baekenough/wiki-rag

Wiki RAG Skill

Query the project wiki to answer questions about the codebase. The wiki is a pre-compiled knowledge base — searching it is faster and more accurate than exploring raw source files.

Usage

/omcustom:wiki-rag "how does orchestration work?"
/omcustom:wiki-rag "what agents handle database tasks?"
/omcustom:wiki-rag "explain the compilation metaphor"

Also triggered automatically by intent-detection when a user asks about project architecture, agent roles, skill purposes, or rule behavior.

Workflow

Step 1: Load Index

Read and parse wiki/index.yaml to get the full page catalog. If wiki/index.yaml does not exist, report:

[wiki-rag] Wiki not initialized. Run /omcustom:wiki first to create wiki/index.yaml.

Then abort.

Step 2: Identify Relevant Pages

From the question and the index, identify 3–7 most relevant wiki pages. Selection criteria:

| Criterion | Detail | |-----------|--------| | Title/summary keyword match | Direct term overlap with the question | | Category relevance | Architecture questions → architecture/; agent questions → agents/ | | Cross-reference chains | Follow [[related]] links to surface adjacent context |

Step 3: Read Pages (Parallel)

Read identified pages in parallel per R009. Extract sections relevant to the question; ignore unrelated sections to control context size (R013).

Step 4: Synthesize Answer

Compose a direct answer citing wiki pages inline:

[wiki/architecture/orchestration.md] 참조 — 메인 대화가 유일한 오케스트레이터...

Answer format:

• Direct answer first (2–3 sentences)
• Supporting details from wiki pages
• Inline citations to specific wiki pages ([wiki/path/to/page.md])
• "See also" links for further reading

Step 5: Evaluate Coverage

After synthesizing, assess whether the wiki alone was sufficient.

| Coverage Level | Condition | Action | |---------------|-----------|--------| | Full | Wiki pages answer the question completely | Return wiki-sourced answer | | Partial | Wiki covers some aspects; gaps remain | Supplement with raw codebase exploration (Read/Glob/Grep); offer to ingest | | Miss | Wiki has no relevant pages | Full raw exploration; offer to create wiki page |

If raw file exploration was needed, offer to persist the new knowledge:

[wiki-rag] This answer required raw file exploration. Save findings to wiki? [Y/n]

If yes: delegate wiki page creation/update to the wiki-curator agent (R010 — orchestrator delegates all wiki writes).

Step 6: Log Query

Insert into wiki/log.duckdb:

duckdb wiki/log.duckdb "INSERT INTO wiki_log (operation, target, pages_consulted, coverage, details) VALUES ('query', '{question}', N, 'full/partial/miss', '{\"fallback\": false, \"new_page_created\": false}');"

Log writes are delegated to a subagent (R010).

Integration with Intent Detection

The following trigger patterns route user questions to wiki-rag automatically:

| Trigger Keywords | Language | Confidence | Action | |-----------------|----------|------------|--------| | "어떻게 작동", "how does X work" | KO/EN | 85% | wiki-rag | | "무슨 에이전트", "which agent" | KO/EN | 80% | wiki-rag | | "규칙", "rule R0XX" | KO/EN | 75% | wiki-rag | | "아키텍처", "architecture" | KO/EN | 90% | wiki-rag | | "워크플로우", "workflow" | KO/EN | 85% | wiki-rag |

Patterns should be registered in .claude/skills/intent-detection/patterns/agent-triggers.yaml when that file is updated.

Fallback Strategy

| Wiki Coverage | Action | |--------------|--------| | Full (answer from wiki alone) | Return wiki-sourced answer | | Partial (wiki + raw files needed) | Supplement with raw exploration, offer to ingest findings | | Miss (wiki has nothing relevant) | Full raw exploration, offer to create new wiki page |

For partial and miss cases: always acknowledge what the wiki covered before pivoting to raw exploration. Do not silently fall back.

Execution Rules

| Rule | Requirement | |------|-------------| | R000 | Answers to user in Korean; wiki content and citations reference English file paths | | R009 | Parallel reads for identified wiki pages (Step 3) | | R010 | All wiki write operations (log.duckdb, new pages) delegated to subagents | | R013 | Ecomode: return concise answers when context pressure is high; suppress "See also" section |

What Makes a Good Wiki-RAG Answer

• Cites specific wiki pages — not vague "the wiki says..." but [wiki/agents/mgr-creator.md]
• Provides actionable information — tells the user what to do, not just what exists
• Links to related pages for readers who want to go deeper
• Acknowledges gaps honestly — never hallucinate content not found in the wiki
• Offers to improve wiki coverage when gaps are found — compounding the knowledge base over time

Integration with /omcustom:wiki

| Skill | Role | |-------|------| | wiki | Builds and maintains the wiki (ingest, update, lint) | | wiki-rag | Queries the wiki as a RAG source; surfaces gaps; triggers ingest offers |

wiki-rag is the read-path; wiki is the write-path. They share wiki/index.yaml and wiki/log.duckdb as the coordination layer.

Integration

| Rule / Skill | Integration | |--------------|-------------| | R009 | Parallel page reads in Step 3 | | R010 | Log writes and wiki page creation delegated to subagents | | R000 | Korean user responses; English wiki content and paths | | R013 | Ecomode: suppress extended "See also" section; summary answer only | | R015 | Intent detection routes architecture/agent/rule questions here at 75–90% confidence | | wiki | Counterpart write-path skill; wiki-rag triggers ingest offers to wiki when gaps found | | intent-detection | Registers trigger patterns for auto-routing user questions to wiki-rag |