The agent is asking Qloo…
Measured on 8 pairs (live Qloo API): the same model without Qloo named a film, artist or show that was in both people's lists 1 time in 24, and repeated one person's own favourite 12 times. One mixed Qloo call picked something outside one person's list 10 times in 24. This agent: in both lists 17 of 24, and "take turns" for the rest. How we measured
Show the agent's work: every Qloo call it made
| # | Why | Endpoint and parameters | ms |
|---|
How it works
- Find each favourite in Qloo. Each name goes to
/search. If the first match has no taste data, the agent tries the next one and tells you. - First pass, all five parts at once. Each person's own Qloo list (
/v2/insights): top 100 places in your city open on Saturday, top 200 films, artists and shows, with age bands and ratings for who is coming. Every candidate is scored for each person withfilter.results.entities. - Checks, then re-plan. Is a favourite's genre missing from that person's list (a trot fan whose list is K-pop)? Is lunch with a child at a place Qloo tags as serving alcohol? Is nothing in both lists? The planner re-asks Qloo where a check fails: within the genre, or restaurants tagged good for kids with no alcohol tags.
- Choose and explain. Shared only if it is in both lists; otherwise you take turns. A language model (NVIDIA Nemotron when configured) picks by reference from Qloo's results and writes one line per pick; every name, tag and rank in that line is checked against Qloo's data before you see it.
- Show the evidence. Both ranks on every card, what one mixed Qloo call would have picked, and every Qloo call one click away.
Every pick is a Qloo entity; the model cannot add a place or a film. With no model configured the same steps run under fixed rules, and the plan says so. Results describe what people with similar tastes tend to like; they are not a promise about any one person.
Who would use this: ticketing, family travel and streaming products that sell one plan to two or more people with different tastes.