FIND · Vibeset

Curation

AI setlist generation over a licensed, deeply-tagged catalog

The engine that started Vibeset. Describe a vibe and it builds the setlist: tempo-matched, harmonically compatible, shaped like an evening. Underneath sits a Postgres + pgvector catalog and an LLM ensemble to finish.

search
1–150 ms
modes
sql · hybrid · semantic
sequencing
bpm + key + energy arc

Open the DJ app ↳

Three ways to answer one query

Most searches don't need an embedding model. The ones that do are ruined by exact matching. So the router reads the query and picks. The axis is logarithmic: the cheap path really is a hundred times cheaper.

sql 1–10 ms

The query names something exactly

artist = "Bicep", 124–128 bpm

Indexed columns answer it. No model runs.

hybrid 50–150 ms

Part literal, part description

"like Bicep but slower"

Text match narrows it down, then vectors rank what is left.

semantic 150–300 ms

No literal to match on

"last hour of a warehouse set"

Embedding search across the catalog. The expensive path, earned.

What's underneath

  • Search router

    Three retrieval strategies behind one query box

    Most catalog searches don’t need an embedding model, and exact matching ruins the ones that do. The router classifies each query and dispatches it: exact SQL for structured filters, hybrid retrieval when the query is partly literal, semantic vector search when it’s a description rather than a name. Latency spans two orders of magnitude across those paths, which is the point: the cheap path stays cheap.

    • Query complexity classification
    • Exact SQL filtering for structured queries
    • Hybrid lexical + vector retrieval
    • Semantic embedding search over pgvector
    • BPM and key awareness
    • Metadata-aware ranking

    Retrieval-Augmented Generation · Recommendation Systems · Data Architecture · Evaluation & Experimentation

  • Setlist reasoning

    Ordering that a DJ would recognise as deliberate

    Retrieval gives you the right tracks. It doesn’t give you an order. Sequencing weighs key compatibility, tempo continuity and where a set should peak, then commits to an arrangement it can justify. The hard part isn’t producing an order. It’s producing one a working DJ wouldn’t immediately rearrange.

    • Multi-factor track ordering
    • Harmonic compatibility between neighbours
    • Energy-arc shaping across a set
    • Tempo continuity constraints

    LLM Orchestration · Evaluation & Experimentation · Recommendation Systems

  • Set extender

    Keep the night going without breaking its shape

    Extending a set is not the same problem as building one. The arc already exists, the room is already somewhere specific, and new tracks have to arrive without resetting either.

    • Extends an existing set in keeping with its arc
    • Energy-curve preservation
    • Tempo-continuous transitions

    LLM Orchestration · Recommendation Systems · Data Architecture

  • Setlist editor

    Hands on the output, not just the prompt

    Generated sets are a starting point, and any tool that treats them as final loses the people who know the most. The editor lets a DJ overrule the model track by track, then re-checks the transitions around whatever they changed rather than regenerating the whole thing.

    • Track-level editing
    • Transition conflict warnings
    • Flow-preserving suggestions after an edit

    Human-Centered AI · Recommendation Systems · Evaluation & Experimentation