SampleKit: A Native macOS Sample Browser with Local ML

· 1 min read

In development.

SampleKit (working title TitledLake) is a native macOS sample browser for music producers. Every ML step runs locally on Apple Silicon, and nothing leaves the machine.

Architecture

SampleKit is a Swift Package Manager project split into targets with clear boundaries:

  • AudioIO: reading and decoding audio files
  • Analysis: feature extraction
  • Embedding: turning audio into vectors
  • IndexStore: persistence
  • Search: similarity queries
  • Metadata is stored with GRDB.swift on SQLite.
  • Embedding vectors live in memory-mapped binary files.
  • Similarity search is a single matrix operation using cblas_sgemm from Accelerate. With normalized vectors, cosine similarity reduces to a dot product, so ranking a whole library is one BLAS call over a mapped buffer, with no per-row loop and no copy into memory.
  • A CLI benchmark harness, slcli, measures indexing and search performance outside the UI.

Problems solved along the way

  • Xcode 26 deployment-target mismatches across the package
  • Build settings overridden at the project level versus the target level
  • Entitlements configuration for a sandboxed app

Status

The package structure, storage layer, and benchmark harness are in place. The audio input path is the current open item: it needs the device.audio-input entitlement before AudioIO is implemented.

More software