Wiki Query

wiki-query

Margot checks the wiki first. External research only fires when the wiki can't answer — or when the answer exists but the data might be stale.

When to invoke

Before every [RESEARCH] sentinel resolution in margot_bot.handle_turn. The wiki-query result either:

• Replaces the external call (confidence=high, not time-sensitive)
• Supplements it (confidence=medium — wiki context injected alongside external findings)
• Defers entirely (confidence=low — skip wiki, go straight to external)

Algorithm

1. Load Wiki/index.md — the page map.
2. Identify relevant pages. LLM call: given the query, name ≤5 pages from the index that are likely to contain an answer.
3. Read those pages. Load them from disk (max 5, max 1 500 chars each to bound prompt size).
4. Synthesise an answer. LLM call: answer the query from the page content. Return the answer plus a confidence assessment.
5. Return. Structured result with answer, pages consulted, confidence, go_external flag, and stale flag.

Confidence → go_external mapping

Time-sensitive topics (market data, competitor moves, pricing, regulatory) always set go_external=True regardless of confidence — the wiki is a starting point, not a substitute for Gemini grounded search on fast-moving topics.

Input schema

{
  "query": "What is CCW's contract value?",
  "time_sensitive": false
}

Output schema

{
  "answer": "CCW signed a 12-month SaaS contract at $2,400/year on 2026-05-08.",
  "pages_consulted": ["ccw.md", "businesses-overview.md"],
  "confidence": "high" | "medium" | "low",
  "go_external": false,
  "stale": false,
  "error": null
}

Python entry point

from swarm.wiki_query import query

result = query("What is CCW's contract value?", time_sensitive=False)
if not result.go_external:
    # use result.answer directly
else:
    # inject result.answer as context, then fire external research

Integration point in margot_bot

In handle_turn, before _run_research_batch:

from .wiki_query import query as wiki_query

wiki_result = wiki_query(research_request.topic,
                         time_sensitive=(research_request.depth == "quick"))
if not wiki_result.go_external and wiki_result.answer:
    # substitute wiki answer — no external call needed
    research_findings.append({
        "topic": research_request.topic,
        "depth": "wiki",
        "summary": wiki_result.answer,
        "error": None,
    })
else:
    # fire external + optionally prepend wiki context
    ...

Safety

• Read-only — never writes to the wiki. Use wiki-ingest for writes.
• Read from config.BRAIN1_WIKI_DIR only.
• Failure is always non-fatal — on error, go_external=True so Margot falls through to external research.

Verification

from swarm.wiki_query import query

r = query("What is CCW's first-response SLA?")
assert r.confidence in ("high", "medium")
assert "15 min" in r.answer or "15" in r.answer
assert "ccw.md" in r.pages_consulted

r2 = query("What did Anthropic announce this week?", time_sensitive=True)
assert r2.go_external is True

References

• Wiki schema: ~/2nd Brain/2nd Brain/CLAUDE.md
• Config: swarm/config.py → BRAIN1_WIKI_DIR
• Companion skills: wiki-ingest (Wave 5), wiki-lint (Wave 5)