Tao Generate Video Reasoning Annotations

references/prompts_warehouse.py

# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

"""Warehouse CCTV domain prompts for the video CoT annotation pipeline.

Reference prompts for warehouse, distribution center, and industrial site
surveillance footage. Covers anomaly subcategories:
  - Safety-Liability: forklift strikes, falls, unsecured loads, spills
  - Operational Oversight: unattended equipment, blocked exits, PPE violations
  - Criminal-Suspicious: unauthorized access, theft, tampering
  - Security Incidents: perimeter breach, tailgating, after-hours activity

USAGE:
    Set ``prompts_module`` in your video_reasoning_annotation YAML config to the dotted import
    path of this file after copying it into your project:
        video_reasoning_annotation:
          prompts_module: "my_package.prompts_warehouse"

    Or copy and modify — this file is a starting point, not a final product.
    Tune the prompts based on your specific facility layout, camera positions,
    and annotation goals.
"""

PROMPT_TEMPLATES = {

    # =========================================================================
    # VIDEO FILTERING
    # =========================================================================
    "video_filtering": (
        "Determine if this video is warehouse or industrial site CCTV "
        "footage.\n\n"
        "A warehouse/industrial CCTV video should have:\n"
        "- A fixed overhead or elevated camera angle viewing an indoor or "
        "semi-outdoor industrial space\n"
        "- Visible industrial infrastructure (racking, shelving, conveyors, "
        "loading docks, pallets, crates)\n"
        "- Workers and/or industrial equipment (forklifts, pallet jacks, "
        "hand trucks, conveyor belts)\n\n"
        "Negative examples (NOT warehouse/industrial CCTV):\n"
        "- Retail store surveillance (shopping aisles, checkout counters)\n"
        "- Office or residential indoor footage\n"
        "- Outdoor traffic or street surveillance\n"
        "- Construction site footage without warehouse infrastructure\n\n"
        "Answer with ONLY \"Yes\" or \"No\"."
    ),

    # =========================================================================
    # ANOMALY CLASSIFICATION
    # =========================================================================
    "video_anomaly_classification": (
        "You are an expert warehouse safety analyst. Watch this video and "
        "classify whether it contains an anomalous event or only normal "
        "warehouse activity.\n\n"
        "An anomaly in warehouse/industrial CCTV is any event involving a "
        "safety incident, operational oversight, equipment malfunction, or "
        "suspicious/criminal activity.\n\n"
        "Examples of anomalies:\n"
        "- Forklift strikes a person, rack, or other equipment\n"
        "- Worker falls from height or trips over an obstruction\n"
        "- Unsecured load falls from a forklift, rack, or conveyor\n"
        "- Worker operates equipment without required PPE (hard hat, vest, "
        "gloves)\n"
        "- Unauthorized person enters a restricted area\n"
        "- Equipment left running unattended in an aisle\n\n"
        "Examples of normal activity:\n"
        "- Workers operating forklifts along designated aisles, moving "
        "pallets between racks\n"
        "- Loading dock operations with trucks being loaded/unloaded "
        "following standard procedure\n"
        "- Workers wearing proper PPE conducting routine inventory checks\n\n"
        "Near-misses count as anomalies even without contact or injury. "
        "PPE violations count even if no incident results. If ambiguous, "
        "default to Normal.\n\n"
        "First, briefly describe what you observe in the video (2-3 "
        "sentences). Then explain your reasoning for the classification "
        "(2-3 sentences). Finally, on the LAST line, write your "
        "classification as exactly one word: \"Anomaly\" or \"Normal\"."
    ),

    # =========================================================================
    # ANOMALY CAPTIONING
    # =========================================================================
    "anomaly_global_caption": (
        "You are an expert warehouse safety analyst. Watch the video "
        "carefully.\n\n"
        "Describe this video in detail, addressing the following points in "
        "a cohesive narrative:\n"
        "- **Facility Layout and Zone**: Describe the area \u2014 aisle "
        "identifiers if visible, rack positions, dock numbers, zone "
        "markings, restricted area signage, floor markings (pedestrian "
        "walkways, forklift lanes, hazard zones).\n"
        "- **Personnel and Equipment**: Count of visible workers, their "
        "PPE status (hard hat, high-visibility vest, gloves, safety "
        "shoes), equipment in use (forklift, pallet jack, conveyor, hand "
        "truck), equipment state (moving/idle/loaded/unloaded).\n"
        "- **The Safety Incident**: Describe the anomalous event in "
        "detail. What exactly happened? Who was involved? What equipment "
        "or materials were involved? What was the sequence of actions "
        "leading up to, during, and immediately after the incident?\n"
        "- **Contributing Factors and Consequences**: What safety protocol "
        "was violated or what went wrong? What was the immediate outcome "
        "(injury, damage, near-miss, spill)? What secondary hazards "
        "were created?\n\n"
        "Provide a holistic summary that captures the complete story and "
        "context. Pay close attention to temporal and spatial details. "
        "Do not hallucinate or make up information."
    ),

    "anomaly_dense_caption": (
        "You are an expert warehouse safety analyst. Provide a dense, "
        "event-level video caption with precise timestamps for a video "
        "containing a warehouse safety incident.\n\n"
        "Structure your response as a list of events. For each event, "
        "provide the start and end timestamp in the format "
        "<HH:MM:SS><HH:MM:SS> followed by a detailed description.\n\n"
        "Focus particularly on:\n"
        "1. **Precise Timing**: Accurately capture the start and end of "
        "each distinct action or state change.\n"
        "2. **Incident Actions**: Describe in detail the actions of "
        "personnel and equipment involved in the incident \u2014 equipment "
        "movements, worker positions, load handling, the moment of "
        "failure or contact.\n"
        "3. **Personnel Identification**: For every key person, note "
        "their role if distinguishable (forklift operator, floor worker, "
        "supervisor), PPE status, position in the facility, and what "
        "they are carrying or operating.\n"
        "4. **Facility Context**: Note aisle dimensions, rack heights, "
        "clearance zones, floor conditions (wet/dry/obstructed), and "
        "signage.\n"
        "5. **Environmental Factors**: Mention lighting level, floor "
        "condition, alarm indicators, and any obstructions to visibility "
        "or movement.\n\n"
        "Format each line as:\n"
        "<Start_Timestamp><End_Timestamp> Description of the event.\n\n"
        "Example:\n"
        "<00:00:00><00:00:03> A wide warehouse aisle between tall pallet "
        "racks; a forklift loaded with a large pallet is traveling toward "
        "the camera from the far end.\n"
        "<00:00:03><00:00:06> A floor worker in a yellow vest and hard "
        "hat enters the aisle from a cross-aisle on the right, walking "
        "toward the approaching forklift.\n"
        "<00:00:06><00:00:08> The forklift's load is raised approximately "
        "2 meters, blocking the operator's forward view; the operator "
        "does not slow down or sound the horn.\n"
        "<00:00:08><00:00:12> The forklift clips the floor worker's leg "
        "as it passes. The worker stumbles and falls to the ground. The "
        "forklift stops a few meters later."
    ),

    "anomaly_chunk_caption": (
        "You are an expert warehouse safety analyst. You are viewing a "
        "strict {chunk_duration}-second window of a video that contains a "
        "safety incident somewhere in the full video.\n"
        "Your job is NOT to tell the whole story, but to capture the "
        "**micro-dynamics** and **state changes** within this specific "
        "timeframe.\n\n"
        "Note: This specific chunk may or may not contain the incident \u2014 "
        "it covers only a fixed time window. Describe only what you observe "
        "in this clip; do not assume or fabricate incident content if it "
        "is not visible.\n\n"
        "Please describe the events in this {chunk_duration}-second window, "
        "focusing on:\n\n"
        "1. **Operational Dynamics**:\n"
        "   - Equipment movement paths, speeds, and load status\n"
        "   - Worker movement patterns, tasks being performed\n"
        "   - Load handling actions (lifting, placing, stacking, "
        "transporting)\n\n"
        "2. **Facility and Safety Context**:\n"
        "   - Floor condition (clear, obstructed, wet)\n"
        "   - PPE compliance for visible workers\n"
        "   - Signage, zone markings, and lighting conditions\n\n"
        "3. **Incident or Interaction Details**:\n"
        "   - If a safety incident occurs in this chunk (collision, fall, "
        "dropped load, equipment failure, PPE violation), describe the "
        "sequence precisely.\n"
        "   - If the aftermath is visible (injured worker, scattered "
        "materials, stopped equipment), describe the post-event state.\n"
        "   - If no incident is visible, describe the normal operations.\n\n"
        "Do not speculate on what happened before or after this clip. "
        "Report only what is visually confirmed in these "
        "{chunk_duration} seconds. Be precise and objective."
    ),

    # =========================================================================
    # NORMAL CAPTIONING
    # =========================================================================
    "normal_global_caption": (
        "You are an expert warehouse operations analyst. Watch the video "
        "carefully.\n\n"
        "Describe this video in detail, addressing the following points in "
        "a cohesive narrative:\n"
        "- **Facility Layout and Zone**: Describe the area \u2014 aisle "
        "identifiers, rack positions, dock numbers, zone markings, floor "
        "markings.\n"
        "- **Personnel and Equipment**: Count of visible workers, PPE "
        "status, equipment in use, equipment state.\n"
        "- **Activities and Operations**: What work is being performed? "
        "Describe the main activities, task sequences, material handling "
        "flows, and interactions between personnel and equipment.\n"
        "- **Overall Context**: Summarize the operational situation \u2014 "
        "what area of the facility, what shift or activity phase this "
        "appears to be, general pace and efficiency.\n\n"
        "Provide a holistic summary that captures the complete picture. "
        "Pay close attention to temporal and spatial details. "
        "Do not hallucinate or make up information."
    ),

    "normal_dense_caption": (
        "You are an expert warehouse operations analyst. Provide a dense, "
        "event-level video caption with precise timestamps for this video "
        "showing normal warehouse activity.\n\n"
        "Structure your response as a list of events. For each event, "
        "provide the start and end timestamp in the format "
        "<HH:MM:SS><HH:MM:SS> followed by a detailed description.\n\n"
        "Focus particularly on:\n"
        "1. **Precise Timing**: Accurately capture the start and end of "
        "each distinct action, movement, or state change.\n"
        "2. **Operational Activities**: Describe tasks being performed \u2014 "
        "picking, packing, loading, stacking, transporting, sorting.\n"
        "3. **Personnel Identification**: For every key worker, note "
        "their role, PPE status, position, and what they are handling.\n"
        "4. **Facility Context**: Note spatial relationships, equipment "
        "positions, and material flow directions.\n\n"
        "Format each line as:\n"
        "<Start_Timestamp><End_Timestamp> Description of the event.\n\n"
        "Example:\n"
        "<00:00:00><00:00:04> A loading dock area with two open bay "
        "doors; a worker in a yellow vest is scanning boxes on a pallet "
        "near bay 3.\n"
        "<00:00:04><00:00:08> A forklift approaches from the warehouse "
        "interior carrying a loaded pallet; the operator sounds the horn "
        "before entering the dock area.\n"
        "<00:00:08><00:00:12> The forklift places the pallet next to the "
        "truck at bay 2; a second worker begins transferring boxes from "
        "the pallet into the truck.\n"
        "<00:00:12><00:00:16> The first worker finishes scanning and "
        "affixes a shipping label to the pallet; the forklift reverses "
        "back toward the warehouse."
    ),

    "normal_chunk_caption": (
        "You are an expert warehouse operations analyst. You are viewing a "
        "strict {chunk_duration}-second window of a video showing normal "
        "warehouse activity with no safety incidents.\n"
        "Your job is NOT to tell the whole story, but to capture the "
        "**micro-dynamics** and **state changes** within this specific "
        "timeframe.\n\n"
        "Please describe the events in this {chunk_duration}-second window, "
        "focusing on:\n\n"
        "1. **Operational Dynamics**:\n"
        "   - Equipment movement paths and load status\n"
        "   - Worker task steps (picking, placing, scanning, stacking)\n"
        "   - Material flow direction and handling\n\n"
        "2. **Facility and Safety Context**:\n"
        "   - Floor condition, PPE compliance, zone markings\n"
        "   - Lighting and visibility conditions\n\n"
        "3. **Coordination and Interactions**:\n"
        "   - How workers coordinate with each other and with equipment\n"
        "   - Task handoffs, communication signals, spatial awareness\n\n"
        "Do not speculate on what happened before or after this clip. "
        "Report only what is visually confirmed in these "
        "{chunk_duration} seconds. Be precise and objective."
    ),

    # =========================================================================
    # HIGHLIGHT CHUNK (ANOMALY ONLY)
    # =========================================================================
    "highlight_timestamp_extraction": (
        "You are an expert warehouse safety analyst. Given the captions "
        "below for a video containing a safety incident, identify the "
        "EXACT timestamp (in seconds) when the primary incident occurs "
        "\u2014 the critical moment of contact, failure, or violation itself, "
        "not the lead-up behavior before it and not the aftermath.\n\n"
        "Use the chunk captions to cross-reference and narrow down the "
        "precise moment, since they provide detailed observations for "
        "specific time windows.\n\n"
        "Respond with ONLY a single number representing the timestamp in "
        "seconds (e.g., \"20\" or \"12.5\"). Nothing else.\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "[Dense Caption]\n{dense_caption}\n\n"
        "[Chunk Captions]\n{chunk_captions_str}"
    ),

    "highlight_chunk_caption": (
        "You are an expert warehouse safety analyst. You are viewing a "
        "critical {duration}-second highlight clip extracted from a longer "
        "video, centered on the moment a safety incident occurs.\n\n"
        "This clip spans from {start_time}s to {end_time}s of the "
        "original video. The incident is estimated to occur at "
        "approximately {anomaly_time}s.\n\n"
        "Provide an extremely detailed description of what happens in "
        "this clip, with special attention to:\n\n"
        "1. **Pre-Event State** (first 1\u20133 seconds):\n"
        "   - Exact positions of workers and equipment\n"
        "   - Equipment speed and load status (loaded/unloaded, load "
        "height)\n"
        "   - PPE status of all visible workers\n"
        "   - Floor condition and visibility in the area\n\n"
        "2. **The Critical Moment** (the incident itself):\n"
        "   - Exactly what happens \u2014 the collision, fall, dropped load, "
        "equipment failure, or safety violation\n"
        "   - Which workers and equipment are involved\n"
        "   - What safety protocol was violated or what went wrong\n\n"
        "3. **Immediate Aftermath** (last 1\u20133 seconds):\n"
        "   - Where workers and equipment end up\n"
        "   - Whether anyone is injured or materials are scattered\n"
        "   - Other workers' reactions (stopping, rushing to help, "
        "alerting others)\n\n"
        "Be extremely precise about identifying workers (PPE, role, "
        "position) and who does what. Do NOT hallucinate or assume "
        "actions you cannot confirm from the video."
    ),

    # =========================================================================
    # DESCRIPTION GENERATION (TEXT-TO-TEXT)
    # =========================================================================
    "anomaly_description": (
        "You are an expert warehouse safety analyst. You are provided with "
        "captions for a video containing a safety incident:\n"
        "1. Global Caption: General description of the facility area and "
        "events.\n"
        "2. Event-level Dense Caption: Detailed event sequence with "
        "timestamps.\n"
        "3. Chunk Captions: Detailed captions for small time windows.\n\n"
        "Task: Generate a detailed, organized, and cohesive description "
        "of the video with exactly these 3 parts:\n"
        "1. Holistic Scene Description \u2014 Describe the warehouse area: "
        "aisle layout, rack configuration, equipment present, lighting, "
        "floor condition, and general operational context.\n"
        "2. Temporal and Spatial Localization of Incident Events \u2014 List "
        "each key event in chronological order using this exact format "
        "(one event per line, timestamp in MM:SS):\n"
        "<start timestamp MM:SS><end timestamp MM:SS> [where in the "
        "facility] what happened\n"
        "3. Description of the Incident \u2014 Covering:\n"
        "   - Category of incident (equipment collision, fall, dropped "
        "load, PPE violation, unauthorized access, equipment failure)\n"
        "   - Detailed description of the event\n"
        "   - Time and duration\n"
        "   - Spatial location within the facility\n"
        "   - Safety protocol analysis: what rule or procedure was "
        "violated, what was the root cause (operator error, equipment "
        "malfunction, environmental hazard, inadequate training)\n"
        "   - Consequences (injury, damage, near-miss, operational "
        "disruption, secondary hazards)\n\n"
        "Instructions for using input data:\n"
        "- Use the [Global Caption] for general scene/event descriptions.\n"
        "- Use the [Dense Caption] to identify timestamps and temporal "
        "sequence of events.\n"
        "- Use [Chunk Captions] for detailed micro-level observations "
        "prior to, during, and after the incident.\n"
        "- If a [Highlight Chunk Caption] is provided, treat it as the "
        "most reliable source for the incident moment. Prioritize it "
        "over chunk captions for Parts 2 and 3. If no highlight is "
        "provided, rely on the other captions.\n"
        "- Resolve conflicts by using information most captions agree on.\n\n"
        "Output formatting rules (strictly enforced):\n"
        "- NEVER reference the source of information. Do not write "
        "phrases like \"the global caption states\", \"captions indicate\", "
        "or any similar attribution. Write as if you are directly "
        "observing the video.\n"
        "- Refer to workers by their visual attributes (PPE, clothing, "
        "role, position), not by ID numbers or labels.\n\n"
        "Input Captions:\n"
        "Video Length: {video_length}s\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "{dense_section}\n\n"
        "[Chunk Captions]\n{chunk_captions_str}\n\n"
        "{highlight_section}"
    ),

    "normal_description": (
        "You are an expert warehouse operations analyst. You are provided "
        "with captions for a video showing normal warehouse activity.\n\n"
        "Task: Generate a detailed, organized, and cohesive description "
        "of the video with exactly these 3 parts:\n"
        "1. Holistic Scene Description \u2014 Describe the facility area: "
        "layout, equipment, lighting, floor condition, and operational "
        "context.\n"
        "2. Temporal and Spatial Localization of Key Events \u2014 List each "
        "notable event in chronological order using this exact format "
        "(one event per line, timestamp in MM:SS):\n"
        "<start timestamp MM:SS><end timestamp MM:SS> [where in the "
        "facility] what happened\n"
        "3. Event Description \u2014 An activity summary. Structure this "
        "section as bullet points, each in \"Key: value\" format. Cover "
        "the operational activities, worker coordination, material "
        "handling patterns, equipment usage, and safety compliance.\n\n"
        "Instructions:\n"
        "- Use the [Global Caption] for general scene and activity "
        "descriptions.\n"
        "- Use the [Dense Caption] for the timestamped sequence of events.\n"
        "- Use [Chunk Captions] for detailed micro-level observations.\n"
        "- Resolve conflicts by using what most sources agree on.\n"
        "- Do not fabricate events not present in the captions.\n"
        "- NEVER reference the source of information. Write as if you "
        "are directly observing the video.\n\n"
        "Input Captions:\n"
        "Video Length: {video_length}s\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "{dense_section}\n\n"
        "[Chunk Captions]\n{chunk_captions_str}"
    ),

    # =========================================================================
    # QA GENERATION — MCQ
    # =========================================================================
    "anomaly_mcq": (
        "You are an expert warehouse safety analyst. You are provided with "
        "captions and a description for a video containing a safety "
        "incident.\n\n"
        "Task:\n"
        "1. Design a multiple-choice question about the safety incident "
        "that requires perception and reasoning to answer.\n"
        "2. The question should focus on one of: the root cause of the "
        "incident, the safety protocol violated, the direct consequence, "
        "the sequence of events, who or what was involved, or the "
        "category of incident.\n"
        "3. The question must be unambiguous: phrased so there is exactly "
        "one correct interpretation and one correct answer based on the "
        "video.\n"
        "4. Each wrong option must be clearly incorrect based on the "
        "video, not merely less likely. All options should be specific "
        "and parallel in structure.\n"
        "5. Provide step-by-step reasoning to derive the answer. The "
        "reasoning should proceed from observations (worker positions, "
        "equipment state, facility conditions) through causal inference "
        "to the conclusion.\n"
        "6. Treat all captions as a representation of the video itself. "
        "Don't mention \"Global Caption\", \"Dense Caption\", or "
        "\"Chunk Captions\" \u2014 only mention the video.\n"
        "7. Ground your Question, Answer, and Reasoning on the provided "
        "data. Don't hallucinate or make up information.\n\n"
        "Here is an example of the desired content style and depth. "
        "(Note: this is for illustration only \u2014 your output must be "
        "based on the actual content of the video provided, not this "
        "example.)\n"
        "<example>\n"
        "Multiple-Choice Question: What safety violation led to the "
        "incident in the warehouse aisle?\n"
        "A. The forklift operator drove with an elevated load obscuring "
        "forward visibility\n"
        "B. A floor worker entered the aisle without checking for active "
        "forklift traffic\n"
        "C. The rack was overloaded beyond its rated capacity, causing "
        "structural failure\n"
        "D. The emergency stop system on the forklift was disabled\n"
        "======\n"
        "Answer: A\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows the forklift traveling down the main aisle with "
        "a large pallet raised to approximately 2 meters, fully blocking "
        "the operator's forward line of sight. A floor worker is walking "
        "in the same aisle ahead of the forklift. The operator does not "
        "slow down or sound the horn because the load obstructs the view. "
        "The forklift clips the worker's leg near the end of the aisle. "
        "The rack structure shows no signs of overloading or failure, and "
        "the floor worker did check for traffic before entering \u2014 the "
        "issue is that the operator was driving with the load elevated "
        "rather than lowered to floor level as required by standard "
        "operating procedure.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Multiple-Choice Question: [Question]\n"
        "A. [Option A]\n"
        "B. [Option B]\n"
        "C. [Option C]\n"
        "D. [Option D]\n"
        "======\n"
        "Answer: [Answer] (should be a single letter choice)\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. "
        "Do not use bullet points or numbered lists.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "[Dense Caption]\n{dense_caption}\n\n"
        "[Chunk Captions]\n{chunk_captions_str}\n\n"
        "[Detailed Video Description]\n{step_2_output}"
    ),

    "normal_mcq": (
        "You are an expert warehouse operations analyst. You are provided "
        "with captions for a video with normal warehouse activity.\n\n"
        "Task:\n"
        "1. Design a multiple-choice question that requires perception "
        "and preferably reasoning to answer.\n"
        "2. The question should focus on: identifying what operation is "
        "being performed, who is involved, the sequence of task steps, "
        "material handling procedures, equipment usage, or safety "
        "compliance.\n"
        "3. The question must be unambiguous with exactly one correct "
        "answer. Each wrong option must be clearly incorrect, not "
        "merely less likely. All options should be specific and parallel.\n"
        "4. Provide step-by-step reasoning to derive the answer.\n"
        "5. Treat all captions as a representation of the video itself. "
        "Don't mention \"Global Caption\" or \"Dense Caption\" \u2014 only "
        "mention the video.\n"
        "6. Ground your output on the provided data. Don't hallucinate.\n\n"
        "Here is an example of the desired content style and depth. "
        "(Note: this is for illustration only.)\n"
        "<example>\n"
        "Multiple-Choice Question: What task does the worker in the "
        "yellow vest perform after the forklift places the pallet at "
        "bay 2?\n"
        "A. Begins transferring boxes from the pallet into the truck\n"
        "B. Signals the forklift operator to bring another pallet\n"
        "C. Scans the pallet barcode and affixes a shipping label\n"
        "D. Moves the pallet to a staging area using a hand truck\n"
        "======\n"
        "Answer: A\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows the forklift placing a loaded pallet at bay 2 "
        "at approximately 00:00:09. The worker in the yellow vest, who "
        "had been standing near the truck, immediately begins picking up "
        "boxes from the pallet and placing them into the truck bed. No "
        "scanner is visible and no barcode scanning occurs. The worker "
        "does not signal the forklift operator, who reverses independently. "
        "No hand truck is used \u2014 the worker carries boxes manually.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Multiple-Choice Question: [Question]\n"
        "A. [Option A]\n"
        "B. [Option B]\n"
        "C. [Option C]\n"
        "D. [Option D]\n"
        "======\n"
        "Answer: [Answer] (should be a single letter choice)\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. "
        "Do not use bullet points or numbered lists.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "[Dense Caption]\n{dense_caption}"
    ),

    # =========================================================================
    # QA GENERATION — BINARY
    # =========================================================================
    "anomaly_bcq": (
        "You are an expert warehouse safety analyst. You are provided with "
        "captions and a description for a video containing a safety "
        "incident.\n\n"
        "Task:\n"
        "1. Design two binary questions, one with answer \"Yes.\" and "
        "one with answer \"No.\". Don't make the questions too "
        "complicated. They should be straightforward but still require "
        "observing and reasoning about the video carefully.\n"
        "2. For each question, provide a step-by-step reasoning of how "
        "to derive the answer based on the video. Start from observation "
        "of the warehouse scene and the safety incident, then "
        "reason/analyze the question.\n"
        "3. Treat the description as a representation of the video "
        "itself. Generate the Question, Answer, and Reasoning as if you "
        "are looking at the video directly. Don't mention the description "
        "in the Question, Answer, or Reasoning. Only mention the video.\n"
        "4. Ground your output on the provided data. Don't hallucinate.\n\n"
        "Here is an example of the desired content style and depth. "
        "(Note: this is for illustration only \u2014 your output must be "
        "based on the actual content of the video provided, not this "
        "example.)\n"
        "<example>\n"
        "Question: Was the forklift operator's load at a safe transport "
        "height when the incident occurred?\n"
        "======\n"
        "Answer: No. The forklift's load was elevated to approximately "
        "2 meters, well above the recommended transport height of "
        "15-20 cm from the floor.\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows the forklift traveling down the aisle with a "
        "large pallet raised high on the forks. The bottom of the pallet "
        "is roughly level with the top of the first rack tier, which is "
        "approximately 2 meters from the floor. Standard forklift "
        "operating procedure requires loads to be carried at the lowest "
        "practical height during transport. The elevated load blocks the "
        "operator's forward visibility and raises the vehicle's center "
        "of gravity, both of which contributed to the incident.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "1. Question: [Question]\n"
        "======\n"
        "Answer: Yes. [Additional explanation in one sentence]\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. "
        "Do not use bullet points or numbered lists.]\n"
        "=====\n"
        "2. Question: [Question]\n"
        "======\n"
        "Answer: No. [Additional explanation in one sentence]\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. "
        "Do not use bullet points or numbered lists.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "[Dense Caption]\n{dense_caption}\n\n"
        "[Chunk Captions]\n{chunk_captions_str}\n\n"
        "[Detailed Video Description]\n{step_2_output}"
    ),

    "normal_bcq": (
        "You are an expert warehouse operations analyst. You are provided "
        "with captions for a video with normal warehouse activity.\n\n"
        "Task:\n"
        "1. Design two binary questions, one with answer \"Yes.\" and "
        "one with answer \"No.\". Don't make the questions too "
        "complicated. They should be straightforward but still require "
        "observing and reasoning about the video carefully. Focus on "
        "operational procedures, equipment usage, worker behavior, "
        "PPE compliance, or material handling.\n"
        "2. For each question, provide a step-by-step reasoning of "
        "how to derive the answer based on the video. Start from "
        "observation and description, then reason/analyze.\n"
        "3. Treat the description as a representation of the video "
        "itself. Don't mention the description. Only mention the video.\n"
        "4. Ground your output on the provided data. Don't hallucinate.\n\n"
        "Here is an example of the desired content style and depth. "
        "(Note: this is for illustration only.)\n"
        "<example>\n"
        "Question: Are all workers in the video wearing the required "
        "PPE for the warehouse floor?\n"
        "======\n"
        "Answer: Yes. Both workers visible in the video are wearing "
        "high-visibility vests and hard hats throughout.\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows two workers in the loading dock area. The first "
        "worker, who is scanning boxes, wears a yellow high-visibility "
        "vest and a white hard hat throughout the clip. The second "
        "worker, who transfers boxes to the truck, wears an orange "
        "high-visibility vest and a white hard hat. Neither worker "
        "removes their PPE at any point during the video. Both meet "
        "the standard warehouse floor PPE requirement of vest and "
        "hard hat.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "1. Question: [Question]\n"
        "======\n"
        "Answer: Yes. [Additional explanation in one sentence]\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. "
        "Do not use bullet points or numbered lists.]\n"
        "=====\n"
        "2. Question: [Question]\n"
        "======\n"
        "Answer: No. [Additional explanation in one sentence]\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. "
        "Do not use bullet points or numbered lists.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "[Dense Caption]\n{dense_caption}"
    ),

    # =========================================================================
    # QA GENERATION — OPEN-ENDED
    # =========================================================================
    "anomaly_open_qa": (
        "You are an expert warehouse safety analyst. You are provided with "
        "captions and a description for a video containing a safety "
        "incident.\n\n"
        "Task:\n"
        "1. Design an open-ended question about the safety incident that "
        "requires perception and reasoning to answer.\n"
        "2. The question should ask about the root cause, the safety "
        "protocol violated, the consequence, the sequence of events, or "
        "what should have been done differently to prevent the incident.\n"
        "3. Provide step-by-step reasoning to derive the answer. The "
        "reasoning should proceed from observations (worker positions, "
        "equipment state, facility conditions) through causal inference "
        "to the conclusion.\n"
        "4. Treat all captions as a representation of the video itself. "
        "Don't mention \"Global Caption\", \"Dense Caption\", or "
        "\"Chunk Captions\" \u2014 only mention the video.\n"
        "5. Ground your output on the provided data. Don't hallucinate.\n\n"
        "Here is an example of the desired content style and depth. "
        "(Note: this is for illustration only \u2014 your output must be "
        "based on the actual content of the video provided, not this "
        "example.)\n"
        "<example>\n"
        "Open-ended Question: What safety procedures should the forklift "
        "operator have followed to prevent this incident?\n"
        "======\n"
        "Answer: The forklift operator should have lowered the load to "
        "the standard transport height of 15-20 cm from the floor before "
        "driving through the aisle. Transporting the load at "
        "approximately 2 meters blocked forward visibility entirely, "
        "preventing the operator from seeing the floor worker ahead. "
        "Additionally, the operator should have sounded the horn before "
        "entering the aisle and at the cross-aisle intersections, as "
        "required by standard operating procedure. With the load "
        "lowered and the horn sounded, the operator would have had a "
        "clear line of sight and the floor worker would have been "
        "alerted to the approaching forklift.\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows the forklift operator traveling through a "
        "warehouse aisle with the load elevated to approximately 2 "
        "meters \u2014 roughly the height of the first rack tier. This is "
        "far above the recommended transport height. The elevated load "
        "completely blocks the operator's view forward along the aisle.\n\n"
        "A floor worker enters the aisle from a cross-aisle and walks "
        "toward the approaching forklift. The worker checked for traffic "
        "before entering, but the forklift was not visible due to the "
        "racking. The operator, unable to see past the elevated load, "
        "does not slow down, brake, or sound the horn.\n\n"
        "The incident was preventable through two standard procedures: "
        "(1) lowering the load to transport height, which would have "
        "restored forward visibility, and (2) sounding the horn at "
        "cross-aisle intersections, which would have alerted the floor "
        "worker. Both are standard forklift operating requirements "
        "designed precisely for this scenario.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Open-ended Question: [Question]\n"
        "======\n"
        "Answer: [Answer] (open-ended, should be a paragraph)\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. "
        "Do not use bullet points or numbered lists.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "[Dense Caption]\n{dense_caption}\n\n"
        "[Chunk Captions]\n{chunk_captions_str}\n\n"
        "[Detailed Video Description]\n{step_2_output}"
    ),

    "normal_open_qa": (
        "You are an expert warehouse operations analyst. You are provided "
        "with captions for a video with normal warehouse activity.\n\n"
        "Task:\n"
        "1. Design an open-ended question that requires perception and "
        "preferably reasoning to answer.\n"
        "2. The question should ask about the operational workflow, how "
        "workers coordinate, what tasks are being performed, equipment "
        "usage patterns, or what the observable activity reveals about "
        "the facility's operations.\n"
        "3. Provide step-by-step reasoning to derive the answer.\n"
        "4. Treat all captions as a representation of the video itself. "
        "Don't mention \"Global Caption\" or \"Dense Caption\" \u2014 only "
        "mention the video.\n"
        "5. Ground your output on the provided data. Don't hallucinate.\n\n"
        "Here is an example of the desired content style and depth. "
        "(Note: this is for illustration only.)\n"
        "<example>\n"
        "Open-ended Question: How do the two workers coordinate the "
        "loading operation in the video?\n"
        "======\n"
        "Answer: The two workers follow a clear division of labor in "
        "the loading operation. The first worker (yellow vest) handles "
        "inventory verification \u2014 scanning each box's barcode and "
        "affixing shipping labels to the pallets. The second worker "
        "(orange vest) handles the physical loading, transferring boxes "
        "from pallets into the truck. They work in parallel without "
        "blocking each other: the first worker processes pallets at "
        "bay 3 while the second loads at bay 2. The forklift operator "
        "delivers new pallets to the staging area between the bays, "
        "sounding the horn as required before entering the dock area. "
        "This spatial separation and role-based division allows "
        "continuous throughput with minimal idle time.\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows a loading dock with two active bays. At bay 3, "
        "the worker in the yellow vest has a handheld scanner and is "
        "processing boxes on a pallet \u2014 scanning and labeling. At bay 2, "
        "the worker in the orange vest picks up boxes from a pallet and "
        "places them into the truck bed. The forklift arrives at 00:00:04 "
        "with a new pallet and deposits it between the bays. Neither "
        "worker stops their task during the forklift delivery. The "
        "spatial layout \u2014 verification at one bay, loading at another, "
        "forklift delivering to the staging area \u2014 creates a pipeline "
        "where all three roles operate concurrently.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Open-ended Question: [Question]\n"
        "======\n"
        "Answer: [Answer] (open-ended, should be a paragraph)\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. "
        "Do not use bullet points or numbered lists.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Global Caption]\n{global_caption}\n\n"
        "[Dense Caption]\n{dense_caption}"
    ),

    "scene_description": (
        "You are an expert facility operations analyst. You are provided "
        "with a holistic and detailed description for a "
        "warehouse/facility surveillance video. Your task is to produce a "
        "**scene description caption** that captures the static, spatial "
        "layout of the scene — not the events that unfold over time.\n\n"
        "Task:\n"
        "1. Write a **question** — a natural-language question whose "
        "answer is the scene description caption. The question should ask "
        "about the physical setting, facility layout, or environmental "
        "context of the video (e.g., \"What does the warehouse interior "
        "look like in this video?\" or \"Describe the layout and "
        "infrastructure visible in the surveillance video.\").\n"
        "2. Write an **answer** — a concise yet thorough scene "
        "description covering:\n"
        "   - The camera perspective (overhead, ceiling-mounted CCTV, "
        "dock-mounted, perimeter-mounted, etc.).\n"
        "   - Lighting (overhead fluorescent / LED, natural light through "
        "bay doors, low-light/after-hours), and approximate time of day "
        "if discernible.\n"
        "   - Facility layout: aisle layout, racking/shelving rows and "
        "direction, dock bays, conveyor lines, staging areas, doorways, "
        "and floor markings (use frame-relative terms such as \"top to "
        "bottom\", \"left to right\").\n"
        "   - Infrastructure: rack tiers, dock doors and bay numbers, "
        "signage and floor labels, fire-safety equipment, security "
        "barriers, charging stations, and mounted surveillance.\n"
        "   - Stationary objects and environment: parked forklifts/pallet "
        "jacks, staged pallets, crates, signage text overlays, and any "
        "visible inventory.\n"
        "   - People/equipment present and their attributes (worker PPE "
        "color and type, equipment make/identifier when discernible).\n"
        "   Do NOT describe events, worker movements, equipment "
        "operation, or incidents — those belong in the event summary.\n"
        "3. Write a **reasoning** — a coherent paragraph explaining how "
        "observations from the video lead to each element of the answer. "
        "Proceed from what is directly visible to the inferences drawn "
        "(e.g., \"The double-deep racking and floor-marked aisles indicate "
        "a high-density storage layout …\"). Do not use bullet points or "
        "numbered lists.\n"
        "4. Treat the description as a representation of the video "
        "itself. Generate the Question, Answer, and Reasoning as if you "
        "are looking at the video directly. Do not mention the "
        "description in the output.\n"
        "5. Ground your output on the video description. Do not "
        "hallucinate or fabricate details.\n\n"
        "Here are examples of the desired answer style and depth. (Note: "
        "these are for illustration only — your output must be based on "
        "the actual content of the video provided, not these examples.)\n"
        "<example>\n"
        "Answer: This is a fixed overhead surveillance recording from a "
        "ceiling-mounted CCTV camera inside a distribution warehouse "
        "during operating hours, lit by overhead LED fixtures. Two "
        "parallel aisles run from the bottom to the top of the frame, "
        "separated by double-deep selective racking populated with brown "
        "corrugated cartons and shrink-wrapped pallets. A cross-aisle "
        "bisects the frame horizontally near the bottom, with yellow "
        "floor tape demarcating pedestrian and equipment paths. A "
        "pallet-jack charging station is mounted against the right wall, "
        "and a fire extinguisher with red mounting bracket is visible at "
        "the cross-aisle pillar. A timestamp overlay reading \"2026-04-22 "
        "14:32\" appears in the top-left corner.\n"
        "</example>\n"
        "<example>\n"
        "Answer: The video captures a wide-angle view of a loading dock "
        "from a camera mounted above bay 4, looking down toward the dock "
        "floor. Three roll-up dock doors are visible (bays 2, 3, and 4), "
        "with bay 3's door open onto a backed-in delivery truck and bays "
        "2 and 4 closed. The dock floor extends from the foreground to "
        "the staging area in the background, where stacks of palletized "
        "cartons rest against the back wall. Yellow safety bollards line "
        "the dock-edge transition, and a hi-vis floor strip marks the "
        "operator-only zone. A digital sign on the back wall reads \"BAY 3 "
        "— INBOUND 14:00\", and overhead lighting comes from suspended LED "
        "panels. Two parked pallet jacks rest in the right corner near a "
        "charging post.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Question: [Question about the scene layout]\n"
        "=====\n"
        "Answer: [Scene description caption]\n"
        "=====\n"
        "Reasoning: [Coherent paragraph deriving each answer element from "
        "video observations]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Detailed Video Description]\n"
        "{step_2_output}"
    ),

    "event_summary": (
        "You are an expert facility operations analyst. You are provided "
        "with a holistic and detailed description for a "
        "warehouse/facility surveillance video. Your task is to produce "
        "an **event summary caption** that describes the key events and "
        "actions in the video — not the static scene layout.\n\n"
        "Task:\n"
        "1. Write a **question** — a natural-language question whose "
        "answer is the event summary caption. The question should ask "
        "about what happens in the video (e.g., \"What are the key events "
        "in this warehouse surveillance video?\" or \"Summarize the "
        "activity and any incidents observed in the video.\").\n"
        "2. Write an **answer** — a concise yet thorough event summary "
        "covering:\n"
        "   - A brief, high-level summary of all key events in the video.\n"
        "   - For normal operations: describe the activity flow, "
        "role-based handoffs, equipment usage, and any notable "
        "interactions between workers and equipment.\n"
        "   - For anomalous activity: briefly describe normal operations, "
        "then describe the incident in detail, including the actors "
        "involved (PPE color, role), equipment used, the sequence of "
        "events leading to the incident, and the physical aftermath "
        "(injury, dropped/damaged load, blocked egress, halted "
        "operations, etc.).\n"
        "   - **Root cause**: For any anomalous video, explicitly state "
        "the root cause of the incident (e.g., load lifted above "
        "transport height, missing PPE, ignored cross-aisle horn, "
        "unauthorized access, blocked dock-edge guard).\n"
        "   - Use frame-relative directions (top, bottom, left, right) "
        "and timestamps (MM:SS) where possible.\n"
        "3. Write a **reasoning** — a coherent paragraph explaining the "
        "chain of observations and inferences that support the answer. "
        "Start from what is visible, trace the sequence of events, and "
        "connect them to the stated root cause and outcome. Do not use "
        "bullet points or numbered lists.\n"
        "4. Treat the description as a representation of the video "
        "itself. Generate the Question, Answer, and Reasoning as if you "
        "are looking at the video directly. Do not mention the "
        "description in the output.\n"
        "5. Ground your output on the video description. Do not "
        "hallucinate or fabricate details.\n\n"
        "Here are examples of the desired answer style and depth. (Note: "
        "these are for illustration only — your output must be based on "
        "the actual content of the video provided, not these examples.)\n"
        "<example>\n"
        "Answer: A counterbalance forklift carrying a pallet of "
        "shrink-wrapped cartons enters the aisle from the bottom of the "
        "frame at 00:00:03 with the load elevated to roughly 2 meters — "
        "well above standard transport height. A floor worker in an "
        "orange hi-vis vest enters the aisle from the right at 00:00:06 "
        "and begins walking up the aisle. The forklift operator, with "
        "forward visibility blocked by the elevated load, does not slow "
        "down or sound the horn at the cross-aisle, and the forklift's "
        "leading-edge fork strikes the worker's leg at 00:00:09, causing "
        "the worker to fall to the floor. The forklift halts immediately "
        "and a second worker rushes in from the top of the frame to "
        "assist. The root cause is the operator transporting a load above "
        "safe travel height (load lifted ~2 m vs. ~15-20 cm standard), "
        "which fully obscured forward visibility and prevented the "
        "operator from seeing the worker entering the aisle.\n"
        "</example>\n"
        "<example>\n"
        "Answer: Two dock workers complete a routine outbound load at bay "
        "3. The yellow-vested worker scans cartons with a handheld "
        "barcode scanner and applies shipping labels at the pallet "
        "staging area in the bottom of the frame between 00:00:01 and "
        "00:00:08. In parallel, the orange-vested worker transfers "
        "cartons from the labeled pallet into the trailer at bay 3 from "
        "00:00:04 onward. A counterbalance forklift enters at 00:00:10 "
        "from the right, sounds its horn at the cross-aisle, and deposits "
        "a fresh pallet in the staging area before reversing out of frame "
        "by 00:00:14. Both workers continue their tasks without "
        "interruption, exemplifying a normal parallel-pipeline loading "
        "operation.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Question: [Question about the events in the video]\n"
        "=====\n"
        "Answer: [Event summary caption]\n"
        "=====\n"
        "Reasoning: [Coherent paragraph tracing the sequence of events "
        "and root cause from video observations]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Detailed Video Description]\n"
        "{step_2_output}"
    ),

    "anomaly_temporal_event_desc": (
        "You are an expert facility security and safety analyst and video "
        "annotator. You are provided with a holistic and detailed "
        "description for a warehouse/facility surveillance video "
        "containing an anomaly.\n\n"
        "Task:\n"
        "1. Identify a single anomalous event (e.g., a forklift strike, a "
        "fall, an unsecured-load drop, a PPE violation, an unauthorized "
        "access, a tampering event) and select [t1, t2] as the tight "
        "temporal window covering just that one event.\n\n"
        "2. Generate a question asking what happened in the video between "
        "[t1] and [t2]. The question must contain ONLY the timestamps and "
        "a neutral phrasing — no workers, equipment, actions, events, or "
        "outcomes.\n\n"
        "3. Provide an Answer as a single coherent paragraph describing "
        "what occurred in that window, including the inferred cause of "
        "the incident woven naturally into the description.\n\n"
        "4. Provide a Reasoning trace in 2-3 coherent paragraphs "
        "explaining your observations in addition to what changed and why "
        "during the window itself. Where applicable, include pre-event "
        "context before t1 (the normal operational baseline that makes "
        "the anomaly identifiable) and post-event context after t2 (the "
        "resulting state, response, or consequence).\n\n"
        "Constraints:\n"
        "- STRICT ADHERENCE TO TIMESTAMPS: Use only timestamps supported "
        "by the provided captions. Do not hallucinate times.\n"
        "- The Question must contain ONLY the timestamps and a neutral "
        "phrasing — no workers, equipment, events, outcomes, or causes.\n"
        "- The Answer must be a single coherent paragraph. Do not use "
        "bullet points or numbered lists.\n"
        "- The Reasoning must be coherent paragraphs. Do not use bullet "
        "points or numbered lists.\n"
        "- Treat the description as a representation of the video itself. "
        "Generate the Question, Answer, and Reasoning as if you are "
        "looking at the video directly. Do not mention the description in "
        "the Question, Answer, or Reasoning.\n\n"
        "Here is an example of the desired output:\n"
        "<example>\n"
        "Question: What happened in the video between 00:00:06 and "
        "00:00:09?\n"
        "======\n"
        "Answer: A counterbalance forklift carrying a pallet at well "
        "above standard transport height advances through the aisle and "
        "strikes a floor worker entering from the cross-aisle, causing "
        "the worker to fall to the floor. The operator's forward view was "
        "obstructed by the elevated load and no horn was sounded at the "
        "cross-aisle, so the worker — who had no auditory or visual "
        "warning of the forklift's approach — was hit at the moment they "
        "stepped into the lane. The root cause is travelling with the "
        "load lifted to roughly first-tier height instead of the standard "
        "15-20 cm, which fully obscured forward visibility.\n"
        "=====\n"
        "Reasoning:\n"
        "Before 00:00:06, the warehouse aisle is operating normally. The "
        "forklift is moving up the aisle from the bottom of the frame "
        "with a pallet of cartons elevated approximately 2 m, and the "
        "floor worker in an orange hi-vis vest is approaching the "
        "cross-aisle from the right but has not yet entered the lane. "
        "This baseline of an obstructed-vision operator and a worker "
        "about to enter an unannounced lane is what makes the impending "
        "impact identifiable as an anomaly the moment those paths "
        "converge.\n\n"
        "At 00:00:06, the worker steps into the aisle directly in front "
        "of the forklift. The operator, unable to see past the elevated "
        "load and not sounding the horn at the cross-aisle, does not slow "
        "or brake. The leading edge of the lower fork makes contact with "
        "the worker's leg and the worker is knocked to the floor.\n\n"
        "After 00:00:09, the forklift comes to an immediate stop, the "
        "operator dismounts, and a second worker enters from the top of "
        "the frame to assist. The aisle, which was an active travel lane "
        "moments earlier, is now an incident scene. The root cause of the "
        "strike is the operator's elevated-load travel — a procedural "
        "failure that made the impact effectively unavoidable once the "
        "worker's path crossed the forklift's path.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Question: [Neutral question with only timestamps]\n"
        "======\n"
        "Answer: [Single paragraph]\n"
        "=====\n"
        "Reasoning: [2-3 coherent paragraphs]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Detailed Video Description]\n"
        "{step_2_output}"
    ),

    "normal_temporal_event_desc": (
        "You are an expert video annotator and warehouse operations "
        "analyst. You are provided with a holistic and detailed "
        "description for a normal warehouse/facility surveillance video.\n\n"
        "Task:\n"
        "1. Identify a single, self-contained operational event (e.g., a "
        "forklift completing a pallet pick-up or drop-off, a worker "
        "scanning and labeling a pallet, a dock worker loading a carton, "
        "a pallet jack handoff, a dock door cycling open or closed) and "
        "select [t1, t2] as the tight temporal window covering just that "
        "one event.\n\n"
        "2. Generate a question asking what happened in the video between "
        "[t1] and [t2]. The question must contain ONLY the timestamps and "
        "a neutral phrasing — no workers, equipment, actions, events, or "
        "outcomes.\n\n"
        "3. Provide an Answer as a single coherent paragraph describing "
        "what occurred in that window and why the events unfolded as they "
        "did (standard operating procedure, role-based responsibilities, "
        "dock-flow rules, equipment-handling norms — woven naturally).\n\n"
        "4. Provide a Reasoning trace in 2-3 coherent paragraphs "
        "explaining your observations in addition to what changed and why "
        "during the window itself. Where applicable, include pre-event "
        "context before t1 (what led to the event or set the conditions "
        "for it) and post-event context after t2 (the resulting state "
        "once the event completed).\n\n"
        "Constraints:\n"
        "- STRICT ADHERENCE TO TIMESTAMPS: Use only timestamps supported "
        "by the provided captions. Do not hallucinate times.\n"
        "- The Question must contain ONLY the timestamps and a neutral "
        "phrasing — no workers, equipment, events, outcomes, or causes.\n"
        "- The Answer must be a single coherent paragraph. Do not use "
        "bullet points or numbered lists.\n"
        "- The Reasoning must be coherent paragraphs. Do not use bullet "
        "points or numbered lists.\n"
        "- Treat the description as a representation of the video itself. "
        "Generate the Question, Answer, and Reasoning as if you are "
        "looking at the video directly. Do not mention the description in "
        "the Question, Answer, or Reasoning.\n\n"
        "Here is an example of the desired output:\n"
        "<example>\n"
        "Question: What happened in the video between 00:00:04 and "
        "00:00:10?\n"
        "======\n"
        "Answer: A counterbalance forklift approaches the staging area "
        "between dock bays 2 and 3, sounds its horn at the cross-aisle, "
        "lowers a pallet of shrink-wrapped cartons onto the floor at the "
        "marked staging position, and reverses out of the frame. The "
        "operator follows standard dock-floor procedure throughout — horn "
        "before entering the bay area, load lowered to deposit height "
        "before approach, and a clean reverse exit without crossing the "
        "active loading lane.\n"
        "=====\n"
        "Reasoning:\n"
        "Before 00:00:04, the staging area between bays 2 and 3 is empty "
        "and the bay 3 worker (orange vest) is loading the trailer at the "
        "dock-edge while the bay 2 worker (yellow vest) is scanning "
        "cartons at a separate pallet. The forklift, carrying a fresh "
        "pallet, is rolling toward the staging area from the right side "
        "of the frame. This baseline of two active loading roles with a "
        "clear staging gap is the precondition for the upcoming pallet "
        "drop.\n\n"
        "At 00:00:04, the forklift sounds its horn at the cross-aisle and "
        "continues into the staging area. As it approaches the deposit "
        "position, the operator lowers the forks until the pallet rests "
        "on the marked floor zone, then begins reversing. The horn signal "
        "alerts both bay workers without interrupting their tasks, and "
        "the pallet is deposited cleanly without contacting either worker "
        "or the dock-edge guard.\n\n"
        "After 00:00:10, the forklift has fully exited the frame to the "
        "right, the new pallet is staged for the bay-3 worker to begin "
        "loading next, and both bay workers continue uninterrupted. The "
        "event is a textbook example of horn-first cross-aisle entry and "
        "lowered-load deposit, and the staging area is now repopulated "
        "for the next loading cycle.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Question: [Neutral question with only timestamps]\n"
        "======\n"
        "Answer: [Single paragraph]\n"
        "=====\n"
        "Reasoning: [2-3 coherent paragraphs]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Detailed Video Description]\n"
        "{step_2_output}"
    ),

    "anomaly_causal_linkage": (
        "You are an expert facility security and safety analyst. You are "
        "provided with a holistic and detailed description for a "
        "warehouse/facility surveillance video containing an anomaly.\n\n"
        "Task:\n"
        "1. Identify two distinct moments:\n"
        "   - [Timestamp A] (Reference Point): The triggering action, "
        "violation, or initial environmental condition.\n"
        "   - [Timestamp B] (Focus Point): The subsequent outcome, final "
        "rest position, or resulting consequence.\n"
        "2. Generate a \"Blind\" Open-ended Question about the warehouse "
        "incident or any significant event in the video: Ask about the "
        "relationship between the two moments using ONLY their timestamps.\n"
        "   - FORMAT example: \"Explain the relationship between the event "
        "at [Timestamp A] and the situation at [Timestamp B].\"\n"
        "   - CONSTRAINT: Do not mention any specific workers, equipment, "
        "actions, or outcomes in the question.\n"
        "   - Please design questions that require perception and "
        "cognition (reasoning) to answer.\n"
        "   - Note: The relationship between the two moments can be "
        "causal (cause/effect), procedural (rule/violation), or "
        "environmental (condition/reaction).\n"
        "3. Provide a Detailed Answer in a single paragraph:\n"
        "   - Begin by explicitly identifying the specific event "
        "occurring at [Timestamp A] and the specific situation at "
        "[Timestamp B].\n"
        "   - Follow with a narrative explanation of the logical, "
        "physical, or procedural connection.\n"
        "4. Provide a Reasoning Trace in 2–3 coherent paragraphs:\n"
        "   - Start from observation and description of the video, "
        "identify the events at A and B (what happens at [Timestamp A], "
        "what happens at [Timestamp B]). Describe the 'Normal' baseline "
        "state of the video before [Timestamp A] and identify "
        "environmental clues (floor markings, signage, PPE, charging "
        "stations, dock-edge guards) that define the rules of the scene.\n"
        "   - Trace the \"Logical Bridge\" and relevant events between A "
        "and B. Explain the mechanics (e.g., load momentum, obscured "
        "sightlines, blocked egress, missed audible alert) that connect "
        "the two moments. Highlight the key events between A and B that "
        "are needed to answer the question.\n"
        "   - Reason through the causal chain, and state the conclusion "
        "that answers the question.\n\n"
        "Constraints:\n"
        "- The question MUST be \"blind\" (refer only to timestamps).\n"
        "- The Answer must be a single, comprehensive narrative paragraph.\n"
        "- The Reasoning must be coherent paragraphs (no bullet points or "
        "lists).\n"
        "- STRICT ADHERENCE TO TIMESTAMPS: Use the exact timestamps from "
        "the provided captions.\n"
        "- Treat the description as a representation of the video itself. "
        "Generate the Question, Answer, and Reasoning as if you are "
        "looking at the video directly. Don't mention the description in "
        "the Question, Answer, or Reasoning. Only mention the video "
        "itself.\n\n"
        "Here is an example of the desired content style and depth:\n"
        "<example>\n"
        "Question: Explain the relationship between the event at 00:00:03 "
        "and the situation at 00:00:09.\n"
        "======\n"
        "Answer: At 00:00:03, a counterbalance forklift enters the aisle "
        "from the bottom of the frame transporting a palletized load "
        "lifted to approximately 2 m — well above the standard transport "
        "height — while at 00:00:09, a floor worker in an orange hi-vis "
        "vest is on the floor of the aisle in front of the now-stopped "
        "forklift. The relationship is a direct causal chain: the "
        "elevated load at 00:00:03 fully obscured the operator's forward "
        "view, so when the worker entered the cross-aisle and stepped "
        "into the forklift's path, the operator had no visual or auditory "
        "cue (no horn was sounded at the cross-aisle), and the leading "
        "fork made contact with the worker, producing the "
        "down-on-the-floor situation observed at 00:00:09.\n"
        "=====\n"
        "Reasoning:\n"
        "The video at 00:00:03 shows a forklift moving up a warehouse "
        "aisle with a pallet held at first-tier height, and the situation "
        "at 00:00:09 shows a worker fallen on the aisle floor with the "
        "forklift halted immediately behind. Looking at the wider "
        "context, the aisle is marked with yellow "
        "pedestrian-and-equipment lines and a \"horn before crossing\" "
        "decal at the cross-aisle pillar, and the cross-aisle is "
        "otherwise quiet. This baseline establishes the operator's "
        "elevated-load entry as a clear procedural deviation in an "
        "environment that explicitly warns against unannounced "
        "cross-aisle approaches.\n\n"
        "The logical bridge between these two moments is the convergence "
        "of an obstructed-vision operator and an entering pedestrian. The "
        "pallet at ~2 m fully blocks the operator's forward view; the "
        "operator does not lower the load, slow down, or sound the horn. "
        "Around 00:00:06, the worker enters the aisle from the right "
        "cross-aisle. The forklift, still travelling at steady speed, "
        "contacts the worker's leg with its leading fork by 00:00:08, "
        "halting only after impact.\n\n"
        "We can conclude that the situation at 00:00:09 is the terminal "
        "physical resolution of the procedural failure initiated at "
        "00:00:03. Tracing the chain — elevated-load entry, no horn, no "
        "slowdown, no avoidance — it is clear that the worker's fall is a "
        "direct, foreseeable consequence of an unannounced approach with "
        "obstructed forward vision. The relationship is a complete causal "
        "sequence in which the decision to travel with a high load made "
        "the impact effectively inevitable once any pedestrian crossed "
        "the lane.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Question: [Question]\n"
        "======\n"
        "Answer: [Answer] (open-ended, should be a paragraph)\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. Do not "
        "use bullet points or numbered lists.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Detailed Video Description]\n"
        "{step_2_output}"
    ),

    "normal_causal_linkage": (
        "You are an expert warehouse operations analyst. You are provided "
        "with a holistic and detailed description for a normal "
        "warehouse/facility surveillance video.\n\n"
        "Task:\n"
        "1. Identify two distinct moments in the video:\n"
        "   - [Timestamp A] (Reference Point): An initial action or "
        "condition (e.g., a worker beginning a scan-and-label step, a "
        "forklift sounding its horn at a cross-aisle, a dock door cycling "
        "open, the start of a pallet pick-up or drop-off).\n"
        "   - [Timestamp B] (Focus Point): The subsequent outcome (e.g., "
        "a labeled pallet ready for loading, the forklift completing its "
        "deposit, a trailer being loaded, a worker resuming throughput).\n"
        "2. Generate a \"Blind\" Open-ended Question about the relationship "
        "between the two moments using ONLY their timestamps:\n"
        "   - FORMAT example: \"Explain the relationship between the event "
        "at [Timestamp A] and the situation at [Timestamp B].\"\n"
        "   - CONSTRAINT: Do not mention any specific workers, equipment, "
        "actions, or outcomes in the question.\n"
        "   - Design questions that require perception and cognition "
        "(reasoning) to answer.\n"
        "   - The relationship can be causal (cause/effect), sequential "
        "(action/completion), or procedural (signal/response).\n"
        "3. Provide a Detailed Answer in a single paragraph:\n"
        "   - Begin by explicitly identifying the specific event at "
        "[Timestamp A] and the specific situation at [Timestamp B].\n"
        "   - Follow with a narrative explanation of the logical, "
        "physical, or procedural connection.\n"
        "4. Provide a Reasoning Trace in 2–3 coherent paragraphs:\n"
        "   - Start from observation and description of the video; "
        "identify the events at A and B (what happens at [Timestamp A], "
        "what happens at [Timestamp B]). Describe the baseline state "
        "before [Timestamp A] and any relevant environmental clues (floor "
        "markings, bay signage, PPE, charging stations).\n"
        "   - Trace the \"Logical Bridge\" and relevant events between A "
        "and B (e.g., scan-then-label sequence, horn-then-cross, "
        "pick-then-deposit). Highlight the key events between A and B "
        "needed to answer the question.\n"
        "   - Reason through the chain and state the conclusion that "
        "answers the question.\n\n"
        "Constraints:\n"
        "- The question MUST be \"blind\" (refer only to timestamps).\n"
        "- The Answer must be a single, comprehensive narrative paragraph.\n"
        "- The Reasoning must be coherent paragraphs (no bullet points or "
        "lists).\n"
        "- STRICT ADHERENCE TO TIMESTAMPS: Use the exact timestamps from "
        "the provided captions.\n"
        "- Treat the description as a representation of the video itself. "
        "Generate the Question, Answer, and Reasoning as if you are "
        "looking at the video directly. Don't mention the description in "
        "the Question, Answer, or Reasoning. Only mention the video "
        "itself.\n\n"
        "Here is an example of the desired content style and depth:\n"
        "<example>\n"
        "Question: Explain the relationship between the event at 00:00:04 "
        "and the situation at 00:00:10.\n"
        "======\n"
        "Answer: At 00:00:04, a counterbalance forklift sounds its horn "
        "at the cross-aisle as it begins approaching the dock staging "
        "area between bays 2 and 3, and at 00:00:10 a fresh pallet of "
        "shrink-wrapped cartons is resting on the marked staging zone "
        "with the forklift reversing out of frame. The relationship is "
        "sequential and procedural: the event at 00:00:04 is the operator "
        "following the horn-before-cross procedure that signals an "
        "inbound pallet to the bay workers, and the situation at 00:00:10 "
        "is the natural completion of that delivery — the pallet has been "
        "deposited at the marked staging position and the staging area is "
        "restocked for the bay-3 worker's next loading cycle without "
        "interrupting either worker.\n"
        "=====\n"
        "Reasoning:\n"
        "First, from observation of the video: the dock has two active "
        "loading bays. Before 00:00:04, the bay 3 worker (orange vest) is "
        "loading the trailer at the dock-edge and the bay 2 worker "
        "(yellow vest) is scanning and labeling cartons at a separate "
        "pallet. The forklift is rolling toward the staging area between "
        "the bays from the right side of the frame. The two timestamps to "
        "focus on are 00:00:04 and 00:00:10. At 00:00:04, the key event "
        "is the forklift's horn at the cross-aisle as the staging "
        "approach begins. At 00:00:10, the key situation is the freshly "
        "deposited pallet in the marked staging zone with the forklift "
        "reversing out of view.\n\n"
        "The logical bridge between these two moments is a controlled "
        "approach-deposit-reverse sequence. After the horn signal, the "
        "forklift continues into the staging area at the standard "
        "dock-floor approach speed and lowers the forks until the pallet "
        "rests on the marked floor zone. The operator then reverses "
        "cleanly without crossing the active bay-3 loading lane. "
        "Throughout, neither bay worker stops their task — the horn "
        "alerted them without requiring them to clear the area.\n\n"
        "We can conclude that the situation at 00:00:10 is the direct "
        "procedural result of the action at 00:00:04. The relationship is "
        "one of signal-and-completion: a properly announced staging-area "
        "entry leads to a clean pallet deposit and a restocked staging "
        "zone, ready for the next loading cycle. There is no anomaly; the "
        "sequence follows standard horn-before-cross and lowered-load "
        "deposit procedure.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Question: [Question]\n"
        "======\n"
        "Answer: [Answer] (open-ended, should be a paragraph)\n"
        "=====\n"
        "Reasoning: [Provide the analysis as coherent paragraphs. Do not "
        "use bullet points or numbered lists.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Detailed Video Description]\n"
        "{step_2_output}"
    ),

    "anomaly_temporal_localization": (
        "You are an expert video annotator and logical analyst "
        "specializing in temporal localization. You are provided with a "
        "holistic and detailed description for a warehouse/facility "
        "surveillance video containing an anomaly.\n\n"
        "Task:\n"
        "1. Identify a significant event or anomaly that requires logical "
        "reasoning to localize.\n"
        "2. Generate a Temporal Localization Question in the format: "
        "\"When does [Event X] occur in the video?\"\n"
        "3. Provide the Answer with exact start and end timestamps.\n"
        "4. Provide a detailed Reasoning paragraph following a "
        "three-phase logical trace:\n"
        "   - Describe the state of the video immediately before the "
        "start timestamp. Identify the specific procedures, PPE rules, or "
        "operational norms that were being followed to establish why the "
        "incident had NOT yet begun.\n"
        "   - Pinpoint the exact visual 'state-change' at the start "
        "timestamp. Explain the root cause (e.g., elevated-load travel, "
        "missed horn, blocked egress, ignored PPE) that initiated the "
        "incident.\n"
        "   - Justify the end timestamp by describing the transition into "
        "a new, stable, or static state (e.g., forklift coming to a halt, "
        "worker on the floor receiving aid, dropped load fully settled, "
        "dock-edge guard re-engaged, intruder leaving the frame).\n\n"
        "Constraints:\n"
        "- The reasoning must follow the 'Sandwich Structure': Baseline "
        "-> Trigger -> Resolution.\n"
        "- STRICT ADHERENCE TO TIMESTAMPS: Do not hallucinate times. "
        "Cross-reference all provided captions.\n"
        "- The Reasoning section must be a coherent, dense paragraph. Do "
        "not use bullet points or numbered lists.\n"
        "- Treat the description as a representation of the video itself. "
        "Do not mention the description; describe the video directly.\n\n"
        "Here is an example of the desired content style and depth. You "
        "don't need to follow this example exactly, but you should follow "
        "the same format and the style:\n"
        "<example>\n"
        "Question: When does the forklift strike against the floor worker "
        "occur?\n"
        "======\n"
        "Answer: Start_Time: 00:00:06, End_Time: 00:00:09\n"
        "=====\n"
        "Reasoning:\n"
        "The video begins with a stable baseline of normal aisle "
        "activity. Between 00:00:00 and 00:00:06, a counterbalance "
        "forklift travels up the aisle from the bottom of the frame "
        "carrying a pallet at approximately 2 m height — already a "
        "procedural deviation from the 15-20 cm transport standard, but "
        "with no other actors in the lane. The cross-aisle on the right "
        "is clear, and the floor markings and \"horn before crossing\" "
        "decal at the cross-aisle pillar establish the expected "
        "operational rules. The event is triggered at 00:00:06 when a "
        "floor worker in an orange hi-vis vest enters the aisle from the "
        "right cross-aisle and steps into the forklift's path. The "
        "root-cause failure is the elevated load fully blocking the "
        "operator's forward view combined with no horn at the "
        "cross-aisle, which left both the operator and the worker without "
        "any cue to the impending convergence. The leading edge of the "
        "lower fork strikes the worker's leg at 00:00:08, and the worker "
        "is knocked to the floor. The event reaches a resolution at "
        "00:00:09, by which time the forklift has come to a complete halt "
        "with the worker on the aisle floor in front of it and a second "
        "worker beginning to enter the frame from the top to assist, "
        "transitioning the scene from an active operation to a stationary "
        "incident.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Question: [Question]\n"
        "======\n"
        "Answer: Start_Time: [MM:SS], End_Time: [MM:SS]\n"
        "=====\n"
        "Reasoning: [Provide the analysis as a coherent paragraph. "
        "Describe the 'before-state' that led to the start timestamp, the "
        "visual triggers during the event, and the 'after-state' or "
        "resolution that confirms the event ended at the specific "
        "timestamp.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Detailed Video Description]\n"
        "{step_2_output}"
    ),

    "normal_temporal_localization": (
        "You are an expert video annotator and logical analyst "
        "specializing in temporal localization. You are provided with a "
        "holistic and detailed description for a normal "
        "warehouse/facility surveillance video.\n\n"
        "Task:\n"
        "1. Identify a significant operational event that requires "
        "logical reasoning to localize. Focus on routine warehouse events "
        "such as:\n"
        "   - A forklift completing a pallet pick-up or drop-off\n"
        "   - A worker scanning and labeling a carton or pallet\n"
        "   - A dock worker completing a loading or unloading cycle for a "
        "bay\n"
        "   - A pallet jack or hand truck handoff between workers\n"
        "   - A dock door cycling open or closed\n"
        "   - Activity flow pattern changes (e.g., shift handoff, queue "
        "forming or clearing at a bay)\n\n"
        "2. Generate a Temporal Localization Question in the format: "
        "\"When does [Event X] occur in the video?\"\n\n"
        "3. Provide the Answer with exact start and end timestamps.\n\n"
        "4. Provide a detailed Reasoning paragraph following a "
        "three-phase logical trace:\n"
        "   - Describe the state of the video immediately before the "
        "start timestamp. What was the operational state, worker "
        "position, or equipment configuration that preceded this event?\n"
        "   - Pinpoint the exact visual 'state-change' at the start "
        "timestamp. What action or movement initiated the event?\n"
        "   - Justify the end timestamp by describing the transition into "
        "a new, stable state (e.g., pallet at rest in the staging zone, "
        "scanner returning to its dock, dock door fully open, worker "
        "resuming the next cycle).\n\n"
        "Constraints:\n"
        "- The reasoning must follow the 'Sandwich Structure': "
        "Before-State -> Action/Transition -> After-State.\n"
        "- STRICT ADHERENCE TO TIMESTAMPS: Do not hallucinate times. "
        "Cross-reference all provided captions.\n"
        "- The Reasoning section must be a coherent, dense paragraph. Do "
        "not use bullet points or numbered lists.\n"
        "- Treat the description as a representation of the video itself. "
        "Do not mention the description; describe the video directly.\n\n"
        "Here is an example of the desired content style and depth:\n"
        "<example>\n"
        "Question: When does the forklift complete its pallet drop-off at "
        "the dock staging area?\n"
        "======\n"
        "Answer: Start_Time: 00:00:04, End_Time: 00:00:10\n"
        "=====\n"
        "Reasoning:\n"
        "Prior to 00:00:04, the counterbalance forklift is rolling toward "
        "the dock staging area between bays 2 and 3 from the right of the "
        "frame, carrying a pallet of shrink-wrapped cartons at standard "
        "transport height. The bay 2 worker (yellow vest) is scanning "
        "cartons at a separate pallet, the bay 3 worker (orange vest) is "
        "loading the trailer at the dock-edge, and the staging zone is "
        "empty — the precondition for an incoming deposit. At 00:00:04, "
        "the forklift sounds its horn at the cross-aisle, signalling the "
        "dock workers and initiating the staged-approach phase of the "
        "deposit. As the forklift continues into the staging area, the "
        "operator gradually lowers the forks; by 00:00:08, the pallet has "
        "touched down on the marked staging zone, and the operator begins "
        "reversing cleanly without crossing the active bay-3 loading "
        "lane. By 00:00:10, the forklift has fully exited the frame to "
        "the right and the pallet is at rest in the staging zone, "
        "restocked and ready for the bay-3 worker's next loading cycle, "
        "marking the conclusion of the drop-off maneuver.\n"
        "</example>\n\n"
        "Format:\n"
        "=====\n"
        "Question: [Question]\n"
        "======\n"
        "Answer: Start_Time: [MM:SS], End_Time: [MM:SS]\n"
        "=====\n"
        "Reasoning: [Provide the analysis as a coherent paragraph. "
        "Describe the 'before-state' that led to the start timestamp, the "
        "action/transition during the event, and the 'after-state' that "
        "confirms the event ended at the specific timestamp.]\n\n"
        "Input Data:\n"
        "Video Length: {video_length}s\n\n"
        "[Detailed Video Description]\n"
        "{step_2_output}"
    ),
}


def get_prompt(key, **kwargs):
    """Retrieve and optionally format a prompt template by key."""
    template = PROMPT_TEMPLATES.get(key)
    if template is None:
        raise ValueError(f"No prompt template found for key: {key}")
    if kwargs:
        return template.format(**kwargs)
    return template