The demo/ application is the browser interface used to inspect offline Duo inference.
It is backend-neutral: model and checkpoint names are supplied at runtime, and
the source contains no service URL, credential, or checkpoint path.
Serve the directory and add ?preview=1 to the URL:
python -m http.server 8000 --directory demoOpen http://127.0.0.1:8000/?preview=1. The preview exercises recording, upload, queue, timeline, settings, and progress layouts without calling a model. Generated media is intentionally unavailable in this mode.
Without ?preview=1, the application connects to API routes one directory
above the frontend. A deployment commonly serves the page at /video/ and the
API at /api/ on the same origin. This preserves camera permissions, cookies,
byte-range video requests, and event streaming without cross-origin setup.
The frontend first requests GET /api/frontend-catalog and falls back to its
local catalog.json. The endpoint returns model labels and Cases:
{
"models": [
{
"id": "duo",
"label": "Duo",
"detail": "Configured checkpoint",
"sampler": "Configured by the backend"
}
],
"cases": [
{
"id": "case",
"label": "Case",
"subtitle": "",
"description": "Assistant and scene description",
"assistant_image": "/api/cases/case/reference-image"
}
]
}The catalog may include a schema object when an existing backend uses
different JSON field names. The frontend normalizes those names to intent,
realization, and codec without embedding model-specific terminology.
Only the fixed assistant reference image and system description belong in the catalog. User media always comes from the current camera recording or uploaded video.
The browser uses these same-origin routes:
| Route | Purpose |
|---|---|
GET /api/frontend-catalog |
Runtime model and Case labels |
GET /api/video/state |
Current queue, session, chunks, and media |
GET /api/video/events |
Server-sent state updates |
POST /api/video/join |
Join the shared session queue |
POST /api/video/accept |
Accept the offered turn |
POST /api/video/leave |
Leave the session or queue |
POST /api/video/capture |
Reserve a recording upload |
POST /api/video/upload |
Upload MP4 or WebM input |
POST /api/video/cancel-capture |
Cancel a reserved upload |
POST /api/reset |
Reset committed history or the full session |
POST /api/model |
Select a backend-defined model ID |
POST /api/case |
Select a Case and fixed conditions |
State chunks use public stage fields named intent, realization, and codec.
Media URLs should support HTTP byte ranges so cumulative video can update
without interrupting active playback.
The API retains a few serialized block field names for compatibility with
existing clients and checkpoints. UI copy and new documentation call each
17-frame interaction interval a chunk.