How to Build a Deep Research Workflow in Dify Using AI Agents

Business Pain Points

Standard search queries often fail for complex problems such as academic papers, market analysis, or code debugging, requiring dozens of separate searches. This is where deep research becomes valuable.

Deep research uses an intelligent feedback loop to identify knowledge gaps, lock onto specific questions, explore systematically, and produce comprehensive reports, unlike fragmented searches.

Workflow Overview

The Dify deep research workflow consists of three stages: Intent Identification, Iterative Exploration, and Synthesis.

Intent Identification : Capture the research topic, collect background, and define clear direction.

Iterative Exploration : Use loop variables to evaluate knowledge, run targeted searches, and build findings.

Synthesis : Compile all information into a structured report with proper citations.

Stage 1: Research Foundations

Start Node

Configure the Start node with basic input parameters:

Research Topic : Core question to explore.

Maximum Loop : Iteration budget for the session.

Background Knowledge Acquisition

Use the Exa Answer tool to gather initial information and ensure the model understands terminology.

Intent Analysis

Employ an LLM node to extract the user's true intent, distinguishing surface issues from deeper information needs.

Stage 2: Dynamic Research Cycle

Loop Node: Research Engine

The loop node drives the entire research, passing information across iterations.

Dify tracks six key variables:

findings : New knowledge discovered each cycle.

executed_querys : Previously used search queries to avoid redundancy.

current_loop : Iteration counter.

visited_urls : Properly cited source URLs.

image_urls : Visual content references.

knowledge_gaps : Identified information needs.

Reasoning Node: Better Questions

The reasoning node outputs structured JSON with fields reasoning, search_query, and knowledge_gaps.

{
    "reasoning": "Detailed justification for the chosen action path...",
    "search_query": "Specific follow-up question targeting knowledge gaps",
    "knowledge_gaps": "Information still needed to answer the original question"
}

Agent Node: Conduct Research

Agent nodes act as autonomous researchers, selecting appropriate tools such as exa_search for web search and exa_content for retrieving full content.

They also use the think tool (inspired by Claude's Think Tool) to reflect on findings and plan next steps.

URL Extraction

The workflow automatically extracts URLs and visual references from agent responses for proper source tracking.

Variable Assignment

After each cycle, a variable assigner node updates the research state, ensuring each iteration builds on previous work.

Stage 3: Research Synthesis

When all exploration cycles finish, the Summary node aggregates accumulated variables to generate a comprehensive report with correct Markdown citations and a bibliography.

Conclusion

This guide demonstrates how Dify can digitize expert research methods and accelerate them through automation.

Future research should focus on smarter data exploration rather than merely more data.

References

https://github.com/dzhng/deep-research

https://github.com/jina-ai/node-DeepResearch

https://github.com/langchain-ai/local-deep-researcher

https://github.com/nickscamara/open-deep-research