Linkup Deep Research
The Research endpoint is an autonomous agent that investigates the web iteratively for minutes, cross-checks claims across sources, and returns a synthesized, cited answer. Reach for it only when a single deep search can't cover the job — many entities, many facets, verification needed, or a report-style deliverable.
This uses the REST API, so it needs LINKUPAPIKEY in the environment:
test -n "$LINKUP_API_KEY" || echo "Missing LINKUP_API_KEY"
Choose mode and depth explicitly
Always set mode and reasoningDepth for predictable latency, cost, and output shape.
- mode —
answer (a precise question with one correct answer, self-verified), investigate (deep analysis of one defined entity), research (broad, multi-angle exploration across many entities). Rule of thumb: ends with "?" and expects one answer → answer; "the state of / analysis of" → research; one entity comprehensively → investigate.
- reasoningDepth —
S (~2-5 min), M (~3-7 min, default for a bounded report), L (~5-10 min, market maps / multi-company), XL (~10-20 min, only when the user wants exhaustive coverage).
- outputType —
sourcedAnswer for reports; structured only with a structuredOutputSchema.
Submit, then poll with backoff
POST /v1/research returns {id, status: "pending"} immediately. Poll GET /v1/research/{id} with exponential backoff (start 2s, cap 10s) until status is completed or failed. Never poll faster than once per second. Failed tasks are not charged.
curl -sS -X POST "https://api.linkup.so/v1/research" \
-H "Authorization: Bearer $LINKUP_API_KEY" -H "Content-Type: application/json" \
-d '{"q":"Compare 2024 cloud revenue growth for Microsoft, Amazon, and Google from primary sources.","mode":"answer","reasoningDepth":"M","outputType":"sourcedAnswer"}'
curl -sS "https://api.linkup.so/v1/research/RESEARCH_ID" -H "Authorization: Bearer $LINKUP_API_KEY"
Write a well-scoped prompt
Specify the angles to cover, leads to pursue, facts to verify, entities to compare, constraints/exclusions, and the output structure. More precise input yields more thorough, aligned output.
Guidance
- Research costs more and runs for minutes ($0.25–$2.50 by depth). Tell the user what you're about to run and why before starting.
- Ground every claim in the returned
sources. When done, summarize in plain English: what was researched, the answer, key sources, and roughly how long it took.
- Retry only transient failures (
429, 5xx) with backoff.
For the full decision matrix, mode examples, patterns, and polling details, read references/LINKUPSPECIALIZEDENDPOINTS.md (Research section) in this skill's directory.
