Skip to main content

Three Modes of Search

Public beta

CORPUS is being built in the open. Some of what you read here is live, some is still design intent — expect it to evolve.

People come to a music corpus knowing what they want in very different ways. A music supervisor with a brief might have a precise sound in mind. A composer might have a reference track and want something like it but darker. A director might have a scene and no idea what music it needs. A single search field treats all three the same; CORPUS treats them as different problems.

The demo at intelligence.corpus.music routes each query to one of three modes. The session locks to its mode so follow-ups stay coherent.

For when you know the sound. The query becomes an explicit set of constraints — genres, moods, instruments, key, vocal treatment — and the corpus returns ranked tracks that match. Refinements ("more sparse", "no piano") update the constraints without leaving the mode.

"Dark jazz, no vocals, only piano, slow."

Best for: briefs with a clear sonic specification, A/B comparisons across known parameters, catalogue navigation by a user who already knows the space.

The CORPUS chat interface in Precision Search mode. The user has typed 'triumphant with something heavy underneath'. The system shows the resolved search query 'triumphant heavy powerful epic', a short composer briefing describing the kind of music being looked for, and a ranked list of matching tracks (Warmaggedon, Gemini Instrumental, two more below) each with BPM, key, and instrument tags.
Precision search at intelligence.corpus.music. The user types a brief in natural language; the system resolves it to a search query, generates a short composer briefing of what it is looking for, and returns ranked tracks with the metadata that lets the user decide.

Audio similarity​

For when you have a reference. Upload a track — your own, a temp from an edit, a piece you wish you could license. The system analyses BPM, key, instrumentation, energy, and vocal treatment, then returns similar tracks from the corpus. Relative shifts work too: "similar but darker", "similar but at half tempo", "similar but instrumental".

Best for: replacing a temp track, finding alternatives to a piece you cannot clear, exploring a corner of the corpus anchored by a known example.

The CORPUS chat interface in Similarity Search mode. The user has uploaded an audio file called 'Landscape Of Hope.wav'. The system shows a Reference Analysis panel reporting BPM 122 and Key C minor, followed by ranked tracks from the corpus (1106462, Home World by Misha Dioxin, 682582) all matching at BPM 120 in minor keys, with instrument tags.
Similarity search. The system analyses the uploaded reference (here a track called Landscape Of Hope) and returns ranked tracks from the corpus that match on tempo, key, and instrumentation — surfacing the underlying acoustic profile rather than relying on text tags.

Story2Music — narrative search​

For when you have a scene, not a sound. Paste a screenplay excerpt, describe a beat in a film, sketch a moment in a game. The system reads the scene and returns three dramaturgically distinct directions, each with:

  • a short title naming the direction,
  • keywords describing the sonic and atmospheric character,
  • a what it adds note explaining how the music transforms the scene,
  • a semantic search query the system runs against the corpus to retrieve matching tracks.

The three directions are deliberately different from each other — orchestral tension versus stripped-down intimacy versus electronic propulsion, for example. The value of the mode is the contrast between them: readings of the scene the user would never have searched for on their own. Rejected directions are remembered so the system does not repeat them.

Best for: creative briefs without a fixed sound, early exploration in a film or game scoring process, finding music for a feeling that does not yet have a vocabulary.

The CORPUS chat interface in Narrative Search mode. The user has typed 'Wedding, first dance, goosebumps'. The system has asked a clarifying question with three options for what kind of moment this is; the user picked option 1 (a tender, intimate moment). The system then returns three columns labelled 'Timeless Waltz', 'Pulse of Now', and 'Fragile Memory' — each with its own keyword set (orchestral/romantic/strings vs. minimalist/pulsing/ambient vs. piano/solo/melancholic), a one-sentence 'what it adds' description, and matching tracks from the corpus under each direction.
Narrative search. A short brief — Wedding, first dance, goosebumps — is resolved through a clarifying question, then returned as three dramaturgically distinct readings of the scene: Timeless Waltz, Pulse of Now, Fragile Memory. The contrast is the point: three musical interpretations of the same moment, each with retrieval from the catalogue underneath.

Refinement carries memory​

Across all three modes, refinements stay in the same session. "Make it more electronic" applied to a precision search updates the constraints; applied to a similarity search shifts the reference profile; applied to a narrative result generates a new direction without retreading the rejected one. The system tracks what the user has rejected so it does not loop.

What the corpus knows about each track — the dimensions all three modes query against — is in What CORPUS Knows About Each Track.