Skip to content

Iterative Research: Human-Guided Exploration

The power of ResearchCrew is in iteration. After each round, you review findings and actively steer the next research direction.

The Iterative Loop

Round 1: Initial Research
Review findings
Identify gaps and areas to explore
Round 2: Guided Research (with feedback)
Review refined findings
Request deeper analysis or new angles
Round 3+: Continue iterating

Step-by-Step Example

Round 1: Initial Exploration

Create input.md:

# Research Topic: The Future of Remote Work

Research the future of remote work post-2025. Include current trends, 
company policies, and predictions from industry experts.

Run:

crewai run

You get <yyyymmdd>.md with initial findings on remote work trends.

Round 2: Provide Feedback & New Direction

After reviewing <yyyymmdd>.md, add feedback at the end:

## User Feedback

Good coverage on company policies and trends. Now explore:

1. **Mental health impact** — Research psychological effects of remote work 
   (isolation, burnout, work-life balance)

2. **Hybrid model adoption** — Focus on companies successfully implementing 
   hybrid (not full remote) models

3. **Tools & infrastructure** — Deep dive into remote work tools and technologies

Please de-prioritize:
- General history of remote work (already covered)
- COVID-era temporary remote mandates (focus on permanent structures)

Run again:

crewai run

The crew:

  • Remembers your previous research
  • Sees your feedback
  • Explores the new topics you highlighted
  • Skips areas you said to de-prioritize
  • Integrates findings from both rounds

You get an updated <yyyymmdd>.md with deeper analysis on your chosen angles.

Round 3: Refine Further

Review the updated report. You notice:

  • Mental health section is strong
  • Hybrid model section needs more company case studies
  • Tools section could include budget considerations

Add new feedback:

## User Feedback

Excellent analysis on mental health. The hybrid model section is good but needs:
- Real case studies (which companies, what results)
- Employee satisfaction metrics
- Implementation challenges and solutions

For tools section, add:
- Cost comparisons between platforms
- SMB vs Enterprise tool choices
- Security and compliance considerations

New area to explore:
- How remote work is affecting real estate and office spaces

Run again:

crewai run

Continue until the research meets your needs.

How the Crew Remembers

ResearchCrew uses LanceDB to persist memory across runs:

What it remembers:

  • Previous research outputs
  • Your input topic
  • Extracted claims and sources
  • URLs already crawled

What it learns:

  • Your feedback patterns (what you want deeper on)
  • Topics you want to skip
  • The full context of prior investigation

What it avoids:

  • Re-crawling already-researched URLs
  • Repeating topics you've marked as complete
  • Redundant searches

This means each iteration builds on prior work rather than starting fresh.

Feedback Format

Types of Feedback You Can Provide

1. Explore deeper:

## User Feedback

The AI ethics section needs more depth:
- Current regulatory approaches (EU, US, etc.)
- Industry self-regulation vs government mandates
- Specific AI governance frameworks

2. New topic area:

## User Feedback

Also research:
- Supply chain risks of outsourced AI
- Labor implications of AI automation

3. De-prioritize topics:

## User Feedback

Good coverage on: history of AI

Skip further research on:
- Ancient history of computing
- Academic theoretical frameworks (unless directly relevant)

4. Fix inaccuracies:

## User Feedback

The report claims "AI adoption is 50% complete" but this seems high. 
Please verify this with recent enterprise adoption surveys.

Also, the section on "AI healthcare applications" missed regulatory barriers.

5. Change focus:

## User Feedback

The report is too academic. Focus more on:
- Practical use cases
- Real-world adoption barriers
- Implementation timelines

De-prioritize:
- Theoretical frameworks
- Historical background

Feedback Best Practices

Good feedback:

  • Specific (name the topic or section)
  • Action-oriented (what to explore)
  • Clear scope (what to skip)

Vague feedback:

  • "Make it better"
  • "More research"
  • "I didn't like the AI section"

Better: "The AI section focuses on hype. Please research actual production deployments, success rates, and lessons learned from real implementations."

Multi-Day Research

ResearchCrew remembers context across sessions:

Day 1:

crewai run  # Initial research

Review, provide feedback.

Day 2:

crewai run  # Iteration continues from prior context

The crew has full history from Day 1, so it knows what was researched and what feedback you gave.

You can distribute feedback and runs across days without losing context.

When Iteration Helps Most

Use iterative research when:

  • Topic is complex and requires multiple angles
  • You want to steer direction based on findings
  • Quality is more important than speed
  • You're building a comprehensive report

Use single-round when:

  • Simple, focused questions ("What is X?")
  • You just need quick facts
  • Topic doesn't require refinement

Optimization Tips

Efficient Iterations

Each iteration is a new full pipeline run. To speed things up:

  1. Batch feedback — Collect multiple feedback items before next run, rather than running after each small note

  2. Specific topics — Name specific sections/topics to explore, so the crew doesn't re-research completed areas

  3. Quality over quantity — 2-3 focused iterations produce better results than 10 scattered ones

Iteration Limit

A practical approach:

  • Round 1: Initial exploration (comprehensive)
  • Round 2: Address major gaps (targeted)
  • Round 3: Polish and verify (refinement)
  • Round 4+: Diminishing returns (fine-tuning)

Most research reaches diminishing returns after 3-4 rounds.

Exporting & Sharing

After you're satisfied with your research:

cp report.md my_research_2025_05_16.md

The report includes full citations, so others can:

  • Read your findings
  • Verify sources by clicking links
  • Trace claims back to original sources

Next Steps

  • Configuration — Optimize LLM choices and performance
  • Examples — See full multi-round workflow examples
  • Architecture — Understand how feedback is incorporated