Back to case study
Research pipeline replay

Enterprise Deep Research Agent (Dify)

A Deep Research system as a Dify workflow: intent gate → topic decomposition → a ReAct agent iteratively searches/extracts evidence via Tavily → DeepSeek/Qwen writes a footnote-cited Markdown report.

A replay of the Dify Deep Research workflow on a sample topic: the three-way intent gate, decomposition into subtopics, a ReAct agent that searches and extracts evidence per subtopic, source dedup with stable IDs, and a footnote-cited Markdown report.

DifyDeepSeekQwenTavilyReActWorkflow
Enterprise Deep Research Agent (Dify)

Why this local version exists

The preview replays the real graph's behavior; no live Dify/Tavily/LLM calls. The running workflow lives on the Dify platform and needs DeepSeek, Tongyi/Qwen, and Tavily API keys.

Interactive Preview

Run the Deep Research workflow

Replays the Dify graph on a sample topic: intent gate → decomposition → an iterative ReAct agent that searches and extracts evidence per subtopic → source dedup → a footnote-cited report.

Research topic

当前主流的大模型对齐(alignment)方法有哪些?

Intent gate (DeepSeek)

Classify input: NeedMoreInfo / Decompose / Execute. Here → Execute.

Decompose (DeepSeek)

main_intent + key_dimensions + 2–3 subtopics with retriever-ready queries.

Research agent (ReAct)

Per subtopic: tavily_search → tavily_extract → {claim, quote, confidence}.

Accumulate (code)

URL→sid dedup, merge findings, append history.

Report (Qwen3-max)

Markdown report with [^sid] footnote citations.

Activity log

Run the workflow to watch the Dify graph execute stage by stage.

Subtopics & evidence

0/3
Subtopics and their extracted evidence appear here as the ReAct agent runs.

What to try

Run the workflow and follow the Dify graph: intent gate → decompose → research → accumulate → report.

Watch the ReAct agent extract {claim, quote, confidence} per subtopic via Tavily search + extract.

See URL→sid dedup (one source reused across subtopics) and the footnote-cited report.

What this demo proves

You can design a non-trivial Dify workflow: branching intent control, iteration, code nodes, multi-model routing.

You take evidence discipline seriously: claim+quote+confidence, stable source IDs, footnote citations.

You know when low-code is the right tool — and can be honest that the value is the orchestration.

Platform

Dify advanced-chat workflow · DeepSeek + Qwen3-max · Tavily

Graph

Intent gate → decompose → ReAct research → dedup → cited report

Best signal

Evidence-first agent workflow design in low-code