GroupTaste

Inspiration

Groups pick a bar by whoever talks loudest. GroupTaste makes the trade-off visible.

Built for: Qloo Agentic Hackathon · MIT licensed

Problem

Groups pick bars and restaurants by whoever talks loudest. That place is great for one person and a bad night for someone else. Nobody can see that until they're already there.

What it does

Each person lists a few things they love (artists, films, shows, books, brands). GroupTaste:

1. Resolves each "love" to a Qloo entity (search, keeping cultural entities only, never places).
2. Proposes candidate venues: each member's top Qloo recommend places in the city (optional bar/restaurant tag). Places whose geocoded city isn't the chosen city are dropped, because filter_location is a radius, not a city boundary.
3. Scores every candidate against each member's taste separately with Qloo rank, so every person gets their own affinity for the same set of real places.
4. Decides by maximin: the venue with the best least-happy-member score, with mean affinity as the tie-breaker.
5. Shows the evidence: a per-member affinity table, plus what "pick for the first person" would have chosen and who that choice would leave out.

Sample run (live API, New York, bar; Ana: Radiohead/Wes Anderson/Sideways · Ben: Metallica/Fast & Furious/Guy Fieri · Chi: Taylor Swift/Gilmore Girls/Pride and Prejudice): Igloo Bar NYC. Affinities: Ana 0.86 · Ben 0.85 · Chi 0.87. The least-happy member is at 0.85.

Qloo workflow used

@qloo/qloo-harness (qloo api search, qloo exec recommend, qloo exec rank) against https://hackathon.api.qloo.com.

• Entity ambiguity: the harness's needs_input candidates are resolved deterministically (city match first, then popularity) and re-run once.
• Request to result: likes → entity ids → recommend {target_type: place, signals: [ids], filter_location, include_tags} → candidate place ids → rank {options: place ids, option_type: place, signals: member ids} per member → maximin.

Run it

npm i -g @qloo/qloo-harness            # needs Node >= 22
export QLOO_API_KEY=...  QLOO_BASE_URL=https://hackathon.api.qloo.com  QLOO_TRUSTED_BASE_URL=$QLOO_BASE_URL
node server.mjs                        # http://localhost:8809
GROUPTASTE_FIXTURES=fixtures/demo.json node server.mjs   # keyless replay of a recorded live run
node test.mjs                          # decision-logic test

Optional GROUPTASTE_CACHE=cache.json persists Qloo answers, so repeated groups don't spend the key's rate budget.

Limits

• Affinity is Qloo's taste signal, not a forecast of anyone's evening. Price, distance and opening hours are not considered.
• The hackathon key is rate-limited, so a brand-new group takes about a minute. Calls are serialized, with backoff on QLOO_RATE_LIMIT.
• Maximin protects the least-happy member, and it can pass over a place two people love that the third merely tolerates. That trade-off is the point, and it's shown in the table.
• No personal data is stored. Member names are free text and only stay in memory for the request.

Built With

• html
• javascript
• node.js
• qloo