| name | deep-research |
| description | Deep web research via LangGraph agent on VPS. Conduct multi-step research with web search, analysis, and report generation. Two modes: "auto" (fire-and-forget, Pi answers clarifications) and "interactive" (relays questions to user). Trigger: "research X", "investigate Y", "look into Z", "find information about", "write a report on", "what does the research say about".
|
Deep Research
Pi has access to a deep research tool that runs a LangGraph research agent on the VPS. The agent uses DeepSeek for reasoning and Tavily for web search, producing comprehensive research reports.
When to Use
Use deep_research when the user asks for:
- Research on a topic: "research the state of X"
- Investigation: "look into Y"
- Reports: "write a report on Z"
- Fact-finding: "what does the research say about..."
- Comparative analysis: "compare A vs B"
Don't use for:
- Simple factual questions (Pi can answer directly)
- Code-related tasks (use your normal tools)
- Real-time data (the agent searches the web, but isn't real-time)
Modes
interactive (default)
The research agent may ask clarifying questions. The tool returns the question. Relay it to the user, get their answer, then call deep_research_continue with the thread_id and answer. Repeat until the agent delivers a final report.
Pattern:
1. Pi: deep_research(query="research AI safety", mode="interactive")
2. Agent: "What aspect of AI safety — technical alignment, policy, or existential risk?"
3. Pi relays question to user
4. User: "Technical alignment"
5. Pi: deep_research_continue(thread_id="...", answer="Technical alignment")
6. Agent: [delivers final report]
auto (fire-and-forget)
The agent handles clarifications automatically — Pi answers them using a small LLM call. The user is not bothered. Best when the user says "just do it" or "background research."
Pattern:
User: "Do some background research on quantum computing advances"
Pi: deep_research(query="quantum computing advances 2026", mode="auto")
Pi: [waits, then returns the complete report]
Tips
- The tool can take minutes to complete (up to 1 hour max) — give the user a heads-up
- Clarification questions are usually 1-2 sentences — relay them verbatim
- Thread IDs are opaque strings — pass them exactly as received
- If the server is down:
systemctl --user start deep-research.service
