Exploring 'HippoMaps', a map app that uses a deterministic parser instead of LLMs (Large Language Models) to perform ultra-fast searches on OpenStreetMap in everyday language, and the implications of its related technology.
Imagine this. You went on a weekend trip to Zurich, Switzerland. While walking down the streets of an unfamiliar city, you suddenly crave a warm cup of coffee. Instead of opening a smartphone map app and searching for store names one by one as usual, you type what comes to mind just as it is: “cafes near the Limmat River in Zurich 100 meters from the lake”. A map app has emerged that can pinpoint this complex sentence—which would likely cause errors or pull up completely wrong locations on any major map app unless it’s a tech giant’s massive map app—right on the map in just a few seconds [Source 1].
This is the story of HippoMaps, an open-source map and navigation app that has recently gained significant attention after being featured on the online tech community Hacker News [Source 1, Source 8]. In today’s tech industry swept by the AI wave, this project—which implemented natural language geographic search solely through sophisticated rules and algorithms without the help of AI models—offers us a refreshing perspective on technology [Source 1].
Why does this matter?
Services we use every day like Google Maps or Naver Maps provide vast amounts of location data and convenient search features worldwide. However, behind them lies a massive treasure trove known as OpenStreetMap (OpenStreetMap, a free worldwide map created together by people around the world, OSM for short), an open-source map ecosystem [Source 5, Source 9]. While OpenStreetMap is a public geographic information repository that anyone can freely use and modify, it has had one critical weakness: its search feature is not as user-friendly or fast as commercial apps [Source 1].
Not only for general users but also for investigative journalists, OpenStreetMap is an immense source of information. However, finding desired data previously required proficiency in Overpass Query Language (OverpassQL for short, a dedicated data search command for OpenStreetMap), a complex database command language [Source 3, Source 5]. Because of this, ordinary people and non-experts found it difficult to properly utilize this vast treasure trove [Source 3, Source 5].
The significance of projects like HippoMaps is that they open a path where ordinary people who don’t know programming or database syntax can speak to the map in their everyday speaking style to accurately find the location they want [Source 1]. It proves that even without massive server costs or the brainpower of giant AI models, smart design alone can allow anyone to enjoy a fast and convenient map service [Source 1].
Understanding Simply: The Secret of the ‘Smart Translator’ Operating Without AI
Most modern technologies today use massive AI models like ChatGPT (LLMs) to understand natural language spoken by humans [Source 1]. However, AI has drawbacks: it occasionally gives odd answers (hallucination), takes time to produce answers, and incurs high server costs.
Instead of a massive cannon like AI, HippoMaps chose a method like a finely tuned “Swiss Army Knife” [Source 1]. To put it simply:
An easy analogy:
If search using a large AI model is a “simultaneous interpreter” who deeply ponders context every time and attempts impromptu translation, HippoMaps’ method is like a “highly mechanical automated classification assistant” that holds strict grammar rules and word cards in hand, breaks down input sentences unit by unit, and finds the correct answer at maximum speed.
HippoMaps breaks down sentences entered by users in the search bar into fine tokens (tokens, minimum units of meaning) [Source 1]. Then, by cross-referencing word lists (famous brands, sports categories, OpenStreetMap tags, etc.), it figures out ‘What’ the user precisely wants to find [Source 1]. At the same time, using the geocoder service Photon (a search service that converts addresses or place names into coordinates), it determines ‘Where’ the user specified [Source 1].
In short, it analyzes the skeletal structure of the sentence to instantly grasp, “Ah, this user is looking for a ‘cafe’ in the region of ‘Zurich’! And they attached conditions like ‘near the Limmat River’ and ‘within 100 meters of the lake’“, and then seamlessly translates this into Overpass Query Language that OpenStreetMap can understand [Source 1]. Because this process bypasses AI inference steps, search speed is remarkably fast and the probability of displaying erratic results is low [Source 1].
Current Status: The Evolution of Open Source Maps and Natural Language Search
Within the OpenStreetMap ecosystem today, various ongoing attempts are being made to help anyone easily perform complex data queries. Representatively, academic and developer communities have actively researched Text-to-OverpassQL tasks that convert natural language questions into OpenStreetMap-dedicated search commands [Source 2, Source 3].
Additionally, open-source tools like SPOT have been released and received positive feedback, helping journalists and researchers search OpenStreetMap data by describing scenes in everyday language [Source 5]. For example, if you input a sentence like “Find the Italian restaurant right next to the bus stop in Paris”, the system understands it and filters the data accordingly [Source 5].
HippoMaps took this trend a step further. Rather than just mimicking search features, it packed almost all essential features expected in a map app into a single open-source package [Source 1]:
- Ultra-fast Custom Search Engine: Equipped with a custom Overpass API engine, it provides fast POI (Point of Interest), location, and route searches that rival commercial map apps [Source 1].
- Diverse Exploration Features: It seamlessly handles creative and complex condition-combined searches, such as “pharmacies currently open in Berlin”, “peaks over 4000 meters in Valais”, or “bridges of the Danube River in Serbia” [Source 1].
- Comprehensive Navigation Features: Moving beyond mere search, it supports voice-enabled turn-by-turn navigation via Valhalla (an open-source routing engine using the Piper speech synthesis engine), European public transit routing (MOTIS), and real-time traffic information in select countries [Source 1].
- Open Source Release: All these high-quality features are released under the GPL-3.0 open-source license, allowing anyone to inspect code and contribute [Source 1, GitHub repository: https://github.com/cdutsov/hippomaps].
What Lies Ahead?
Experts and open-source developers have high expectations for the advancement of this deterministic natural language search approach [Source 1]. Even without consuming massive server costs and power to maintain large AI models, combining rule-based smart parsers and well-organized geographic databases can create lightweight and powerful local navigation apps even in smartphone environments [Source 1].
In the future, even in remote areas or travel destinations with unstable internet connections, independent map apps will increasingly emerge where users can comfortably whisper directions in their native language to instantly receive customized routes based on worldwide OpenStreetMap data [Source 1, Source 16]. It is a time of change where public map assets created through cooperation among global developers are stepping closer into daily life, standing against the monopoly of big tech corporations [Source 1, Source 9].
MindTickleBytes AI Reporter’s Perspective
Implementing fast and accurate geospatial search without relying solely on complex AI models, but through sophisticated rules and token decomposition, demonstrates the fine craftsmanship of software engineering. It is a meaningful attempt leading the popularization of map data.
References
- Show HN: Fast deterministic natural language search for OpenStreetMap data — https://hippomaps.ch/?q=cafes+by+the+Limmat+100+m+from+the+lake+in+zurich#map=15.02/47.36708/8.54935, https://github.com/cdutsov/hippomaps [Source 1]
- Text-to-OverpassQL: A Natural Language … - ar5iv — https://ar5iv.labs.arxiv.org/html/2308.16060 [Source 2]
- Text-to-OverpassQL: A Natural Language Interface … - arXiv — https://arxiv.org/abs/2308.16060 [Source 3]
- Text-to-OverpassQL: A Natural Language Interface for Complex … (PDF) — https://aclanthology.org/2024.tacl-1.31.pdf [Source 4]
- New natural language search interface for OSM (public demo, open source) — https://community.openstreetmap.org/t/new-natural-language-search-interface-for-osm-public-demo-open-source/134816 [Source 5]
- Text-to-OverpassQL: A Natural Language Interface for Complex … (ACL Anthology) — https://aclanthology.org/2024.tacl-1.31/ [Source 6]
- Text-to-OverpassQL: A Natural Language Interface for Complex … (Schumann Pub) — https://www.schumann.pub/assets/uploads/publications/pdfs/staniek_2023.pdf [Source 7]
-
HN.watch Hacker News with explainer videos — https://hn.watch/ [Source 8] - OpenStreetMap — https://www.openstreetmap.org/ [Source 9]
- Open source geocoding with OpenStreetMap data — https://nominatim.org/ [Source 10]
- Inference using Large Language Models (LLMs)
- A deterministic parser that does not use AI
- Cloud-only artificial neural network translation
- Overpass Query Language (OverpassQL)
- SQL database language
- Python geo-query
- Professional programmers
- Professional journalists
- Navigation hardware engineers