Unified Ask (`!ask`)

The !ask command is an agentic search engine that queries all community knowledge sources, synthesizes a single AI-powered answer with citations, and uses multi-round gap detection for thorough results.

Quick Start

!ask <question>              Search all relevant sources, get synthesized answer
!ask <question> --broadcast  Search + broadcast answer to multiple groups
!ask --sources               List available search sources and their status

Aliases: !a

Examples:

!ask What are the claude code plugins and lessons learned?
!ask What link was posted about AI in the DoD?
!ask How do I set up ATAK?
!ask --broadcast What's the latest on drone regulations?

How It Works

!ask uses a 4-stage agentic pipeline that intelligently searches across 7 knowledge sources:

Round 1 (~3-5s):
  ├─ ROUTER — AI picks 2-4 relevant sources for the question
  │  Bot sends progress: "Searching wiki, outline docs, wikipedia..."
  ├─ PARALLEL SEARCH — Selected sources queried simultaneously
  └─ GAP DETECTION — AI reviews results: "Have enough info?"

Round 2 (only if gaps detected, ~3-5s):
  ├─ Search additional sources identified by gap detection
  └─ GAP DETECTION again

Round 3 (final, if still gaps):
  └─ Search remaining sources, then synthesize regardless

Synthesis (~2-3s):
  AI combines ALL gathered context → single answer with citations

Max 3 rounds. After round 3, the bot synthesizes whatever context has been gathered. If no relevant sources are found, it falls back to a general AI answer with a disclaimer.

Search Sources

The router (a fast AI call) selects which sources to query based on the question:

Source	When Selected	Search Method	Speed
IrregularPedia Wiki	Technical/community topics, how-to, guides	Hybrid (keyword + semantic embeddings)	~2-3s
Outline Docs	Team documentation, meeting notes, policies	Outline API full-text search	~1-2s
File Archive	Looking for specific files, documents, PDFs	Smart search + grep	~2-4s
RSS Links	News, recent articles, shared links	PostgreSQL term-based search	~0.5s
Q&A History	Previously asked/answered questions	PostgreSQL term-based search	~0.5s
Wikipedia	General knowledge, concepts, people, organizations	MediaWiki Action API	~1-2s
Google Knowledge Graph	Entity disambiguation, quick facts	Knowledge Graph API	~1s

Router Heuristics

The AI router uses these rules to pick sources:

Technical questions about community tools → wiki first, then outline
General knowledge or “what is X” → wikipedia, knowledge_graph
“Has anyone asked about…” → qa_history
Looking for a specific file → files
Recent news or articles → rss
AI/security/tech topics → always includes wiki
Topics that may not be community-specific → includes wikipedia

Response Format

🤖 Ask: "what are the claude code plugins and lessons learned?"

Claude Code supports several plugin types for extending its capabilities:

1. **MCP Servers** - Model Context Protocol servers that provide
   additional tools. Configure in `.claude/settings.json`...

2. **Custom Slash Commands** - Define in `.claude/commands/`...

**Key lessons learned from production projects:**
- Always establish CLAUDE.md files early
- Use project rules to enforce coding standards
- ...

📚 Sources (searched: wiki, outline, wikipedia):
• Claude Code - IrregularPedia: https://irregularpedia.org/ai-ml/claude-code/
• Project Rules & Lessons Learned: https://irregularpedia.org/ai-ml/project-rules-lessons-learned/
• Claude Code Setup - Outline: https://outline.irregularchat.com/doc/...

Format details:

Header echoes the original question
AI-synthesized answer (max ~1500 chars for Signal readability)
Source list at bottom with clickable URLs
Sources searched noted parenthetically

Comparison with Other Commands

!ask is an orchestrator — it calls into the same search backends as the individual commands, but combines and synthesizes results.

Command	What It Does	When to Use
`!ask`	Unified search across all sources + AI synthesis	General questions, when you don’t know which source to check
`!wikiask` / `!wa`	Search IrregularPedia wiki only	Specific wiki knowledge
`!ol ask`	Search Outline docs only	Specific team documentation
`!files -s`	Smart file search only	Looking for a specific file
`!ai`	Pure AI (no source search)	Opinions, creative tasks, general chat
`!qa`	Submit question to Q&A database for community answers	Want human answers tracked over time

When to use !ask vs specific commands:

Use !ask when you’re not sure which source has the answer
Use specific commands when you know exactly where to look (faster, more targeted)
!ask adds ~5-15 seconds of overhead for routing and gap detection

Relationship to Q&A System

The original !ask command (submit a question for community answers) has been renamed to !qa. See Q&A System for documentation.

Old Command	New Command	Purpose
`!ask <question>`	`!qa <question>`	Submit question for community tracking
`!q <question>`	`!q <question>`	Still works (alias for `!qa`)
`!question <hash>`	`!question <hash>`	View question details (unchanged)
(new)	`!ask <question>`	Unified agentic search

Error Handling

Scenario	Behavior
No sources return results	Falls back to pure AI answer with disclaimer
One source times out	Continues with remaining sources, notes timeout
All sources timeout	Returns error suggesting specific commands
OpenAI rate limit	Queues synthesis, retries once after 5s
Wikipedia API down	Skips, uses remaining sources
Google KG API key missing	Skips Knowledge Graph silently

Per-source timeout: 5 seconds Total timeout: 30 seconds max — synthesizes with whatever is gathered Rate limit: 10 !ask calls per hour per user

Configuration

Required

Variable	Purpose
`OPENAI_API_KEY`	Powers the router, gap detection, and synthesis (gpt-4o-mini)

Optional

Variable	Purpose	Default
`GOOGLE_KG_API_KEY`	Google Knowledge Graph API	Disabled if not set
`QA_WEB_URL`	Base URL for Q&A web links in results	None

Wikipedia search requires no API key or authentication (uses the public MediaWiki Action API).

Architecture

Pipeline Stages

Router — Single gpt-4o-mini call (temperature: 0.1) picks 2-4 sources
Parallel Search — All selected sources searched simultaneously via Promise.all()
Gap Detection — gpt-4o-mini evaluates gathered context, decides if more sources are needed
Synthesizer — gpt-4o-mini combines all context into a cohesive answer (temperature: 0.3 with context, 0.7 without)

Key Files

File	Purpose
`src/src/utils/unified-search.ts`	Router, parallel search, gap detection, synthesizer
`src/src/utils/wikipedia-client.ts`	Wikipedia MediaWiki Action API client
`src/src/utils/knowledge-graph-client.ts`	Google Knowledge Graph API client
`src/src/utils/rss-search.ts`	RSS links term-based search
`src/src/bot/command-handler.ts`	`handleUnifiedAsk()` method, command routing

Search Term Extraction

For RSS and Q&A history, natural language questions are split into individual search terms with stop words removed. For example:

Question: "was there a link posted last week about ai in the dod?"
Stop words removed: ["link", "posted", "last", "week", "about"]
Search terms: ["ai", "dod"]
SQL: WHERE (title ILIKE '%ai%' OR summary ILIKE '%ai%')
      OR (title ILIKE '%dod%' OR summary ILIKE '%dod%')

This dramatically improves recall compared to matching the entire question as a single pattern.

Q&A System — Community Q&A tracking (!qa)
Wiki & Knowledge — Wiki search commands (!wikiask, !wikisearch)
Outline Docs — Team documentation (!ol ask)
RSS Reader — Link aggregation and search
Integrations — External service configuration (Wikipedia, Knowledge Graph)
Architecture Overview — System architecture