[Feature] Auto-generate consistent retrieval tags for notes (type/domain/year/status, plugin-free)

[jiang4wqy]

English first, 中文见下 ⬇︎

Summary

DeepPaperNote produces excellent single-paper notes, but it is deliberately single-paper by design (SKILL.md: "it handles one paper at a time, it does not update daily reading lists"). Once a user accumulates dozens or hundreds of paper notes in their vault, a new problem appears: the notes don't connect and are hard to retrieve across the collection.

This proposal adds a small, deterministic step that auto-generates a consistent set of native Obsidian tags on each note, so the vault becomes filterable — no plugin required.

Proposed feature

When writing a note, automatically emit a set of clearly-namespaced, consistently-spelled native tags into its frontmatter. First version covers 4 facets:

Tag	Source	Controlled vocabulary
type/ paper type	existing note_plan.paper_type	AI-method / benchmark-dataset / clinical-empirical / humanities-social / survey
domain/ domain	reuses the archived domain folder (dominant_domain)	consistent with the vault's existing domain folders
year/ year	publication year from metadata	e.g. year/2026
status/ personal status	default to-read, user edits later	to-read / reading / read / to-reproduce / reproduced / archived

Example (a note's frontmatter):

tags:
  - type/AI-method
  - domain/medical-ai
  - year/2026
  - status/to-read

The user can then click type/survey, year/2026, or status/to-reproduce in Obsidian's native tag pane to aggregate matching papers — with no plugin such as Dataview.

Why this design (fits the project's philosophy)

• Deterministic normalization → scripts. Per architecture.md, deterministic/repeated normalization belongs in the script layer. The tags are synthesized deterministically from artifacts the pipeline already computed (domain/ reuses the resolve_domain_subdir result, so the tag and the archive folder can never diverge). Consistency is locked in, not left to the model's discretion.
• Contracts centralized, not buried in prompts. The controlled vocabulary lives in contracts.py, following "don't bury essential contracts only in prompt text."
• Grounding-safe. All four values come from the paper / metadata / existing archive — nothing requires "world knowledge," so it does not conflict with the fail-closed grounding discipline.
• Zero external dependency. Uses only Obsidian's native tags; users are not required to install any plugin.

Benefits

1. The note collection turns from "a pile of documents" into "a click-to-filter mini-database," greatly improving reuse/retrieval.
2. Consistency is guaranteed by the tool — the same concept always maps to the same tag, avoiding the drift of hand-typed tags.
3. Zero disruption to existing single-note quality and workflow; purely additive.

Scope & non-goals (first version)

• ✅ Only the 4 tag facets above, with deterministic generation + consistency guarantee.
• ❌ No plugin dependency (Dataview etc.).
• ❌ No method/ or dataset/ tags yet (their vocabularies are the messiest / most drift-prone; deferred).
• ❌ No cross-paper comparison tables or vault-level dashboards — those are separate future proposals.

This feature is the foundation of a larger "knowledge-base retrieval layer" (optional follow-ups: a vault-level index/dashboard, and cross-paper auto-linking — both would consume the structured tags this feature produces).

Implementation sketch (for discussion)

• Synthesize the 4 tags in the save path (write_obsidian_note.py) and write them into the frontmatter; keep the vocabulary + mapping in contracts.py.
• Add a lightweight check in lint_note.py: the 4 tags exist, values are within the controlled vocabulary, and domain/ matches the archive folder.
• Backward compatible: keep the existing tags/aliases conventions; new tags are additive and don't break old notes. Works on both POSIX and Windows.

---

---

背景 / 痛点（中文）

DeepPaperNote 目前把「写好单篇笔记」做到了很高的水准，但它明确是单篇工具（SKILL.md："it handles one paper at a time, it does not update daily reading lists"）。当用户在 vault 里攒下几十上百篇论文笔记后，会遇到一个新问题：笔记之间不连通、不好检索。

想找「我读过的、所有用了某类方法 / 属于某个领域 / 某年 / 我标了想复现的论文」，现在只能一篇篇翻或全文搜（不准）。笔记写得再好，攒多了也难以二次利用。

本提案新增一个小而确定性的步骤：自动为每篇笔记生成一组一致的原生 Obsidian 标签，让笔记库可筛选——零插件。

提议的功能

写笔记时，自动在其 frontmatter 生成一组分类清晰、写法统一的原生标签。首版含 4 类：

标签	取值来源	受控词表
type/ 论文类型	note_plan 已有的 paper_type	AI方法 / 基准数据 / 临床实证 / 人文社科 / 综述
domain/ 领域	复用归档用的领域文件夹（dominant_domain）	与 vault 已有领域夹一致，不新造
year/ 年份	元数据中的发表年份	如 year/2026
status/ 个人状态	默认 待精读（用户后续手改）	待精读 / 精读中 / 已读 / 想复现 / 已复现 / 存档

示例：

tags:
  - type/AI方法
  - domain/医学AI
  - year/2026
  - status/待精读

用户即可在 Obsidian 原生标签面板点击 type/综述、year/2026、status/想复现 一键聚合，无需任何插件（如 Dataview）。

为什么这样设计（契合项目哲学）

• 确定性归一化 → 放脚本：按 architecture.md，这类确定性、可重复的归一化逻辑应放脚本层。标签由脚本从管线已经算好的结构化产物确定性生成（domain/ 直接复用 resolve_domain_subdir 的结果，因此标签与归档目录永远一致、不会分裂），一致性是「锁死」的，而非靠模型自觉。
• 契约集中、不藏在提示词里：受控词表放进 contracts.py，符合 "don't bury essential contracts only in prompt text"。
• 接地安全：4 类取值全部来自论文 / 元数据 / 既有归档，不引入需要"世界知识"的内容，不与 fail-closed 的接地纪律冲突。
• 零外部依赖：只用 Obsidian 原生标签，不要求用户装插件。

好处

1. 笔记库从「一堆文档」变成「可点击筛选的小数据库」，显著提升二次检索/复用价值。
2. 一致性由工具保证——同一概念永远同一个标签，避免手打标签漂移。
3. 对现有单篇笔记质量与流程零破坏，纯增量。

范围与非目标（首版）

• ✅ 仅做上述 4 类标签的确定性生成 + 一致性保证。
• ❌ 不引入 Dataview 等插件依赖。
• ❌ 暂不做 method/ dataset/ 标签（词表最杂、最易漂，留待后续）。
• ❌ 不做跨论文对比表、vault 级看板——那是后续独立提案。

本功能是更大的「知识库检索层」的地基（后续可选：② vault 级索引看板、③ 跨论文自动关联，二者都消费本功能产出的结构化标签）。先立地基。

实现草图（供讨论）

• 在保存路径（write_obsidian_note.py）合成 4 类标签并写入 frontmatter；词表与映射放 contracts.py。
• lint_note.py 增加一道轻量检查：4 类标签存在、值在受控词表内、domain/ 与归档目录一致。
• 向后兼容：保留现有 tags/aliases 约定，新标签为叠加，不破坏旧笔记；POSIX/Windows 均适用。