wiki-context-pack

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Wiki Context Pack — Bounded Token Retrieval

Wiki上下文包——受限Token检索

You are producing a focused, token-bounded context pack from the wiki. Unlike

wiki-query

(which answers a question), this skill packages the most relevant wiki knowledge into a single markdown block that a downstream agent, skill, or user can consume directly.

你需要从wiki生成一个聚焦的、受token限制的上下文包。与

wiki-query

（用于回答问题）不同，该skill会将最相关的wiki知识打包成单个Markdown块，供下游Agent、skill或用户直接使用。

Before You Start

开始之前

Resolve config — follow the Config Resolution Protocol in
```
llm-wiki/SKILL.md
```
(walk up CWD for
```
.env
```
→
```
~/.obsidian-wiki/config
```
→ prompt setup). This gives
```
OBSIDIAN_VAULT_PATH
```
and any QMD variables.
Read
```
$OBSIDIAN_VAULT_PATH/hot.md
```
if it exists — gives instant context on recent activity.
Read
```
$OBSIDIAN_VAULT_PATH/index.md
```
— the full page inventory.

解析配置 — 遵循
```
llm-wiki/SKILL.md
```
中的配置解析协议（从当前工作目录向上查找
```
.env
```
→
```
~/.obsidian-wiki/config
```
→ 提示设置）。这会获取
```
OBSIDIAN_VAULT_PATH
```
及所有QMD变量。
若存在
```
$OBSIDIAN_VAULT_PATH/hot.md
```
，请读取该文件——它提供了近期活动的即时上下文。
读取
```
$OBSIDIAN_VAULT_PATH/index.md
```
——完整的页面清单。

Invocation Forms

调用格式

/wiki-context-pack "transformer attention mechanism" --budget 16000
/wiki-context-pack "my-project architecture decisions" --budget 8000
/wiki-context-pack --recent --budget 4000   # recent activity pack from hot.md
/wiki-context-pack "authentication patterns"          # default budget: 8000 tokens

Parse the user's invocation to extract:

topic — the query string (required unless
```
--recent
```
)
--budget N
— token budget in tokens (default:
```
8000
```
; max:
```
100000
```
)
--recent
— pack the most recently updated/ingested pages instead of a topic query

/wiki-context-pack "transformer attention mechanism" --budget 16000
/wiki-context-pack "my-project architecture decisions" --budget 8000
/wiki-context-pack --recent --budget 4000   # 从hot.md生成近期活动包
/wiki-context-pack "authentication patterns"          # 默认预算：8000 tokens

解析用户的调用内容，提取以下信息：

主题 — 查询字符串（除非使用
```
--recent
```
，否则为必填项）
--budget N
— token预算（默认：
```
8000
```
；最大值：
```
100000
```
）
--recent
— 打包最近更新/导入的页面，而非基于主题查询

Algorithm

算法流程

Step 1: Relevance Pass (cheap)

步骤1：相关性筛选（低成本）

Without opening page bodies:

Scan
```
index.md
```
and frontmatter for topic match. Score each page:
- +5 exact title or alias match
- +3 tag match
- +2
```
summary:
```
  field contains the query term
- +1
```
index.md
```
  entry description contains the query term
For
```
--recent
```
mode: sort pages by
```
updated:
```
frontmatter descending. Take top 20 as candidates.
For topic mode: collect the top 20 candidates by score. If QMD is configured (
```
QMD_WIKI_COLLECTION
```
set), run a semantic pass and merge with the frontmatter score (QMD rank adds +4 to the page's score).

无需打开页面内容：

扫描
```
index.md
```
和frontmatter，匹配主题。为每个页面打分：
- +5分：标题或别名完全匹配
- +3分：标签匹配
- +2分：
```
summary:
```
  字段包含查询关键词
- +1分：
```
index.md
```
  中的条目描述包含查询关键词
若为
```
--recent
```
模式：按frontmatter中的
```
updated:
```
字段降序排序页面，取前20个作为候选。
若为主题模式：按分数收集前20个候选页面。若已配置QMD（
```
QMD_WIKI_COLLECTION
```
已设置），执行语义检索并将结果与frontmatter分数合并（QMD排名为页面额外加**+4分**）。

Step 2: Tier-Aware Selection

步骤2：分层筛选

Within the candidate set, sort by relevance score, then apply tier ordering within each score bucket (see

llm-wiki/SKILL.md

, Importance Tiering section):

All
```
core
```
-tier matches first
Then
```
supporting
```
Then
```
peripheral
```
(only if budget allows)

Maintain this ordering when filling the budget in Step 3.

在候选集中，先按相关性分数排序，再在每个分数区间内按层级排序（参考

llm-wiki/SKILL.md

中的重要性分层章节）：

优先选择所有
```
core
```
层级的匹配页面
其次是
```
supporting
```
层级
最后是
```
peripheral
```
层级（仅当预算允许时）

在步骤3填充预算时，需保持此排序顺序。

Step 3: Compression

步骤3：压缩处理

For each selected page (in tier/relevance order), compute its compressed representation — not a full read, but a structured distillation:

Required: title,
```
tier:
```
,
```
tags:
```
,
```
summary:
```
(from frontmatter — cheap, no body read needed)
If budget allows: add the page body, but stripped of:
- Frontmatter block (already captured above)
- The
```
## Sources
```
  section (keep source names in a one-liner instead)
- Duplicate wikilinks that are already mentioned in included pages
- Boilerplate headers with no content following them
Dedup overlapping content — if two selected pages share a paragraph (or near-identical claim), keep it only in the more relevant page. Mark the removal:
```
_(content also in [[other-page]])_
```
.

Estimate tokens for each page representation as

len(text_chars) / 4

对每个选中的页面（按层级/相关性顺序）生成压缩表示——并非完整读取，而是结构化提炼：

必填内容：标题、
```
tier:
```
、
```
tags:
```
、
```
summary:
```
（来自frontmatter——低成本，无需读取页面内容）
若预算允许：添加页面内容，但需移除以下部分：
- Frontmatter块（已在上方捕获）
- ```
## Sources
```
  章节（改为用单行保留来源名称）
- 已在其他包含页面中提及的重复wikilinks
- 无后续内容的模板化标题
去重重叠内容：若两个选中页面包含相同段落（或近乎相同的表述），仅保留相关性更高页面中的内容，并标记移除：
```
_(内容也见于[[other-page]])_
```
。

按

len(text_chars) / 4

估算每个页面表示的token数量。

Step 4: Budget Enforcement

步骤4：预算管控

Fill the pack greedily in tier/relevance order until the budget is exhausted:

Always include the frontmatter summary block for every selected page, even if the body doesn't fit.
If a page body doesn't fit in full, include a compressed excerpt: the first non-header paragraph plus the "Key Ideas" section (if present).
Drop
```
peripheral
```
-tier pages first when trimming.
Keep a running token count. Stop adding pages when the next page would exceed the budget.
Track how many pages were dropped and note it in the header.

按层级/相关性顺序贪婪填充包内容，直至预算耗尽：

始终包含每个选中页面的frontmatter摘要块，即使页面内容无法完全容纳。
若页面内容无法完整容纳，添加压缩摘录：第一个非标题段落加上“Key Ideas”章节（若存在）。
裁剪时优先移除
```
peripheral
```
层级的页面。
持续跟踪token计数，当添加下一个页面会超出预算时停止。
统计被移除的页面数量，并在标题中注明。

Step 5: Render Output

步骤5：输出渲染

Emit a single markdown block:

markdown

undefined

生成单个Markdown块：

markdown

undefined

Context Pack: <topic>

上下文包: <主题>

Generated: <ISO timestamp>

生成时间: <ISO时间戳>

Budget: <budget> tokens | Actual: <actual> tokens | Pages: <N included> / <M candidates>

预算: <预算> tokens | 实际使用: <实际> tokens | 页面: <已包含数量> / <候选数量>

Methodology: 4 chars/token estimate

估算方法: 4字符/Token

[[<category/page-name>]] (<tier>, ~<tokens> tokens)

[[<分类/页面名称>]] (<层级>, ~<token数> tokens)

tags: #tag1 #tag2 summary: <summary field text>

tags: #tag1 #tag2 summary: <摘要字段文本>

<压缩内容或摘录>

[[<next-page>]] (<tier>, ~<tokens> tokens)

[[<下一页>]] (<层级>, ~<token数> tokens)

...


If `--recent` mode, the header reads:

...


若为`--recent`模式，标题为：

Context Pack: Recent Activity (last N pages)

上下文包: 近期活动（最近N页）


**Empty result:** If no pages scored above 0 and `--recent` produced no results, output:


**空结果处理**：若没有页面得分高于0，且`--recent`模式无结果，输出：

Context Pack: <topic>

上下文包: <主题>

No relevant pages found. Consider running /wiki-ingest to add sources about this topic.

undefined

未找到相关页面。建议运行/wiki-ingest添加该主题的相关来源。

undefined

Step 6: Log

步骤6：日志记录

Append to

$OBSIDIAN_VAULT_PATH/log.md

- [TIMESTAMP] CONTEXT_PACK topic="<topic>" budget=<N> actual_tokens=<M> pages_included=<K> pages_dropped=<D>

在

$OBSIDIAN_VAULT_PATH/log.md

中追加内容：

- [时间戳] CONTEXT_PACK topic="<主题>" budget=<N> actual_tokens=<M> pages_included=<K> pages_dropped=<D>

Use Cases

使用场景

Feed into
/wiki-research
— pass the pack as context to avoid re-discovering known facts
Pass to
/wiki-synthesize
— scoped input for a specific synthesis task
Provide to external agents via MCP or clipboard — bounded, structured, citation-ready
Checkpoint context before a long multi-step task — know what the wiki already knows before starting

传入/wiki-research — 将包作为上下文，避免重复发现已知信息
传入/wiki-synthesize — 为特定合成任务提供限定范围的输入
通过MCP或剪贴板提供给外部Agent — 受限、结构化、可直接引用的内容
长多步骤任务前的上下文检查点 — 在开始任务前了解wiki已有的相关知识

Notes

注意事项

The
```
4 chars/token
```
heuristic matches
```
wiki-status
```
's token footprint estimate — consistent across skills
The pack is a snapshot; it is not written to the vault. Re-run to refresh.
For very large budgets (> 50K tokens), warn the user: "This pack is large. Consider narrowing your topic or using wiki-query for a targeted answer instead."

"4字符/Token"的估算方法与
```
wiki-status
```
的token占用量估算一致——确保各skill间的一致性
包为快照形式，不会写入vault。需重新运行以刷新内容
对于超大预算（>50K tokens），需向用户发出警告："此包体积较大。建议缩小主题范围，或使用wiki-query获取针对性答案。"