Skip to content

feat: multimodal source units for extract (keyframes / audio alongside transcript) #66

Description

@mroops0111

Problem

Source loaders currently materialize text files that claude reads, and view.ts already notes binary artifacts are deferred to v2. Extraction from video needs a source unit that carries non-text payloads: sampled keyframes (slides, diagrams, whiteboard math) and/or audio, alongside the transcript.

Where it came from

Surfaced by newledge: transcript-only extraction misses visual-heavy educational content (a lecturer pointing at a slide, a derivation drawn on a board). A vision-capable extract pass would recover it.

Proposal

  • Let a source unit bundle text + referenced binary assets (frames / audio).
  • Let the extract agent receive them (the default claude-code agent is vision-capable).

Acceptance

  • An extract run consumes a unit with a transcript + N keyframe images.
  • A resulting node can cite a specific frame as evidence (composes with the media-fragment provenance issue).

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions