kash-media

See the main kash repo for general instructions.

To run kash with the media kit features enabled, ensure you have uv set up then:

# Pin GIL 3.13. Bare `uv tool install` / `uv python find 3.14` may pick
# freethreaded 3.14t, which is unsupported (cydifflib sdist and OpenCV).
uv tool install kash-media --upgrade --force --python 3.13
kash

GIL CPython 3.14 also works if you pass that interpreter explicitly. --python 3.14 is not enough when uv resolves 3.14t.

Or for dev builds from within this git repo:

# Install all deps and run tests:
make
# Run kash with all media kit features enabled:
uv run kash

Model Configuration

This kit inherits the current model profiles from kash-shell. Speaker identification uses the configurable fast_llm workspace parameter instead of requiring a specific provider. Media transcription uses Deepgram nova-3 with the newest generally available batch diarizer through diarize_model=latest; it does not send the deprecated diarize=true option. See the main model configuration documentation for the current Anthropic defaults and equivalent OpenAI settings.

Transcription consumes generic title, description, and additional_context item metadata. Feature-specific hints use the extensible Item.extra.transcription mapping:

extra:
  transcription:
    key_terms: [Alice Chen, SignalFlow]
    speaker_hints:
      "0": Alice Chen

Nova-3 receives each key term separately. Speaker identification uses all descriptive metadata as untrusted reference context, and explicit speaker hints override model guesses. Unknown fields in extra.transcription are preserved for downstream extensions.

For how to install uv and Python, see installation.md.

For development workflows, see development.md.

For instructions on publishing to PyPI, see publishing.md.

This project was built from simple-modern-uv.