| name | llmwiki:search |
| description | Search the LLM Wiki using qmd (hybrid BM25/vector search). Finds relevant wiki pages, reads them, and synthesizes an answer with citations. Use when asking questions against the knowledge base. |
| argument-hint | <search query> |
| user-invokable | true |
| allowed-tools | Read, Write, Edit, Bash, Glob, Grep, Agent, mcp__plugin_qmd_qmd__query, mcp__plugin_qmd_qmd__status, mcp__plugin_qmd_qmd__get, mcp__plugin_qmd_qmd__multi_get, TaskCreate, TaskUpdate, TaskList, TaskGet |
LLM Wiki: Search
Search the wiki and synthesize an answer from the knowledge base.
Step 1: Parse Query
The user's argument is the search query. If no argument is provided, ask the user what they'd like to search for.
Step 2: Search Strategy
Use a multi-pronged search approach:
2a. qmd Search (Primary)
Use the qmd MCP tools to search the wiki collection:
{
"searches": [
{ "type": "lex", "query": "<key terms extracted from query>" },
{ "type": "vec", "query": "<full natural language query>" }
],
"collections": ["wiki"],
"intent": "<what the user is looking for>",
"limit": 15,
"minScore": 0.3
}
For complex queries, add a hyde sub-query with a hypothetical answer (50-100 words).
2b. Index Scan (Fallback/Supplement)
Read wiki/index.md and scan for relevant page titles and descriptions. This catches pages that qmd might miss due to indexing lag.
2c. Grep (Precision)
If the query contains specific terms, names, or identifiers, use Grep to find exact matches across wiki pages.
Step 3: Read Relevant Pages
From the search results, identify the most relevant pages (typically 3-10). Read them in full using the Read tool or qmd get.
For each page, note:
- How it relates to the query
- Key facts or claims
- Which sources it draws from
- What wikilinks it contains that might lead to more relevant pages
If the initial pages reference other wiki pages that seem highly relevant, follow those links too (one level deep).
Step 4: Synthesize Answer
Compose an answer that:
- Directly answers the query based on wiki content
- Cites sources using wikilinks:
[[Page Title]]
- Notes confidence level — how well-covered is this topic in the wiki?
- Identifies gaps — what's missing from the wiki that would help answer this better?
- Suggests follow-ups — related questions the user might want to explore
Format the answer in clean markdown. Use headers, bullet points, and tables where they aid clarity.
Step 5: Offer to File
If the answer is substantial and adds value beyond what's already in the wiki, offer to file it:
"This analysis could be saved as a wiki page. Would you like me to file it in wiki/comparisons/ or wiki/synthesis/?"
If the user agrees, write the page following standard wiki conventions (frontmatter, wikilinks, proper filename) and update the index.
Append to wiki/log.md:
## [YYYY-MM-DD] query | <short query description>
- Query: <full query>
- Pages consulted: [[Page1]], [[Page2]], ...
- Filed as: wiki/path/to/new-page.md (if applicable)
Tips
- Combine search types. Lexical search finds exact terms; vector search finds related concepts. Use both.
- Follow the links. Wiki pages are richly interlinked — the best answer often comes from reading 2-3 connected pages.
- Check source freshness. Note the
updated date in frontmatter. If a page hasn't been updated recently, mention this.
- Be honest about gaps. If the wiki doesn't cover something well, say so. Don't hallucinate from outside the wiki unless the user asks.
