wiki-query

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Query — Ask about the wiki

查询——询问维基相关内容

wiki/CONVENTIONS.md

for format conventions, links, and language.

请遵循

wiki/CONVENTIONS.md

中的格式规范、链接和语言要求。

Language

语言

Write the artifact in the user's language. Apply correct grammar and any required diacritics or script-specific characters. If the user's language is unclear, ask before generating output.

使用用户的语言撰写输出内容。确保语法正确，并使用所需的变音符号或特定脚本字符。如果用户的语言不明确，请先询问再生成输出。

Query language alignment

查询语言对齐

When the user's language differs from the wiki language, translate the retrieval query into the wiki language before searching. Determine the wiki language from

.wiki-guardrails.yml

(

query_language

language

), then from wiki frontmatter/index if guardrails are absent. Keep exact product names and code identifiers unchanged.

Write
```
intent:
```
,
```
vec:
```
, and
```
hyde:
```
in the wiki language.
Use
```
lex:
```
for exact terms from both languages when useful.
Answer the user in the user's language unless they ask otherwise.
If translation is uncertain, say which terms were translated and keep the original terms in
```
lex:
```
.

当用户使用的语言与维基语言不同时，先将检索查询翻译成维基语言再进行搜索。从

.wiki-guardrails.yml

文件（

query_language

或

language

字段）确定维基语言；如果没有该配置文件，则从维基的前置元数据/索引中确定。产品名称和代码标识符保持不变。

用维基语言编写
```
intent:
```
、
```
vec:
```
和
```
hyde:
```
字段。
当有用时，使用
```
lex:
```
来包含两种语言中的精确术语。
除非用户另有要求，否则用用户的语言回答问题。
如果对翻译不确定，请说明哪些术语已翻译，并将原始术语保留在
```
lex:
```
中。

Retrieval — prefer QMD when available

检索——优先使用QMD（如果可用）

The recommended retrieval engine is QMD (local hybrid search: BM25 + vector + LLM reranking). Setup is one-time per repo — see

docs/wiki/qmd-setup.md

in this skills repo.

Detect availability before each query:

MCP available (
```
mcp__qmd__status
```
and
```
mcp__qmd__query
```
tools registered) → use the MCP tools.

CLI available (

which qmd

succeeds) → use

qmd query --json --files --min-score 0.4

Neither available → fall back to the grep/index path described below. Tell the user once that QMD is not configured and point them to
```
docs/wiki/qmd-setup.md
```
.

Never run

qmd embed

qmd update

, or

qmd collection add

automatically — those are owner-run setup commands. The skill only queries an already-built index.

推荐的检索引擎是**QMD**（本地混合搜索：BM25 + 向量 + LLM重排序）。每个仓库只需一次性设置——请参阅本技能仓库中的

docs/wiki/qmd-setup.md

。

每次查询前检测可用性：

MCP可用（已注册
```
mcp__qmd__status
```
和
```
mcp__qmd__query
```
工具）→ 使用MCP工具。

CLI可用（

which qmd

执行成功）→ 使用

qmd query --json --files --min-score 0.4

命令。

两者都不可用→ 回退到下文描述的grep/索引路径。告知用户一次QMD未配置，并引导他们查看
```
docs/wiki/qmd-setup.md
```
。

切勿自动运行

qmd embed

、

qmd update

或

qmd collection add

命令——这些是管理员执行的设置命令。本技能仅查询已构建好的索引。

Steps

步骤

Plan the search — turn the user's question into a structured query document. Always provide an
```
intent:
```
line so reranking and snippet extraction stay anchored to the actual goal:
```
intent: <one-sentence framing of what the user actually wants>
lex: <exact terms / phrases / -negations>
vec: <natural-language paraphrase of the question>
hyde: <one short paragraph that looks like the ideal answer>
```
Drop a sub-query if it does not add signal (a glossary lookup needs only
```
lex
```
; a "how does X work" is mostly
```
vec
```
/
```
hyde
```
).
Retrieve via QMD
- MCP: call
```
mcp__qmd__query
```
  with the structured
```
searches
```
  array (or the multi-line
```
q
```
  string) and the
```
intent
```
  field. Restrict by
```
collections
```
  when the question is scoped to a single collection vs cross-repo.
- CLI:
```
qmd query "$(cat <<EOF intent: ... lex: ... vec: ... hyde: ... EOF )" --json --files -n 8 --min-score 0.3
```
- Read the top 3-5 hits with
```
mcp__qmd__get
```
  /
```
mcp__qmd__multi_get
```
  (or
```
qmd get
```
  /
```
qmd multi-get
```
  on the CLI). Fetch the full body when the snippet is borderline; honor the
```
context
```
  field returned by QMD — it carries collection-level guidance the owner curated.
Fallback path (no QMD)
- Read
```
wiki/index.md
```
  and use the "By topic" table to locate likely pages.
- Use
```
grep
```
  /
```
Glob
```
  with the keywords distilled in step 1.
- Read the identified pages in full.
Synthesize the answer:
- Direct answer to the question.
- Cite the wiki pages used with links:
```
[page](wiki/path.md)
```
  (use the path returned by QMD, not the docid).
- If sources contradict each other, present the different viewpoints and flag the conflict.
- If the information does not exist in the wiki, say so explicitly and suggest sources that could be ingested.
Assess whether the answer has lasting value
- If the answer synthesizes multiple pages in a new and useful way → offer to save it as a page in
```
wiki/sources/
```
  (cross-cutting summary) or as a new section in an existing page.
- If it is a simple one-off answer → no need to save.

Log (only if the query resulted in a wiki update):

## [YYYY-MM-DD] query | <question summary>
- Retrieval: qmd | grep-fallback
- Pages consulted: ...
- Result saved: yes/no (which page, if yes)

规划搜索——将用户的问题转换为结构化查询文档。始终提供
```
intent:
```
行，以便重排序和片段提取始终紧扣实际目标：
```
intent: <用户实际需求的一句话描述>
lex: <精确术语/短语/否定词>
vec: <问题的自然语言改写>
hyde: <一段类似理想答案的简短段落>
```
如果子查询没有增加有效信息，可以省略（比如词汇表查询只需要
```
lex
```
；“X如何工作”类问题主要用
```
vec
```
/
```
hyde
```
）。
通过QMD检索
- MCP：调用
```
mcp__qmd__query
```
  ，传入结构化的
```
searches
```
  数组（或多行
```
q
```
  字符串）以及
```
intent
```
  字段。如果问题限定在单个集合而非跨仓库，则通过
```
collections
```
  参数进行限制。
- CLI：
```
qmd query "$(cat <<EOF intent: ... lex: ... vec: ... hyde: ... EOF )" --json --files -n 8 --min-score 0.3
```
- 使用
```
mcp__qmd__get
```
  /
```
mcp__qmd__multi_get
```
  （或CLI中的
```
qmd get
```
  /
```
qmd multi-get
```
  ）读取前3-5条结果。如果片段内容不够明确，则获取完整正文；遵循QMD返回的
```
context
```
  字段——它包含管理员整理的集合级指导信息。
回退路径（无QMD）
- 读取
```
wiki/index.md
```
  ，使用“按主题分类”表格定位可能的页面。
- 使用
```
grep
```
  /
```
Glob
```
  工具结合步骤1中提炼的关键词进行搜索。
- 完整读取找到的页面。
合成答案：
- 直接回答用户的问题。
- 使用链接引用所使用的维基页面：
```
[页面名称](wiki/path.md)
```
  （使用QMD返回的路径，而非docid）。
- 如果来源之间存在矛盾，呈现不同观点并标记冲突。
- 如果维基中不存在相关信息，请明确说明，并建议可纳入的来源。
评估答案是否具有持久价值
- 如果答案以新的有用方式整合了多个页面的内容→ 提议将其保存为
```
wiki/sources/
```
  中的页面（跨领域摘要）或现有页面的新章节。
- 如果只是简单的一次性答案→ 无需保存。

记录（仅当查询导致维基更新时）：

## [YYYY-MM-DD] query | <问题摘要>
- 检索方式: qmd | grep回退
- 查阅页面: ...
- 结果已保存: 是/否（如果是，说明保存到哪个页面）

Rules

规则

Answer first based on the wiki — do not re-synthesize from scratch using code or external sources when the answer already exists.
Always supply an
```
intent:
```
line to QMD — it is the strongest single lever for relevance.
Trust QMD's
```
context
```
field — it reflects curated metadata (e.g.
```
raw/
```
is not the source of truth). Do not "correct" it from training intuition.
If you cannot find the information, suggest which sources could be ingested to cover the gap.
Always cite the wiki pages used.
If the question reveals a gap, suggest pages that could be created (but do not create them automatically — leave that for ingest).

首先基于维基内容回答——当答案已存在时，不要使用代码或外部资源重新生成。
始终向QMD提供
```
intent:
```
行——这是提升相关性最有效的手段。
信任QMD的
```
context
```
字段——它反映了经过整理的元数据（例如
```
raw/
```
不是权威来源）。不要凭经验“修正”它。
如果找不到信息，建议可纳入哪些来源来填补空白。
始终引用所使用的维基页面。
如果问题暴露出内容缺口，建议可创建哪些页面（但不要自动创建——留待纳入流程处理）。