data-flow.md

Data Flow

Two independent pipelines share the generic worker pool (processing/pool.rs). The frontend starts a run via a command, then reacts to a stream of events.

Photo pipeline (queue.rs)

User clicks Start
  └─ commands.startProcessing(jobs, settings, profile)
       └─ ProcessingQueue::start()
            ├─ build the prompt once (shared by all workers)
            └─ worker pool (max_concurrent)
                 └─ per job:
                    ├─ emit "job-update" -> processing
                    ├─ AI call (SSE)  → emit "stream-chunk" (reasoning|content|usage)*
                    ├─ parse + dedupe tags
                    ├─ if apply_automatically: write EXIF/IPTC/XMP
                    │  else:                   keep tags, status "donePending"
                    ├─ emit "job-update" -> done | donePending | error
                    └─ emit "progress" + "progress-timer"

When auto-apply is off, results wait as donePending; the user edits title/tags/genres inline and later calls apply_photo_metadata, which performs the actual write.

Video pipeline (video_queue.rs)

User clicks Start
  └─ commands.startVideoProcessing(jobs, settings, profile, videoSettings)
       └─ VideoProcessingQueue::start()
            ├─ check ffmpeg/ffprobe
            ├─ build the JSON prompt once
            └─ worker pool (max_concurrent, clamped 1..4)
                 └─ per job, three phases:
                    ├─ 1 EXTRACT   emit "video-job-update" -> extracting
                    │              video_decoder extracts frames -> emit "video-frame-extracted"
                    ├─ 2 ANALYZE   emit "video-job-update" -> processing
                    │              AI call (SSE) → emit "stream-chunk"*  → parse VideoMetaOutput (JSON)
                    │              (one retry on failure; comma-parse fallback if JSON invalid)
                    └─ 3 WRITE     write title/description/genres/keywords via ffmpeg
                                   (or hold as "donePending" for manual apply)
                    └─ emit "video-job-update" -> done | donePending | error
                       emit "video-progress" + "progress-timer"

Events: photo vs. video

Photo	Video
job-update	video-job-update
progress	video-progress
import-progress	video-frame-extracted
progress-timer (shared)	progress-timer (shared)
stream-chunk (shared)	stream-chunk (shared)

Notes

1. Separate modules (queue.rs / video_queue.rs), each with its own pool config.
2. stream-chunk is shared; the frontend routes it by jobId to the right worker buffer (ThinkingPanel).
3. progress-timer is shared; the frontend routes it by appMode.
4. Video has three phases (extract → analyze → write); photo has two (analyze → write).
5. Video returns structured JSON (title, description, genres, keywords); photo returns a comma-separated tag list.
6. Pause/resume/stop are handled inside the pool via a pause control and a CancellationToken; stopping leaves unprocessed jobs in the queue.