Vss E2e Smoke

VSS end-to-end smoke tests

Use this skill for the Video Search & Summarization sample application when a user wants a quick, real deployment check rather than unit tests. The bundled scripts exercise the same public APIs exposed by the running app.

Environment setup (run first)

This skill drives the Video Search & Summarization app through its real source files, so the VSS application must be present and you must run commands from its app root. Do this before anything else, and it works whether or not the VSS source is already in your workspace.

Run the bundled bootstrap. It first tries to find an existing VSS checkout - walking up from the current directory and inspecting the enclosing git repo - and reuses it without ever re-cloning. Only when no checkout is found does it do a shallow, single-branch, sparse checkout of just sample-applications/video-search-and-summarization from main. It prints the resolved app root on stdout:

# SKILL_DIR is THIS skill's own directory (shown to you when the skill loads);
# in-repo it is .github/skills/vss-e2e-smoke. Works the same if the skill is installed standalone.
SKILL_DIR=".github/skills/vss-e2e-smoke"
APP_ROOT="$(bash "$SKILL_DIR/scripts/vss-bootstrap.sh")"
cd "$APP_ROOT"

Every command below assumes the working directory is this APP_ROOT. To pull from a fork/branch or reuse a specific checkout dir, override VSS_REPO_URL, VSS_REPO_BRANCH, or VSS_CLONE_DIR before running it.

Which smoke test to run

• Summary mode (source setup.sh --summary): run scripts/e2e_summary.sh. It uploads a video through Pipeline Manager, starts the summary pipeline, polls state, verifies chunking, and requires a non-empty final summary.
• Search mode (source setup.sh --search): run scripts/e2e_search.sh. It uploads a video through Pipeline Manager, creates search embeddings, queries the search service, and requires non-empty hits.
• Dual mode (source setup.sh --summary --search or --dual): run both scripts. Summary and frame-search are independent capabilities.
• Unified mode (source setup.sh --summary-and-search, --search-and-summary, --all, or --unified): run scripts/e2e_summary.sh first, then scripts/e2e_search.sh. Unified mode uses both summary and search services; if search is configured over summary text, the summary pass confirms the source content exists.

Prerequisites

• Start VSS first with source setup.sh --summary, --search, --dual, or --unified.
• Default external API base is http://localhost:12345/manager through nginx (APP_HOST_PORT=12345). The scripts also accept a host root such as http://localhost:12345 and will detect /manager.
• curl, bash, and python3 must be available.
• Provide a streamable MP4 if the repository sample fallback is not suitable. The checked APP_ROOT/data directory currently contains no sample videos; see references/sample-assets.md.

Commands

From this skill directory:

# Summary smoke; optional first arg is a video path, optional second arg is Pipeline Manager API base
./scripts/e2e_summary.sh [video.mp4] [http://localhost:12345/manager]

# Search smoke; optional first arg is a video path, optional second arg is Pipeline Manager API base
./scripts/e2e_search.sh [video.mp4] [http://localhost:12345/manager]

Useful environment overrides:

VSS_API_BASE=http://localhost:12345/manager   # upload, summary, embeddings, manager search
VSS_SEARCH_BASE=http://localhost:7890         # optional direct video-search service base for /query
VSS_QUERY='person walking near shelves'       # natural-language query for search smoke
VSS_TIMEOUT_SECONDS=1800                      # summary/search polling timeout
VSS_POLL_INTERVAL_SECONDS=5                   # poll interval
VSS_CHUNK_DURATION=10                         # summary sampling.chunkDuration
VSS_SAMPLING_FRAME=5                          # summary sampling.samplingFrame
VSS_FRAME_OVERLAP=0                           # summary sampling.frameOverlap; multiFrame is derived
VSS_EVAM_PIPELINE=object_detection            # summary evam.evamPipeline; API also allows video_ingestion

Interpreting results

• PASS means the script reached the expected final condition using real HTTP calls.
• FAIL means the deployment is not ready, the wrong mode is running, the video was rejected, processing timed out, or a backend component returned an error.
• For summary, check the printed videoId, stateId, chunk counts, statuses, and summary excerpt.
• For search, check the printed videoId, embedding response, query text, and hit count.
• A failing health check usually points at APP_HOST_PORT, nginx, or Pipeline Manager. A failing search query with successful embedding usually points at video-search, VDMS, or embedding/dataprep services.

Grounding

These scripts are grounded in the current app code:

• Pipeline Manager OpenAPI: docs/user-guide/_assets/vss-api.yaml
• nginx API prefix: docs/user-guide/api-reference.md says Pipeline Manager paths are prefixed with /manager/ through nginx
• Upload: POST /manager/videos multipart field video
• Summary: POST /manager/summary, poll GET /manager/states/{stateId}
• Search embeddings: POST /manager/videos/search-embeddings/{videoId}
• Search query: direct search service POST /query with a list of {query_id, query}, or manager shim POST /manager/search/query