Lumenstay
2–3× revenue from engaged sessions
Travel · Enterprise · Cloud · North America
Lumenstay’s planner used to keyword-match reviews and then hand a pile of snippets to a model. Guests who engaged with the planner spent more, but the retrieval step was the slow and noisy part.
They store one point per review passage, with a second vector for the hero photo. Payload fields carry city, season, and language. The query filters those fields while the graph is walked.
A rerank step prefers passages that mention the constraint the guest actually typed, not just a nearby neighborhood. The index updates as new reviews land, so a weekend’s bookings are visible on Monday.
Engaged sessions now produce two to three times the revenue of sessions that never open the planner. The figure is part of this fictional story.