| name | deep-research |
| description | Multi-source deep research through Centra AI web search, fetch, and the router-owned Exa MCP. Searches, extracts, synthesizes, and delivers cited reports with explicit grounding. |
| origin | Centra AI/Paxlabs |
Deep Research
Produce a cited research answer from multiple web sources using ordinary discovery plus Centra AI's router-owned Exa evidence tools.
Tool Boundary
Use the shipped web_search, web_news, fetch, and exa__* tools. Exa vendor credentials and run ownership stay in the Centra router. Do not install a direct vendor MCP or ask the user for an API key.
Workflow
- Turn the user's objective into 3-5 bounded research questions. Use reasonable defaults unless a missing choice materially changes the result.
- Use
web_search / web_news for cheap discovery. Add exa_search where semantic matching, domain/date controls, financial reports, or synthesis improves the evidence set.
- Read selected ordinary-search URLs with
fetch. Use exa_contents when extracting a bounded claim from known URLs; preserve every per-URL status and disclose partial evidence.
- For genuinely multi-step synthesis, call
exa_research_start with a bounded output schema and medium-or-lower effort by default, then poll with exa_research_get. Only terminal grounding authorizes the generated output.
- Cross-reference claims, separate facts from inference, label uncertainty, and deliver the answer with citations near each supported claim.
Quality Rules
- Every material current claim needs a source.
- Prefer official, primary, academic, and reputable reporting sources.
- Flag claims supported by only one weak source.
- Say when evidence was not found; never fill a gap with a similarly named entity or a different evidence class.
- Exa highlights and Contents extracts are source evidence. Exa answers and Agent output are generated synthesis.
- Never cite Agent previews or intermediate sources as final support; terminal grounding is authoritative.
- Report partial Contents failures instead of silently dropping them.
- Do not start expensive asynchronous research merely because a surface opened; it must be explicitly useful to the user request.
Report Shape
- Lead with the answer and key findings.
- Organize evidence into the smallest useful set of themes.
- Put caveats beside the affected finding.
- End with actionable implications when the user is making a decision.
