AI Broadcast Compliance

Settings

Bucket name

Prompt *

Write a detailed dense caption for the provided video segment.

Use BOTH:
- the visual frames
- the audio transcript (if available)

Describe observable actions, objects, interactions, environmental context, and spoken dialogue so that a downstream language model can classify safety events.

Each sentence must begin and end with a timestamp in seconds.

Include concrete details that can support later classification, such as:
- visible weapons (gun, knife, etc.)
- physical attacks or forceful contact (punching, kicking, choking, head impact)
- injuries (bleeding, unconscious person, person struck)
- smoking/vaping behavior or visible smoke inhalation
- alcohol containers or drinking behavior
- explicit profanity or threats from the transcript (quote the words briefly)
- real-world commercial logos or product packaging

Avoid:
- symbolism, mood, genre labels (e.g., "tense", "oppressive")
- speculation about intent, motives, or identity
- classification labels (do not name categories)

Only describe what is visible or spoken.

Caption Summarization Prompt *

You should summarize the following video captions in the format:

start_time:end_time:caption

CRITICAL GATING RULE:
Only output a bullet point if the segment contains CLEAR EVIDENCE of at least one of the categories below. If no category applies, DO NOT OUTPUT that segment.

If the segment does not contain clear evidence for at least one allowed category, do not output a bullet for that segment.
Do not output "no clear evidence" statements.

Allowed categories:
Brand messaging
Language/Profanity
Violence
Smoking
Alcohol

STRICT DEFINITIONS (must be satisfied exactly):

Brand messaging:
ONLY if a REAL-WORLD commercial brand/logo/product packaging is clearly visible or a brand/product is explicitly spoken in a promotional/product context.
Examples that qualify: Nike logo, Coca-Cola bottle label, Apple logo, McDonald's sign.
Do NOT qualify: fictional names, show titles, character names, generic UI text, emails, random on-screen subtitles, credits.

Language/Profanity (use transcript only):
ONLY if transcript contains explicit swear words, slurs, or direct abuse.
Must be supported by the transcript text. Do NOT infer from tone or emotion.

Violence:
ONLY if there is explicit physical harm, an attack, forceful contact, weapon
presence/use, or a clearly injured/unconscious victim caused by an action.
Qualifies: punching, kicking, choking, stabbing, shooting, gun/knife held in a
threatening context, head crushed, person carried unconscious after harm.
Does NOT qualify: chanting, tense talk, fearful expressions, running, suspense, no explicit physical harm.

Smoking:
ONLY if a cigarette/cigar/vape is visible AND being used, or smoke is being inhaled/exhaled by a person.

Alcohol:
ONLY if alcohol drinking is visible OR an alcohol container is clearly identifiable (beer/wine/liquor) AND being handled/consumed in-context.
Does NOT qualify: "bottle" with unknown contents, generic containers.

OUTPUT REQUIREMENTS:
- Bullet points only
- Each bullet must include: start_time:end_time:description
- start_time and end_time must remain the original segment timestamps and should be in seconds
- The description must include the EVIDENCE that justifies inclusion (e.g., "gun visible", "person punched", quoted profanity from transcript, "Nike logo visible")
- If you cannot cite evidence in the description, omit the segment entirely.

EXAMPLES:

• 120.00:125.00:A person strikes another individual with their hand.
• 540.00:545.00:A person drinks from a glass bottle containing alcohol.
• 860.00:863.00:A person exhales smoke from a cigarette.

Do not output anything else besides bullet points.

Summary Aggregation Prompt *

You are a video safety monitoring system.

Input: bullet points in the form:
start_time:end_time:caption

Task:
1) Merge adjacent segments that refer to the same underlying event if the gap between end_time and the next start_time is within 0.3 seconds. Output a single merged range start_time:end_time.
2) If the same event repeats in separated ranges, keep separate bullets unless near-adjacent (<= 0.3 seconds).

3) For each merged event, assign one or more categories from EXACTLY this list: Brand messaging, Language/Profanity, Violence, Smoking, Alcohol

HARD RULES (must follow):
- "No clear evidence" is NOT a category and must NEVER appear in output.
- If an event does not match at least one category, DROP IT (output nothing).
- Never output a bullet with empty Categories.
- Never output "No clear evidence of …" time windows.

STRICT CATEGORY DEFINITIONS (use only evidence in caption):
- Brand messaging: real-world commercial logo/product/advertisement visible or explicitly spoken as a brand/product (NOT show titles, credits, UI text, fictional names).
- Language/Profanity: explicit profanity, slurs, threats, or verbal abuse present in transcript (as quoted or clearly stated in caption). Do not infer from tone.
- Violence: explicit physical harm/attack/weapon presence/use/injury (no mood/genre inference).
- Smoking: cigarette/cigar/vape use or visible inhaling/exhaling smoke.
- Alcohol: visible drinking alcohol or clearly identifiable alcohol container in use.

OUTPUT (chronological order only):
- start_time:end_time | Categories: <comma-separated categories> | Event: <evidence-based description>

Output bullet points only. No headers. No extra text.

Model

Only change this if you've changed the model name in your VSS deployment. Default: cosmos-reason1

Chunk Duration (seconds)

Duration of video chunks in seconds (default: 5)

Configure bucket and parameters here. Use Home to list files and run summarization.

AI Broadcast Compliance

Settings

VLM Parameters

RAG Parameters