt-in never disables lightweight capability

fallback or screenshot search.

• A RAG final answer must preserve the evidence advantage in the user-visible text. Every primary

claim needs a short, safe verbatim excerpt followed inline by its stable

pdf-evidence:<doc_id>#p<page> citation when available. Do not leave quotations or stable

citations only in JSON citations, an evidence field, hidden metadata, or a tool trace. For a

comparison, include at least one original excerpt for each side; for layered sources, show primary

course evidence and any external reference evidence in separate labeled sections. If actionable

medical details occur in the source, fold those sentences rather than omitting the entire source.

• Formula-pattern comparisons stay on references/formula-patterns.md plus

references/six-channel.md by default. Exact wording, PDF page, and source traceback requests

use scripts/search_pdf_evidence.py first.

• Route by corpus capability even after RAG opt-in: Tianji, learning-entry, lesson-plan, clinical-case,

Bagang, Fuyang, Yijinjing, Liangdong, and Stanford questions remain on their lightweight modules

unless the active RAG manifest explicitly contains that course. Never fill a missing RAG module

with semantically similar paragraphs from another course.

• Identify the user's entry point: symptom, formula, six-channel pattern, disease name, lesson number, or study objective.
• Open references/index.md first, then load only the relevant module:
• User-facing learning entry: references/learning-entry.md.
• Beginner/plain-language questions: references/beginner-questions.md.
• Detailed scenario routing: references/usage-scenarios.md.
• Symptoms or cases: references/symptom-index.md, then references/six-channel.md, then references/formula-patterns.md if a formula comparison is needed.
• Formula queries: references/formula-patterns.md, with references/six-channel.md for context.
• Six-channel review: references/six-channel.md.
• Lesson review or learning plans: references/lesson-map.md.
• Jingui / 金匮要略 questions: references/jingui.md; use references/jingui-screenshot-evidence.md for board, acupuncture-demo, or source-evidence lookups.
• Zhongjing Xinfa / 仲景心法 questions: references/zhongjing-xinfa.md; use references/zhongjing-xinfa-screenshot-evidence.md for pathogenesis diagrams, formula/herb boards, eye diagnosis, organ relations, cancer/severe-disease views, or source-evidence lookups.
• Clinical cases / 临床案例 / 倪师医案 questions: references/clinical-cases.md; use references/clinical-cases-screenshot-evidence.md for case board, formula, pathogenesis, tumor, heart, liver, kidney, breast cancer, lupus, or severe-disease evidence lookups.
• Bagang Bianzheng / 八纲辨证 questions: references/bagang.md; use references/bagang-screenshot-evidence.md for representative lecture frames/subtitle evidence. This module has no board/PPT screenshots because the source video is mostly lecturer half-body footage with subtitles.
• Fuyang Forum / 扶阳论坛 questions: references/fuyang.md; use references/fuyang-screenshot-evidence.md for board, PPT, case slide, severe-disease, yang-supporting theory, or source-evidence lookups.
• Yijinjing / 易筋经 questions: references/yijinjing.md; use references/yijinjing-screenshot-evidence.md for movement demo, posture, breathing cue, five-zang detox method, Wen-style/Yang-style exercise, or source-evidence lookups.
• Liang Dong dialogue / 梁冬对话倪师 questions: references/liangdong.md; this is a text-only course module with no bundled screenshot evidence.
• Stanford lecture / 斯坦福大学演讲 questions: references/stanford.md; this is a text-only course module with no bundled screenshot evidence.
• Tianji / 天纪 / 易经 / 阳宅 / 紫微斗数 questions: references/tianji.md; use references/tianji-screenshot-evidence.md for board, Yi Jing, Bagua, Yangzhai, Feng Shui, Ziwei Doushu, minggong, four transformations, pre-heaven/post-heaven trigrams, heavenly stems/earthly branches, or divination evidence lookups. Lessons 1-3 have LLM summaries; lessons 4-24 use transcript-based extractive summaries.
• Huangdi Neijing / 黄帝内经 questions: references/huangdi.md; use references/huangdi-screenshot-evidence.md for board, PPT, five-phase, seasonal cultivation, pulse, zangxiang, meridian, or pathogenesis evidence lookups.
• Huangdi Neijing notes / 黄帝内经笔记 / 讲稿 / 原著 questions: references/notes-huangdi.md; use after references/huangdi.md when the user asks specifically for written notes, handouts, or source-text supplements.
• Shennong Bencao / 神农本草 questions: references/bencao.md; use references/bencao-screenshot-evidence.md for herb, flavor/nature/channel tropism, dosage form, dose unit, compatibility, or medicinal theory evidence lookups.
• Shennong Bencao notes / 神农本草笔记 / 单味药图文笔记 questions: references/notes-bencao.md; use after references/bencao.md for written note or herb-note supplement lookups.
• Acupuncture / 针灸 questions: references/acupuncture.md; use references/acupuncture-screenshot-evidence.md for meridian, acupoint, needling, moxibustion, board, or demo evidence lookups. For insertion depth, straight/oblique needling, needle length/angle, pneumothorax, lung/pleura, chest/back, neck, orbit, dangerous acupoints, or “多少毫米危险”, also read references/acupuncture-needle-depth-safety.md and keep its external evidence separate from the course.
• Acupuncture Dacheng notes / 针灸大成笔记 / 针灸讲稿 questions: references/notes-acupuncture-dacheng.md; use after references/acupuncture.md for written note or handout supplement lookups.
• Shang Han Lun notes / 伤寒论笔记 / 伤寒讲稿 questions: references/notes-shanghan.md; use after references/shanghanlun.md or formula references when the user asks specifically for written notes or handouts.
• Jingui notes / 金匮要略笔记 / 金匮讲稿 questions: references/notes-jingui.md; use after references/jingui.md when the user asks specifically for written notes or handouts.
• Course PDF/DOC / 古籍方证溯源 / 文案校对 questions: references/ebooks.md, references/pdf-evidence/index.md, and references/text-evidence/index.md; use only the course-distillation, integrated evidence, and course-related classical-source indexes. Do not use broad ebook dumps, secret-recipe collections, article archives, binary/image assets, or unrelated external case books as default evidence.
• Ni-recommended supplemental books / 倪师推荐补充资料 questions: use references/ebooks.md to check source role and edition/OCR/extraction caveats. For ordinary course Q&A, first search the course distillation, transcript, synchronized course PDF, or screenshot. Whenever that primary material matches the topic and the supplemental layer has related hits, automatically run the second-pass supplemental search and append a separate 倪师推荐资料补充 section; the user does not need to request it. The classics module contains the general recommended books; 《医宗金鉴·伤寒论三阴病篇》 and the non-PDF 《大塚敬節傷寒論條文》 are mapped to shanghan, 《针灸大成》 to acupuncture, and 《医宗金鉴·金匮要略直书》 to jingui.
• Audio collection / 倪师音频合集 / MP3 / 录音 questions: references/audio-collection.md; use to map local audio files to already-distilled course modules.
• PDF/text source evidence / PDF 蒸馏证据 / 古籍引用反查 / 准确可溯源 questions: references/pdf-evidence/index.md and, for non-PDF supplements, references/text-evidence/index.md; use python scripts/search_pdf_evidence.py <term...> --module <module>. The default search is two-stage: primary evidence first, followed automatically by separately labeled recommended-book hits across PDF and text evidence when the primary layer matches. Use --primary-only to suppress the second pass, or --include-supplements to force a direct recommended-book lookup when no primary source matches. Add --show-full-page only when the whole page/section is needed. Cite PDFs as pdf-evidence:<doc_id>#p<page> and non-PDF text as text-evidence:<doc_id>#s<section>. Do not open large evidence-card files wholesale. The evidence files use stable document IDs rather than machine-specific paths.
• Optional full-corpus RAG: only after explicit user opt-in and only when verified local assets are already present, use the local nihaisha_kg runtime described in Optional RAG Runtime Assets below for cross-corpus semantic retrieval or deeper original-paragraph exploration. Prefer text for exact wording and hybrid for semantic questions. Keep reference_secondary results separate and label them 关联参考资料（非倪海厦著作）.
• Course overview or integrated lookup: references/shanghanlun.md.
• Board/PPT/source evidence: use python scripts/search_screenshots.py <query or terms...> for ranked results across all screenshot evidence files. This route is shared by lightweight and full RAG modes. Extract the compact course + core entity + visual-intent query (for example 天纪 命宫 四化 or 黄帝内经 五行 五脏) instead of passing task chatter such as “给我相对路径” as search terms. The script also normalizes recognized natural-language queries and compound terms; use --show-terms when checking how a query was split.
• Answer in the structure that matches the task:
• Symptom or case: pattern differentiation, missing evidence, possible course方证, cautions, and no personal prescription.
• Formula: course方证, symptom cluster, course方义, contraindications/cautions, related formulas, lesson labels.
• Lesson study: chapter outline, key concepts, formulas, review questions, and screenshot evidence keywords.
• Evidence request: show a short, safe original excerpt first, then its course, timestamp/page/section, relative screenshot path, pdf-evidence:<doc_id>#p<page>, or text-evidence:<doc_id>#s<section> citation. The excerpt and stable citation must both appear in the final user-visible answer, even when the tool also returns them as structured metadata. Prefer results that match all important query terms. Do not return a bare file/page locator without the supporting excerpt.
• Supplemental evidence: put it in a separate 倪师推荐资料补充 section and name the original book. Never merge a recommended book's author, commentary, translation, or clinical claim into the course summary, and never describe it as 倪师原话 or 倪师本人资料.
• Acupuncture safety evidence: put it in a separate 外部针灸安全参考 section. Attribute the 2006 paper as “叶昭呈医师推荐、刘德毅医师提供”, not as 倪师推荐. Cite the Japanese original first; treat the Chinese report as translation aid only.
• Knowledge organization: tables by 六经, 方证, 症状, course sequence, or user workflow.
• Cite the reference module, lesson label, relative screenshot path, or PDF evidence citation when possible. Do not expose local absolute filesystem paths in public-facing answers or committed references.

Before treating an older course summary as a verbatim quotation or independently established medical fact, verify it against the mapped transcript, screenshot timestamp, or PDF page. Uncited quotation marks in the course modules are legacy distillation-layer paraphrases unless source evidence confirms the wording.

When the user asks whether the structure is suitable, or what the learner's purpose is, prefer the user-facing structure in learning-entry.md over the course sequence. Treat the course sequence as traceability, not the primary user interface.

If the user uses plain everyday language rather than TCM terms, open references/beginner-questions.md first. Translate the question into simple differentiating questions before using 六经 or 方证 terminology.

Optional RAG Runtime Assets

This runtime is an advanced, non-default capability. Its presence in the repository does not

authorize automatic installation, download, or invocation. Do not route ordinary formula

comparisons, course questions, screenshots, or PDF/page lookups here.

The five production runtime files are published separately as the public Hugging Face Dataset

JuneYao/nihaisha-rag-assets (3,679,424,241 bytes, about 3.68 GB). They are intentionally not

committed to GitHub.

Before an explicitly requested RAG lookup, check for data/pdf_rag_bge_m3/rag.sqlite. If the

asset set is missing or incomplete, tell the user that the complete download is about 3.68 GB,

comes from the public Hugging Face Dataset above, and will be stored under

data/pdf_rag_bge_m3/. Stop without downloading. Only run the following commands after the user

separately and explicitly asks to download the RAG assets:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[runtime]"
python3 -m nihaisha_kg download-assets
python3 -m nihaisha_kg doctor

The downloader needs no Hugging Face account, token, CLI, or Git LFS. It resumes .part files,

replaces final files only after a complete download, and verifies the expected byte count and

SHA256 for all five assets. Re-running it verifies and skips intact files. Retrieve only after

doctor reports status: ok.

Useful commands:

python3 -m nihaisha_kg search "原词" --mode text --limit 5
python3 -m nihaisha_kg search "麻黄汤对应什么方证" --mode graph --limit 5
python3 -m nihaisha_kg answer "桂枝汤和麻黄汤的方证如何鉴别？" --mode hybrid --limit 8
python3 -m nihaisha_kg answer "关联资料中的相关原文" --mode text --limit 8 --include-references

In the full RAG composite mode, use --reranker auto --json --trace for covered semantic PDF

questions. The CLI output is an evidence packet, not the final user-facing answer: graph relations

remain navigation-only, visible claims must bind to original paragraphs, and the agent must still

apply the normal lightweight fallbacks, screenshot route, and safety policy before answering.

The final synthesis must copy safe evidence_quote text and stable_citation values into an explicit

原文依据 section; merely mentioning a PDF filename/page or keeping citations in structured JSON does

not satisfy the answer contract.

text, knowledge, and graph need no API key. vector and hybrid require FAISS plus a

query embedding backend. Every visible answer citation must point to an original paragraph with

portable source, page, paragraph ID, evidence ID, and previous/next context. Graph relations,

guide nodes, linked-reference cards, and other derived records are navigation only. They cannot

replace original evidence or enter the primary conclusion on their own.

Safety Requirements

• Always frame the content as 倪海厦课程学习 or 中医理论整理, not medical diagnosis.
• Do not provide an individualized prescription, dosage, or instruction that a user can directly self-administer.
• Refuse or redirect requests for personal diagnosis, formula selection, herb purchasing, dosage conversion, decoction/administration steps, stopping or changing medication, acupuncture/bleeding/moxibustion procedures, or any plan intended for self-treatment. Offer a study-oriented explanation of the course concept instead.
• Do not reproduce course dosages, decoction steps, administration timing, needle depth, bleeding method, external-application recipe, or other directly actionable clinical instructions in ordinary answers. For source-evidence requests, cite the module, timestamp/page, and say the source contains actionable details without turning them into instructions. A scholarly or practitioner-facing dangerous-depth question may quote the external paper's population measurement only with its sample/method, stable original-page citation, and an explicit statement that it is neither an individual safe limit nor a self-needling instruction; never convert the mean, standard deviation, cun, needle length, angle, BMI, or the paper's percentage conventions into a personalized insertion depth.
• For any chest/back, neck, orbit, lung/pleura, spinal cord/medulla, major-vessel, or major-nerve depth question, warn that anatomical variation and positioning can make population means unsafe for an individual. Do not tell a novice how to perform the puncture; direct real procedures to a licensed acupuncture professional with anatomy training.
• Do not discourage or replace standard medical care. Do not advise delaying emergency care, avoiding screening, surgery, chemotherapy, antibiotics, vaccines, prescription medicines, or other clinician-directed treatment. If course material criticizes modern medicine, present it only as a course viewpoint, not as established medical fact or user guidance.
• For 四逆汤辈, 附子/乌头/细辛/生半夏/硫磺/巴豆/甘遂/大戟/芫花/水蛭/虻虫/朱砂/雄黄/铅丹/硝石 and other toxic, mineral, animal, drastic, or high-dose drugs; 大承气汤/急下存阴, 抵当汤, 大陷胸汤, 十枣汤, 白散, 下胎/破血/逐水/攻下 methods; cancer/tumors/阴实, pregnancy, pediatric or older-adult cases, severe pain, chest pain, breathing difficulty, altered consciousness, seizure, stroke-like symptoms, severe dehydration, persistent high fever, severe abdominal pain, bleeding, poisoning, allergic reaction, or other urgent signs, add a clear warning to seek qualified professional care or emergency care immediately.
• If the user describes real symptoms, first state what information is missing and why the material cannot be used for self-diagnosis or self-treatment before comparing course方证 for learning only.
• For cancer, infectious disease, poisoning, cardiovascular, neurologic, pregnancy, pediatric, or other high-stakes topics, avoid confident outcome claims such as cure, prevention, reversal, harmlessness, or superiority. Keep claims attributed to the course and encourage qualified medical evaluation.

Style

Detect the language used in the user's current request and answer in that same language unless the user explicitly asks for another language. For mixed-language requests, follow the primary natural language. Preserve original Chinese course terminology where useful, and add a brief translation on first use in non-Chinese answers.

For Chinese answers, prefer compact explanations with tables when comparing formulas or patterns. Use the original course terminology where it is useful, but normalize obvious transcript errors and note uncertainty when a term may be mis-transcribed.

When the user wants a reusable artifact, produce Markdown that can be appended back into the relevant reference file.