Depth Search

Comprehensive multi-source research skill. Searches across academic databases, semantic web search, and local knowledge before asking the user for help.

Search Order

Execute searches in this order, using parallel subagents where possible:

1. Local Knowledge Base (~/.topos)

Search ~/.topos directory first for existing research, notes, and cached data:

• Use glob and Grep to find relevant files
• Check .md, .org, .jl, .py, .json files
• Look in subdirectories: skills/, archived/, Gay.jl/, etc.

2. Academic MCPs (parallel)

Launch parallel subagents to search all 4 academic sources:

MCP	Tools	Best For
arxiv	search_papers, get_paper, download_paper	Preprints, CS/physics/math papers
semantic-scholar	paper_relevance_search, paper_details, paper_citations	Citation analysis, author profiles
paper-search	search_arxiv, search_pubmed, search_biorxiv, etc.	Multi-source aggregation
deepwiki	read_wiki_structure, read_wiki_contents, ask_question	GitHub repo documentation

3. Exa Semantic Search

Use Exa MCP for high-quality web search:

• web_search_exa - Semantic web search
• crawling_exa - Extract web content
• company_research_exa - Company research
• deep_researcher_start / deep_researcher_check - Deep research tasks

4. Ask User for Help

If all sources fail to find what's needed:

• DO NOT fall back to web_search - it's basic keyword matching only
• Instead, ask the user:
  • "I couldn't find [X] in academic databases, Exa, or local files. Can you provide a link, paper title, or more context?"
  • Suggest specific sources they might check manually
  • Offer to try different search terms

Critical Rules

1. NEVER use web_search as a fallback - it's not equivalent to Exa
2. NEVER use web_search in Task subagents - use Exa tools instead
3. Always search local ~/.topos first - may have cached/annotated versions
4. Use parallel subagents for academic MCPs to maximize speed
5. Ask user for help rather than guessing or using inferior search

Example Workflow

User: "Find papers on world models for LLMs"

1. Search ~/.topos for existing notes/papers
2. Launch 4 parallel Task subagents:
   - arxiv: search_papers("world models LLM")
   - semantic-scholar: paper_relevance_search("world models language models")
   - paper-search: search across all sources
   - deepwiki: check relevant GitHub repos
3. If needed, use Exa: web_search_exa("world models LLM research")
4. Synthesize results from all sources
5. If still not found: ask user for clarification

Parallel Subagent Template

When searching academic sources, use this pattern:

Launch 4 parallel Task subagents:
- Task 1: Use arxiv MCP to search for [query]
- Task 2: Use semantic-scholar MCP to search for [query]  
- Task 3: Use paper-search MCP to search for [query]
- Task 4: Use deepwiki MCP to find related repos/docs

What NOT To Do

❌ web_search as fallback when Exa fails
❌ Single-source search when multiple are available
❌ Skipping local ~/.topos search
❌ Guessing answers without exhausting sources
❌ Sequential searches when parallel is possible

What TO Do

✅ Search ~/.topos first for cached knowledge
✅ Parallel subagents for academic MCPs
✅ Exa for semantic web search
✅ Ask user when sources are exhausted
✅ Synthesize results from multiple sources