AI_MODEL_COMPARISON.md

# AI Model Comparison

## Current Implementation: Hugging Face (FREE)

### ✅ Advantages
- **100% Free** - No API costs ever
- **Offline** - Works without internet after initial download
- **Privacy** - All data stays on your server
- **No Rate Limits** - Analyze unlimited submissions
- **Open Source** - Full transparency

### ⚠️ Considerations
- **First Run**: Downloads ~1.5GB model (one-time)
- **Speed**: ~1-2 seconds per submission on CPU
- **Memory**: Needs ~2-4GB RAM
- **Accuracy**: ~85-90% for clear submissions

### Model Details
- **Model**: `facebook/bart-large-mnli`
- **Type**: Zero-shot classification
- **Size**: 1.5GB
- **Speed**: Fast on CPU, very fast on GPU

## Alternative: Anthropic Claude (PAID)

If you want to use Claude instead:

### 1. Install Anthropic SDK
```bash
pip install anthropic
```

### 2. Update `.env`
```
ANTHROPIC_API_KEY=your_api_key_here
```

### 3. Replace in `app/routes/admin.py`

**Remove this import:**
```python
from app.analyzer import get_analyzer
```

**Add this import:**
```python
from anthropic import Anthropic
import json
```

**Replace the analyze function (around line 234):**
```python
@bp.route('/api/analyze', methods=['POST'])
@admin_required
def analyze_submissions():
    data = request.json
    analyze_all = data.get('analyze_all', False)

    api_key = os.getenv('ANTHROPIC_API_KEY')
    if not api_key:
        return jsonify({'success': False, 'error': 'ANTHROPIC_API_KEY not configured'}), 500

    client = Anthropic(api_key=api_key)

    if analyze_all:
        to_analyze = Submission.query.all()
    else:
        to_analyze = Submission.query.filter_by(category=None).all()

    if not to_analyze:
        return jsonify({'success': False, 'error': 'No submissions to analyze'}), 400

    success_count = 0
    error_count = 0

    for submission in to_analyze:
        try:
            prompt = f"""Classify this participatory planning message into ONE category:

Categories: Vision, Problem, Objectives, Directives, Values, Actions

Message: "{submission.message}"

Respond with JSON: {{"category": "the category"}}"""

            message = client.messages.create(
                model="claude-sonnet-4-20250514",
                max_tokens=100,
                messages=[{"role": "user", "content": prompt}]
            )

            result = json.loads(message.content[0].text.strip())
            submission.category = result.get('category')
            success_count += 1

        except Exception as e:
            error_count += 1
            continue

    db.session.commit()

    return jsonify({
        'success': True,
        'analyzed': success_count,
        'errors': error_count
    })
```

### Claude Pros/Cons

✅ **Advantages:**
- Slightly higher accuracy (~95%)
- Faster (API response)
- No local resources needed

❌ **Disadvantages:**
- Costs money (~$0.003 per submission)
- Requires internet
- API rate limits
- Privacy concerns (data sent to Anthropic)

## Other Free Alternatives

### 1. Groq (Free Tier)
- API similar to Anthropic
- Free tier: 30 requests/min
- Very fast inference

### 2. Together AI (Free Credits)
- $25 free credits monthly
- Various open source models

### 3. Local Llama Models
- Use Ollama or llama.cpp
- Slower but powerful
- Need more RAM (8GB+)

## Recommendation

**For most users**: Stick with **Hugging Face** (current implementation)
- Free forever
- Good accuracy
- Privacy-focused
- No API complexity

**For mission-critical**: Use **Anthropic Claude**
- Higher accuracy
- Professional support
- Worth the cost for important decisions

**For developers**: Try **Groq free tier**
- Fast
- Free (with limits)
- Easy to integrate