🚀 Improved Ollama Review Guide

What's New

1. ✅ Project context - AI now understands your Pacific Disaster Platform
2. ✅ Better prompts - Focuses on real bugs, not style opinions
3. ✅ False positive filtering - Removes known bad advice automatically
4. ✅ Ensemble mode ready - Multi-model consensus for better accuracy

---

Quick Start

Best Quality (Recommended)

# Enable ensemble mode
echo "USE_ENSEMBLE=true" >> .env.local

# Run review
npm run ollama:review

Fast Review (Good for quick checks)

npm run ollama:fast

Review Specific File

npm run ollama:review -- src/components/YourComponent.tsx

---

Quality Improvements

Before (❌ Old Issues)

• False claims about aria-label
• Nitpicky style suggestions
• Generic "add comments" advice
• Wrong framework assumptions

After (✅ New Behavior)

• Focused on real bugs only
• Understands Next.js/React patterns
• Knows your project context
• Filters out false positives

---

How to Customize

1. Adjust Project Context

Edit .ollama-context.md to add:

• New patterns to ignore
• Specific things to check
• Project-specific conventions

2. Add More False Positive Filters

In scripts/ollama-review.js, add patterns to filterFalsePositives():

const falsePositivePatterns = [
  /your-pattern-here/i,
  // Add more...
];

3. Change Review Focus

Modify getReviewFocus() in the script to emphasize different areas.

---

Best Practices

✅ DO Use For:

• Pre-commit checks - Catch bugs before pushing
• Complex logic - Get second opinion on tricky code
• New features - Quick sanity check
• Learning - Understand potential issues

❌ DON'T Rely On For:

• Final say - Human review is still essential
• Framework expertise - AI can be wrong about React/Next.js
• Architecture - Complex decisions need human judgment
• Security audits - Use dedicated security tools

---

Troubleshooting

"Too many false positives"

1. Check .ollama-context.md has your patterns
2. Add more filters to filterFalsePositives()
3. Use ensemble mode for consensus

"Review too generic"

1. Add more specific context to .ollama-context.md
2. Update getReviewFocus() for your file types
3. Try different models (qwen2.5:14b is best)

"Missing real issues"

1. Enable ensemble mode
2. Reduce SKIP_TRIVIAL threshold
3. Review more files at once (context helps)

---

Model Recommendations

Task	Model	Why
Deep review	qwen2.5:14b-instruct	Best reasoning, catches real issues
Fast check	gemma2:2b	Quick, good for obvious bugs
Consensus	Multiple (ensemble)	Best accuracy, reduces false positives

---

Metrics

Track your improvements:

• Bugs caught before production
• False positives reduced
• Time saved in reviews
• Team agreement with suggestions

Good luck! 🚀