Wiki

Skill: Wiki Retrieval

Retrieves wiki pages relevant to the current message using BM25 + PageRank search and injects them into context.

Trigger

• Session start: If the host respects Alexandria's MCP instructions field or loads the project AGENTS.md, use this before answering the first wiki-relevant message of the session.
• On-demand: User or LLM calls /wiki <query> mid-session to retrieve context on a new topic.

Flow

1. Search (Tool: mcp__alexandria__search)

Input: { query: string, wikiPath?: string, limit?: number }
Output: { results: [{ path, title, snippet, score }], recentLog: string[] }

1. Parse the user's message as a query string
2. Call the mcp__alexandria__search MCP tool with the query
3. Tool returns ranked results: [{ path, title, snippet, score }]
4. Also returns last 5 log entries for recency context

2. Context Assembly

Script output is assembled into a prompt section (~1000 tokens):

## Relevant Wiki Context

### From: [[wiki/path/to/page]]
**Title:** Page Title
**Summary:** One-line description of what this page covers
**Match:** "...matching snippet with query terms highlighted..."

### Recent Activity
[Last 5 log entries]

Based on the above, the most relevant pages to load are:
1. [[wiki/path/one]] — [reason why]
2. [[wiki/path/two]] — [reason why]

3. Page Selection (LLM Decision)

LLM reviews the search results and decides which 2-4 full pages to load based on relevance to the query.

4. Page Loading

LLM reads the selected wiki pages in full via the host agent's file read capability.

5. Response

LLM responds with full continuity — referencing wiki context naturally in the response rather than announcing "based on my wiki search."

Output Format

No structured output. The LLM integrates wiki knowledge into its response naturally.

Edge Cases

• No results: Return a message indicating no relevant wiki pages found. LLM falls back to general knowledge.
• Query too vague: Return top-5 most recently updated pages as general context.
• Search tool error: Log error, return empty results, allow LLM to proceed without wiki context.

Notes

• The LLM should search the wiki first when the user asks about something that seems like it would be documented, rather than asking the user to re-explain.
• The LLM should NOT announce "I searched the wiki and found..." — it should just use the knowledge naturally.
• Tokens are estimated at ~3000 total for auto-triggered retrieval. On-demand calls are cheaper since the query is focused.