Imported from NVIDIA-AI-Blueprints/video-search-and-summarization (skills/operations/vss-ask-video/SKILL.md). Install upstream with npx skills add NVIDIA-AI-Blueprints/video-search-and-summarization --skill vss-ask-video. Copyright stays with the author (Apache-2.0).

Ask a VSS video question

Answer from the cheapest grounded source that can satisfy the question. For a running VSS deployment, use the project-local vss CLI. Do not call an OpenAI-compatible /chat/completions endpoint directly or fall back to raw REST when a CLI command fails.

This skill does not call POST /generate on the VSS agent. It requires a deployed VSS with vss configure already run.

Hard rule — never substitute a hand-built HTTP call for the CLI. Specifically, do not:

• POST to /v1/chat/completions yourself. vss vlm run owns that call.
• Build VIOS clip URLs by hand (e.g. /vst/api/v1/storage/file/<id>/url). --sensor resolves the sensor, recorded window and clip URL internally.
• POST to http://<host>:8000/generate or /v1/summarize.

If vss vlm run fails, report the exit code. Do not retry the question by hand-rolling the request.

Prerequisites

A deployed VSS stack with rt_vlm reachable through the configured origin. Run vss configure once per deployment. Bootstrap, exit codes, and common CLI rules live in AGENTS.md.

vss configure check
# Expected: rt_vlm   ok   http://<origin>/rtvi-vlm   HTTP 200

Instructions

Bootstrap the CLI once

VSS_REPO_ROOT="${VSS_REPO_ROOT:-$HOME/video-search-and-summarization}"
vss() { uv run --project "${VSS_REPO_ROOT}/services/agent" --no-dev --extra cli vss "$@"; }
vss --version

Never construct an endpoint or replace a failed CLI call with raw HTTP.

Choose exactly one initial route

Apply these routes in order:

1. Hot conversation context -> answer directly. If current messages or current-turn tool output already contain the answer, answer from that evidence. Do not query memory or run a model.
2. Explicit stored summary/result -> vss memory get, vss summarize get, or vss memory query. For a known job_id, read the stored parent with vss memory get or, for a summarize job, vss summarize get. To find or list stored results by text, sensor, type, status, or time, use query.
3. General video question where past memory may exist -> vss memory introspect. Use this for a substantive question about prior video analysis when hot context does not answer it and the user did not request one exact stored record or fresh visual verification.
4. Exact sensor/time or explicit fresh visual verification -> vss vlm run. Bypass introspection when the user supplies a grounded VIOS sensor plus exact ISO-8601 UTC start/end times, or explicitly asks to watch, inspect, re-check, or freshly verify that exact window. A user-confirmed vss-search-archive handoff with a pre-resolved bounded VIDEO_URL uses this route as Path A; do not rerun search or resolve a different interval.
5. introspect returns no_memory -> conditional vss vlm run. Run the VLM only when a grounded sensor and exact ISO-8601 UTC start/end window are already available from the request, hot context, or trusted tool output. Otherwise explain that no matching memory was found and that an exact recorded sensor/window is needed. Never invent or broaden a window.

Do not call memory query, memory introspect, and vlm run speculatively or in parallel. The only escalation is the specified no_memory fallback.

For a confirmed search-result handoff, use only the caller-supplied VIDEO_URL and visual question. Treat that URL as Path A; do not rerun search or resolve a different interval. Do not consume similarity scores, filenames, object IDs, or other retrieval metadata as visual evidence, and do not rerun search, resolve a sensor, broaden the clip, or choose another interval. The caller owns verdict validation and any fallback after this skill returns.

Run the selected command

For an explicit stored parent:

vss memory get --job-id '<job-id>'
# summarize jobs may also use:
vss summarize get --job-id '<job-id>'

For a known child, add both --record-type event|search_hit|incident and --record-id '<record-id>'. For discovery, apply only relevant filters:

vss memory query --query '<search text>' --sensor-id '<sensor-name>' --limit 20

For bounded introspection, preserve the user's question verbatim:

vss memory introspect --query '<user question>' --sensor '<sensor-name>'

introspect requires grounded, useful scope: --sensor, --job-id, --record-id, or a complete UTC time range. Add only grounded selectors. A time range requires both --start-time and --end-time. --record-type and --group refine scope but do not establish it. The command may perform its own bounded VLM follow-ups; do not duplicate them manually. no_memory returns JSON with status "no_memory" and exit code 5; this is an expected not-found result, not a general command or backend failure. Do not automatically run a VLM afterward unless an exact sensor and exact ISO-8601 UTC start/end window were already grounded before introspection returned.

For one fresh inspection, use vss vlm run — never a hand-built VLM request.

Path A — URL or local file (default when the user or search handoff provides media directly; skip the sensor check):

VSS=(uv run --project "${VSS_REPO_ROOT:-$HOME/video-search-and-summarization}/services/agent" --no-dev --extra cli vss)
# Exit 6 means the answer was produced but could not be written to memory.
check_rc() { [ "$1" -eq 0 ] || [ "$1" -eq 6 ] || { echo "vss vlm run failed (exit $1)" >&2; exit "$1"; }; }

RC=0
RESULT=$("${VSS[@]}" vlm run --prompt "${USER_QUESTION}" --media-url "${VIDEO_URL}") || RC=$?
check_rc "${RC}"

# Local file (inlined as base64):
# RESULT=$("${VSS[@]}" vlm run --prompt "${USER_QUESTION}" --file "${VIDEO_FILE}") || RC=$?

Path B — named VIOS sensor (optional window). List sensors first even when the user names the sensor. Then:

RC=0
RESULT=$("${VSS[@]}" vlm run \
  --prompt "${USER_QUESTION}" \
  --sensor "${SENSOR_NAME}" \
  --start-time "${START_TIME}" \
  --end-time "${END_TIME}") || RC=$?
check_rc "${RC}"

A question that names a sensor is Path B and MUST use --sensor. Do not hand-build a /storage/file/<streamId>/url call. The window must be fully recorded and start before end. Cite the returned job_id, sensor, and window. Do not substitute vss vios clip, direct VLM HTTP, or a local copy for a failed vss vlm run.

Return a grounded answer

State whether the answer came from hot context, stored memory, introspection, or a fresh VLM job when that distinction matters. Preserve uncertainty and cite available handles. Extract .answer from the CLI JSON. On CLI failure, report the diagnostic and a useful recovery step; do not fabricate an answer.

If vss vlm run exits non-zero, stop and report the error:

• 2 — invalid input. Fix the request; do not retry it unchanged.
• 3 — backend unreachable. Retrying is reasonable.
• 4 — required service missing from the recorded config. Re-run vss configure.
• 5 — the sensor name is not in VIOS. List sensors and confirm the name.
• 6 — the answer was produced but could not be written to memory. Keep the answer.
• 7 — timeout (raise --timeout).

Examples

• Hot conversation: The previous turn says, "A forklift crossed the loading aisle at 10:14 UTC." User: "When did the forklift cross?" -> answer 10:14 UTC directly; run no command.
• Explicit stored parent: "Show me the summary from job sum-01JXYZ." -> vss memory get --job-id sum-01JXYZ or vss summarize get --job-id sum-01JXYZ.
• Stored-result discovery: "Find stored search results about forklifts on dock_cam." -> vss memory query --query 'forklifts' --sensor-id dock_cam.
• General memory-aware question: "Was anyone missing PPE on warehouse_safety_0001?" -> vss memory introspect --query ... --sensor warehouse_safety_0001.
• Exact fresh verification: "Freshly verify whether the worker wore a hard hat on dock_cam from 2026-08-13T20:00:00Z to 2026-08-13T20:00:30Z." -> vss vlm run with that exact sensor/window.
• Search handoff: confirmed unverified hit with bounded VIDEO_URL -> Path A --media-url.
• No-memory with scope: Introspection returns no_memory, while trusted context provides dock_cam and 2026-08-13T20:00:00Z through 2026-08-13T20:00:30Z -> run one vss vlm run for exactly that interval.
• No-memory without scope: "Did a forklift enter the loading area last week?" returns no_memory, with no exact sensor/window -> explain no matching memory/window exists and ask for the sensor and exact UTC window; do not run the VLM.

Explicit non-VSS local-file fallback

This separate fallback applies only when the user explicitly asks about a standalone local file/base64 video and no VSS deployment or VIOS sensor is in scope. It may use a caller-provided OpenAI-compatible VLM according to that service's documented media format. Label the result as non-VSS.

For an MP4, send the native MP4 bytes as one video input. Read the file directly, base64-encode the complete byte sequence, and construct exactly:

data:video/mp4;base64,<base64 of the complete MP4 file>

Pass that URI in one OpenAI-compatible video_url content part:

{"type":"video_url","video_url":{"url":"data:video/mp4;base64,<complete MP4 base64>"}}

Do not run ffmpeg, OpenCV, or any frame extractor. Do not convert the video to JPEG/PNG images or send an image_url array: extracted frames are not the requested native video input and can discard motion, timing, and audio. If the caller-provided VLM does not support a native MP4 video_url data URI, report that incompatibility instead of silently changing the media format.

Never enter this fallback after vss memory introspect, use it for a named VIOS sensor or stored VSS result, or combine local/base64 media with introspection. If the request could refer to VSS memory, ask the user to choose the standalone file or the VSS sensor. If a VSS deployment is configured, use Path A vss vlm run --file / --media-url instead.

Negative triggers

• Archive/semantic similarity retrieval ("find videos of ...") -> /vss-search-archive. This skill may inspect only the pre-resolved bounded clip that search hands off after confirmation; it never performs the retrieval itself.
• Long-form summarization -> /vss-summarize-video.
• Structured reports -> /vss-generate-video-report.
• Existing analytics incidents or metrics -> /vss-query-analytics.
• Deployment/profile changes -> /vss-deploy-profile.

Cross-Reference

• /vss-manage-video-io-storage — optional Path B upload semantics.
• /vss-deploy-dense-captioning — optional standalone RT-VLM. Do not re-run vss configure against a standalone RT-VLM URL if you also need VIOS.
• /vss-generate-video-report — timestamped reports; this skill returns an ad-hoc answer.
• /vss-query-analytics — already-computed incidents/metrics.