Learned search

An optional embedding service adds learned ways to find material in an indexed sample library: audio-to-audio similarity and natural-language retrieval. This page explains the two embedding spaces, how a learned description combines with structured Sample Explorer queries, and how to read the results.

The optional embedding service supplies two spaces:

Learned search, generation, and compatibility ranking are separate tools. PHALAR stem pairing addresses a different question from similarity: compatibility between recordings.

Sample Explorer indexes configured sample roots and cached metadata rather than requiring filenames in every composition. Cmd+P opens it in Studio. Structured queries can combine fields such as BPM, duration, kind, and tags:

bpm:120-140 kind:drum
dur:0.1-2 tag:vocals

An Explorer query can also combine structured constraints with a learned description, written after nl::

bpm:120-140 nl:dark ambient pad

Context-aware insertion and the separate audition action are described with the rest of the sample workflow in Samples and slicing.

Zero-shot descriptions, tags, and similarity rankings come from a model. They are not replacements for BPM or key metadata; learned search adds ways to find material alongside the library’s own metadata.

Without embeddings, indexed and feature-based search remains available. WAV, AIFF, and FLAC headers supply basic metadata, and optional external analysis adds derived features.

These models are optional: patterns, sound output, and many visuals do not need an ML model. The model environment is set up separately from the Lisp dependencies that make deps installs; see Installation and setup.

The MuQ and MuQ-MuLan weights are licensed CC-BY-NC-4.0, as identified on the official MuQ card and MuQ-MuLan card. Mora’s MIT source license does not cover these weights; they keep their own terms.

See Licensing and assets for how Mora’s license relates to dependencies, models, and samples.