Deep Research node

The Deep Research node runs a multi-round research loop: several search rounds run in parallel, each covering a different angle of the question, then a final synthesis pass writes a single cited markdown report.

How it works

1. The query is split into complementary angles (core facts, what sources miss, verification, broader context).
2. Each angle runs as an independent search round with the deep search lever.
3. A synthesis call merges the findings into a report with numbered inline citations and a source list.

Configuration

• Provider — the search backend (Tako, Exa, Perplexity, Parallel).
• Model — the language model (reasoning-capable models work best).
• Query — the research question (may reference {input.*} templates).
• Max iterations — number of parallel rounds (1–8).
• Max results — results per round.

Example

Query: What are the current best practices for RAG evaluation? with maxIterations: 3. The node returns:

{
  "report": "# RAG evaluation\n\n…cited markdown…",
  "sources": [{ "url": "https://…", "title": "…", "quote": "…" }]
}

Notes

• Rounds run in parallel, so maxIterations: 3 costs roughly the time of one round, not three.
• Each LLM call is billed separately and attributed to the run.
• Use Web Search for a quick single-pass answer; use Deep Research for a thorough, sourced report.

Related

• Web Search — the single-pass variant.
• Output — rendering the report.
• Use cases — one research pass feeding a summary, a table and a deck.