Inspired by Karpathy's LLM Wiki, this project builds a PDAC-first, long-term evolving medical wiki framework that works with agents (OpenClaw / Hermes), and is designed for knowledge reuse across model eras.
Traditional RAG is query-time retrieval. We focus on a persistent, evolving wiki:
- LLM + human co-maintained knowledge
- traceable evidence and versioned updates
- reusable by different future agent/runtime stacks
- Disease focus: PDAC (pancreatic ductal adenocarcinoma)
- Product focus:
- Patient-facing: Web Chat + Feishu/WeChat bot
- Builder-facing: knowledge production & review workflow
- Storage strategy:
- Feishu as source-of-truth
- GetNote as mirror
- local standard core (Markdown + JSON cards + graph)
- M1 (Week 1-4)
- finalize schema + folder conventions
- Feishu -> Markdown/Cards sync POC
- first 30-50 PDAC knowledge cards
- M2 (Week 5-8)
- launch Web + Feishu/WeChat entry
- integrate 4 MVP skills:
patient_intake,plan_generate,evidence_trace,risk_check - force citation-backed answers
- M3 (Week 9-12)
- lint for contradiction/staleness/orphans
- enable evolution-log and review loop
- community workflow online (PR -> lint -> review -> merge)
We are actively recruiting collaborators.
- Backend / Platform: sync pipeline, adapters, APIs, reliability
- Agent / Skill Engineer: OpenClaw/Hermes integration, skill contracts, routing
- Data / Knowledge Engineer: schema, card normalization, provenance, linting
- Frontend Engineer: patient chat UX, evidence display, timeline views
- Clinical content reviewers (doctor/senior patient volunteers)
- Python/TypeScript, Markdown automation, API integration
- Feishu API / bot integration experience is a plus
- medical evidence traceability mindset
- Open an Issue with title prefix:
[JOIN] <role> - Introduce your background + available time + sample work
- Start from a good-first-task in roadmap/issues
- Canonical design doc:
docs/architecture-v1.1.md - Docs index:
docs/README.md - Reference draft (kept for context):
reference_design.md
- v1.1 design consolidated
- repository initialized and synced
本项目受 Karpathy 的 LLM Wiki 启发,目标是构建一个以胰腺癌(PDAC)为起点、可长期演化的医疗 Wiki 框架,可与 OpenClaw / Hermes 等智能体协同工作,并在不同模型时代持续复用知识资产。
传统 RAG 主要在“提问时检索”。本项目更关注一个持续演化的 Wiki:
- 由 LLM 与人类协同维护知识
- 每条结论尽量具备证据追溯与版本历史
- 可以被未来不同 Agent / Runtime / 应用层复用
- 病种聚焦:PDAC(胰腺导管腺癌)
- 产品聚焦:
- 面向患者:Web Chat + 飞书/微信机器人
- 面向建设者:知识生产与审核工作流
- 存储策略:
- Feishu 作为主库(source of truth)
- GetNote 作为镜像
- 本地标准核心库(Markdown + JSON cards + graph)
- M1(第 1-4 周)
- 定稿 schema 与目录规范
- 打通 Feishu -> Markdown/Cards 同步 POC
- 形成首批 30-50 张 PDAC 知识卡片
- M2(第 5-8 周)
- 上线 Web 与飞书/微信入口
- 接入 4 个 MVP skills:
patient_intake、plan_generate、evidence_trace、risk_check - 强制答案附带证据引用
- M3(第 9-12 周)
- 建立矛盾/过期/孤儿页 lint 机制
- 启用 evolution-log 与审核回路
- 上线社区协作流程(PR -> lint -> review -> merge)
我们正在持续招募协作者。
- 后端 / 平台工程师:同步链路、适配器、API、可靠性
- Agent / Skill 工程师:OpenClaw/Hermes 集成、skill 契约、路由
- 数据 / 知识工程师:schema、卡片规范化、来源追溯、lint
- 前端工程师:患者聊天体验、证据展示、时间线视图
- 医学内容审核者(医生 / 资深患者志愿者)
- Python / TypeScript、Markdown 自动化、API 集成
- 有 Feishu API / Bot 集成经验更好
- 对医疗证据追溯与知识治理有意识
- 提交 Issue,标题前缀:
[JOIN] <role> - 简述你的背景、可投入时间、相关作品
- 从 roadmap / issues 中认领一个 good-first-task 开始
- 主设计文档:
docs/architecture-v1.1.md - 文档索引:
docs/README.md - 参考草案:
reference_design.md
- 已完成 v1.1 设计整合
- 仓库已初始化并同步