mappity
Map where you ask anything in plain words and Jev gives every nearby place a probability, lighting up those likely to match.
Overview
mappity lets you ask a map anything, such as 'my kid is melting down and needs sugar immediately', and about a second later nearby chocolate shops light up. There are no categories, filters, search index or chat model. Cactus Needle, a small local model, extracts commands and hard constraints from the sentence, such as 'near the aquarium'. The judge, hosted Jev by default or a self-hosted Shingi 27B, answers typed questions with probabilities: what kind of request this is, whether the street matters, and whether each thing Needle extracted is really a place (the README shows 'sugar' at 1% and 'aquarium' at 92%).
Code geocodes landmarks with Nominatim, turns walking minutes into a radius, fetches places from OpenStreetMap and, when the street matters, imagery from Mapillary, then asks the judge about each place and combines scores. Jev handles wide parallel batches; with Shingi, mappity sends one place per request with up to eight in flight, and the README notes Shingi's calibration has not been validated for this app. A recorded demo and a short narrated film are available. It needs Node 22+ and a Mapillary token.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What kind of request is this? | — |
| Yes / no | Is this extracted term really a place? | — |
| Yes / no | Does this place satisfy the user's wish?(inferred) | — |
How it works
- Needle extracts commands and constraints
- Judge classifies the request and validates extracted places
- Code geocodes and computes a walking radius
- Places are fetched from OpenStreetMap (and Mapillary when needed)
- Judge scores each place; the map glows where the answer is yes
Strengths
- No predefined categories needed
- Clear division between extractor, judge and code
- Self-hosted judge option
Caveats
- Requires a Mapillary token and judge provider
- Shingi calibration not validated
- Demo-stage
Getting started
npm install, cp .env.example .env (add keys), ./run.sh and open http://localhost:4321.At a glance
- Type
- Demo
- Status
- prototype
- Language
- JavaScript
- Runtime
- Node 22+ server and web map
- Has code
- yes
Use cases and integrations
How curated lists describe it
Ask the map anything. Every place answers with a probability. OpenStreetMap + Mapillary + TypeSafe Jev + Cactus Needle.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
The request type, whether extracted terms are places, and how likely each place matches your wish.
Is there a chat model?
No, a small extractor, a judge and code do the work.
Can I avoid TypeSafe?
Yes, set JEV_PROVIDER=shingi to use a self-hosted Shingi 27B.
Related projects
- Jev Search★ 534
Plain-language web search where a decision model like Jev picks query, sources and time range, then ranks Search1API results.
AppproductionTypeScript - jev-seo★ 610
Live SEO audit from one homepage URL: crawl, 52 rules, Core Web Vitals and Jev page judgments, reported as PDF, XLSX and Markdown.
PluginbetaPython Self-hostable Reddit buyer-intent finder that has Jev judge every post and comment it reads, with a written reason and cost per lead.
ProjectproductionTypeScriptLocal-first feed that merges arXiv, GitHub, Hacker News and Bilibili results, with optional Jev title-relevance scores as a filter.
ProjectbetaTypeScriptSelf-hosted MCP research server that filters search results and page blocks with Jev, returning compact evidence with provenance.
MCP serverbetaMCP server that scores every search query, result and fetched page an agent sees for relevance, credibility and steering risk.
MCP serverbetaTypeScript