Mapbox Search Patterns Skill

Expert guidance for AI assistants on using Mapbox search tools effectively. Covers tool selection, parameter optimization, and best practices for geocoding, POI search, and location discovery.

Available Search Tools

1. search_and_geocode_tool

Best for: Specific places, addresses, brands, named locations

Use when query contains:

• Specific names: "Starbucks on 5th Avenue", "Empire State Building"
• Brand names: "McDonald's", "Whole Foods"
• Addresses: "123 Main Street, Seattle", "1 Times Square"
• Chain stores: "Target"
• Cities/places: "San Francisco", "Portland"

Don't use for: Generic categories ("coffee shops", "museums")

2. category_search_tool

Best for: Generic place types, categories, plural queries

Use when query contains:

• Generic types: "coffee shops", "restaurants", "gas stations"
• Plural forms: "museums", "hotels", "parks"
• Is-a phrases: "any coffee shop", "all restaurants", "nearby pharmacies"
• Industry terms: "electric vehicle chargers", "ATMs"

Don't use for: Specific names or brands

3. reverse_geocode_tool

Best for: Converting coordinates to addresses, cities, towns, postcodes

Use when:

• Have GPS coordinates, need human-readable address
• Need to identify what's at a specific location
• Converting user location to address

Tool Selection Decision Matrix

Parameter Guidance

Proximity vs Bbox vs Country

Three ways to spatially constrain search results:

1. proximity (STRONGLY RECOMMENDED)

What it does: Biases results toward a location, but doesn't exclude distant matches

Use when:

• User says "near me", "nearby", "close to"
• Have a reference point but want some flexibility
• Want results sorted by relevance to a point

Example:

{
  "q": "pizza",
  "proximity": {
    "longitude": -122.4194,
    "latitude": 37.7749
  }
}

Why this works: API returns SF pizza places first, but might include famous NYC pizzerias if highly relevant

Critical: Always set proximity when you have a reference location! Without it, results are IP-based or global.

2. bbox (Bounding Box)

What it does: Hard constraint - ONLY returns results within the box

Use when:

• User specifies an area: "in downtown", "within this neighborhood"
• Have a defined service area
• Need to guarantee results are within bounds

Example:

{
  "q": "hotel",
  "bbox": [-122.51, 37.7, -122.35, 37.83] // [minLon, minLat, maxLon, maxLat]
}

Why this works: Guarantees all hotels are within SF's downtown area

Watch out: Too small = no results; too large = irrelevant results

3. country

What it does: Limits results to specific countries

Use when:

• User specifies country: "restaurants in France"
• Building country-specific features
• Need to respect regional boundaries
• Or it is otherwise clear they want results within a specific country

Example:

{
  "q": "Paris",
  "country": ["FR"] // ISO 3166 alpha-2 codes
}

Why this works: Finds Paris, France (not Paris, Texas)

Can combine: proximity + country + bbox or any combination of the three

Decision Matrix: Spatial Filters

Setting limit Parameter

category_search_tool only (1-25, default 10)

Performance tip: Lower limits = faster responses

types Parameter (search_and_geocode_tool)

Filter by feature type:

Example combinations:

// Only POIs and addresses, no cities
{"q": "Paris", "types": ["poi", "address"]}
// Returns Paris Hotel, Paris Street, not Paris, France

// Only places (cities)
{"q": "Paris", "types": ["place"]}
// Returns Paris, France; Paris, Texas; etc.

Default behavior: All types included (usually what you want)

auto_complete Parameter (search_and_geocode_tool)

What it does: Enables partial/fuzzy matching

Example:

// User types "starb"
{ "q": "starb", "auto_complete": true }
// Returns: Starbucks, Starboard Tavern, etc.

Use for:

• Search-as-you-type interfaces
• Handling typos ("mcdonalds" -> McDonald's)

• Incomplete queries

Don't use for:

• Final/submitted queries (less precise)
• When you need exact matches

Anti-Patterns to Avoid

Don't: Use category_search for brands

// BAD
category_search_tool({ category: 'starbucks' });
// "starbucks" is not a category, returns error

// GOOD
search_and_geocode_tool({ q: 'Starbucks' });

Don't: Use search_and_geocode for generic categories

// BAD
search_and_geocode_tool({ q: 'coffee shops' });
// Less precise, may return unrelated results

// GOOD
category_search_tool({ category: 'coffee_shop' });

Don't: Forget proximity for local searches

// BAD - Results may be anywhere globally
category_search_tool({ category: 'restaurant' });

// GOOD - Biased to user location
category_search_tool({
  category: 'restaurant',
  proximity: { longitude: -122.4194, latitude: 37.7749 }
});

Don't: Geocode ambiguous place names without proximity (REST too)

This applies to Mapbox Geocoding API v5 / Search Box in browser apps — not only MCP tools.

// BAD — limit=1 without proximity can resolve "Lincoln Memorial" to Illinois
fetch(`https://api.mapbox.com/geocoding/v5/mapbox.places/${encodeURIComponent(q)}.json?access_token=${token}&limit=1`);

// GOOD — bias to map center (and optional bbox)
fetch(
  `https://api.mapbox.com/geocoding/v5/mapbox.places/${encodeURIComponent(q)}.json` +
    `?access_token=${token}&proximity=-77.0369,38.9072&bbox=-77.15,38.79,-76.90,38.99&limit=1`
);

Also debounce search inputs (clearTimeout + setTimeout) so every keystroke does not fire a geocode.

Don't: Use bbox when you mean proximity

// BAD - Hard boundary may exclude good nearby results
search_and_geocode_tool({
  q: 'pizza',
  bbox: [-122.42, 37.77, -122.41, 37.78] // Tiny box
});

// GOOD - Bias toward point, but flexible
search_and_geocode_tool({
  q: 'pizza',
  proximity: { longitude: -122.4194, latitude: 37.7749 }
});

Don't: Request ETA unnecessarily

// BAD - Costs API quota for routing calculations
search_and_geocode_tool({
  q: 'museums',
  eta_type: 'navigation',
  navigation_profile: 'driving'
});
// User didn't ask for travel time!

// GOOD - Only add ETA when needed
search_and_geocode_tool({ q: 'museums' });
// If user asks "how long to get there?", then add ETA

Don't: Set limit too high for UI display

// BAD - Overwhelming for simple dropdown
category_search_tool({
  category: 'restaurant',
  limit: 25
});
// Returns 25 restaurants for a 5-item dropdown

// GOOD - Match UI needs
category_search_tool({
  category: 'restaurant',
  limit: 5
});

Quick Reference

Tool Selection Flowchart

User query contains...

-> Specific name/brand (Starbucks, Empire State Building)
  -> search_and_geocode_tool

-> Generic category/plural (coffee shops, museums, any restaurant)
  -> category_search_tool

-> Coordinates -> Address
  -> reverse_geocode_tool

-> Address -> Coordinates
  -> search_and_geocode_tool with types: ["address"]

Essential Parameters Checklist

For local searches, ALWAYS set:

• proximity (or bbox if strict boundary needed)

For category searches, consider:

• limit (match UI needs)
• format (json_string if plotting on map)

For disambiguation, use:

• country (when geographic context matters)
• types (when feature type matters)

For travel-time ranking:

• eta_type, navigation_profile, origin (costs API quota)

Common Mistakes

1. Forgetting proximity -> Results are global/IP-based (or wrong state for ambiguous memorial/park names)
2. Using wrong tool -> category_search for "Starbucks" (use search_and_geocode)
3. Invalid category -> Check category_list first
4. Bbox too small -> No results; use proximity instead
5. Requesting ETA unnecessarily -> Adds API cost
6. Limit too high for UI -> Overwhelming user
7. Not filtering types -> Get cities when you want POIs
8. No debounce on typeahead -> Quota burn and racy UI

Reference Files

Load these for deeper guidance on specific topics:

• references/advanced-params.md — poi_category, ETA, format, and language parameters
• references/workflows.md — Common patterns: Near Me, Branded, Geocoding, Category+Area, Reverse, Route-Based, Multilingual
• references/optimization-combining.md — Performance optimization, combining tools, handling no results, category list resource