Photo Geolocator

Extract date, time, and GPS location from photos via EXIF metadata, then reverse geocode coordinates to street addresses and venue names using OpenStreetMap Nominatim.

When to use this skill

• User uploads photos and asks "where was this taken?" or "when was this?"
• Any workflow that needs to go from photo files → dates and locations
• Another skill (like the wine journal) needs date/location extraction as a building block
• User wants to map a set of photos to venues, restaurants, cities, or trips

How it works

Step 1: Extract EXIF data

Run the bundled extraction script on the photo directory or individual files:

pip install Pillow --break-system-packages -q
python3 <skill-path>/scripts/extract_exif.py <directory_or_file>

This outputs JSON with file, date, lat, lon for each image. Dates are in YYYY:MM:DD HH:MM:SS format. GPS coordinates are decimal degrees (positive = N/E, negative = S/W).

Not all photos have EXIF data — phone cameras almost always do, but screenshots, downloaded images, and some edited photos won't. When GPS is missing, note it as "Unknown location" rather than guessing.

Step 2: Cluster nearby coordinates

Photos taken at the same venue will have slightly different GPS readings (within ~50m). Before reverse geocoding, cluster coordinates that are within 0.001° of each other (roughly 100m) to avoid redundant API calls.

# Cluster logic: group points within 0.001° lat AND 0.001° lon
clusters = {}
for point in exif_results:
    if point['lat'] is None:
        continue
    matched = False
    for key in clusters:
        if abs(point['lat'] - key[0]) < 0.001 and abs(point['lon'] - key[1]) < 0.001:
            clusters[key].append(point)
            matched = True
            break
    if not matched:
        clusters[(point['lat'], point['lon'])] = [point]

Step 3: Reverse geocode

Run the bundled reverse geocoding script on the unique coordinate clusters:

python3 <skill-path>/scripts/reverse_geocode.py <coords.json>

Where coords.json is a JSON array of {"lat": x, "lon": y} objects. The script respects Nominatim's 1-request-per-second rate limit automatically.

The output includes name (venue name if OSM knows it), road, house_number, neighbourhood, city, country, and amenity.

Step 4: Identify specific venues

Nominatim gives you the street address, but often not the specific restaurant or bar. To identify the actual venue:

1. Check the name and amenity fields first — sometimes Nominatim returns the venue directly (e.g., "Andrew Edmunds", "Le Dive")
2. For addresses without a venue name, use WebSearch to search for restaurants/bars at that specific address: restaurant "123 Main Street" [city]
3. If the photo itself shows a venue name (menu, signage, receipt), that's the most reliable identification
4. For coordinates in residential areas visited repeatedly across many dates, these are likely the user's home — flag as "Home ([neighborhood])" rather than searching for restaurants
5. For coordinates on park drives or waterfront walkways, these are likely outdoor picnics — flag as "[Park name] (picnic)" or "Waterfront (outdoor)"

Step 5: Present results

Return a structured summary with: date (reformatted to YYYY-MM-DD), city/neighborhood, specific venue (if identified), and the street address as fallback.

Tips for accuracy

• Recurring locations: If the same GPS cluster appears across many different dates (5+ visits), it's almost certainly the user's home or office, not a restaurant.
• Same-day clusters: Multiple photos at the same coordinates on the same day are the same venue visit. Group them.
• Sequential photos: Photos taken minutes apart at the same coordinates are the same event (e.g., front and back of a wine label, multiple dishes at dinner).
• Travel patterns: Look at the dates and cities holistically to understand trips. E.g., Dublin → London on consecutive days = a Europe trip.

Dependencies

• Python 3 with Pillow (pip install Pillow --break-system-packages)
• Internet access for Nominatim reverse geocoding and WebSearch for venue identification