# Enabling AI Prompt Generation

The Daily Check-in system can use AI to automatically:
- Parse freeform text ("I take vitamin D every morning")
- Create structured entries (medications, vitals, exercise, etc.)
- Generate daily tracking prompts with pre-filled values
- Learn patterns from repeated entries

## Current Status

✅ Code is ready and deployed
❌ Gemini API key not configured

## Setup Instructions

### 1. Get a Gemini API Key

1. Go to https://aistudio.google.com/app/apikey
2. Create a new API key (free tier available)
3. Copy the key

### 2. Configure on Staging

```bash
# On staging server (192.168.1.253)
echo "GEMINI_API_KEY=YOUR_KEY_HERE" > /tank/inou/api/anthropic.env

# Restart API service
ssh johan@192.168.1.253 "sudo systemctl restart inou-api"
# Or use the standard restart:
ssh johan@192.168.1.253 "/tank/inou/stop.sh && /tank/inou/start.sh"
```

### 3. Configure on Production

```bash
# On production server (192.168.100.2)
echo "GEMINI_API_KEY=YOUR_KEY_HERE" > /tank/inou/api/anthropic.env

# Restart API service
ssh johan@192.168.100.2 "sudo systemctl restart inou-api"
```

### 4. Verify

Check the API logs to confirm the key loaded:

```bash
ssh johan@192.168.1.253 "tail -20 /tank/inou/logs/api.log | grep -i gemini"
```

You should see: `Gemini API key loaded.`

## Testing

Once configured, test by entering freeform text on the prompts page:

**Test inputs:**
- "I take vitamin D 5000 IU every morning"
- "Blood pressure 120/80"
- "Walked 30 minutes today"

**Expected behavior:**
1. Entry is created immediately
2. System analyzes the pattern
3. If it detects a trackable pattern, it creates a new daily prompt
4. You see a notification: "✓ Saved! Added new tracking prompt."
5. The new prompt appears on the page for tomorrow

## How It Works

### Architecture

```
User Input → Triage (category detection) → Extraction (structure) → Create entries + prompts
```

### Prompts

AI prompts are stored in `/tank/inou/prompts/`:
- `triage.md` - Determines category (medication, vital, exercise, etc.)
- `extract_*.md` - Category-specific extraction (one per category)

### Code

- `api/api_llm.go` - Main LLM integration
- `api/api_prompts.go:192` - `tryGeneratePromptFromFreeform()` entry point
- `lib/llm.go` - Gemini API wrapper

## Cost Estimate

Gemini 1.5 Flash (used by default):
- **Free tier**: 15 requests/minute, 1500 requests/day
- **Paid tier**: $0.00001875/1K characters input, $0.000075/1K characters output

For typical use (10-20 prompts/day), free tier is sufficient.

## Alternative: Claude API

The code can be adapted to use Claude instead of Gemini. Would need:
1. Replace `lib.CallGemini()` with Claude API calls
2. Update prompt formatting (Claude uses different message format)
3. Set `ANTHROPIC_API_KEY` instead

## Fallback Without AI

If no API key is configured:
- Freeform entries still work (saves as plain notes)
- No automatic prompt generation
- Users can manually create prompts via the UI (not yet implemented)
- Or add prompts directly to database

## Next Steps

1. **Get API key** - Start with Gemini free tier
2. **Test on staging** - Verify prompt generation works
3. **Monitor usage** - Check if free tier is sufficient
4. **Consider Claude** - If Gemini isn't accurate enough
5. **Build manual prompt creator** - Fallback for users without AI