Tao Generate Video Reasoning Annotations

references/prompts_traffic.py

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

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

Reference prompts for traffic intersection and highway surveillance footage.
Covers anomaly types: collisions, near-misses, stalled vehicles, and traffic
rule violations (red-light running, illegal turns, wrong-way driving).

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_traffic"

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

PROMPT_TEMPLATES = {

    # =========================================================================
    # VIDEO FILTERING
    # =========================================================================
    "video_filtering": (
        "Determine if this video is traffic CCTV footage.\n\n"
        "A traffic CCTV video should have:\n"
        "- A fixed overhead or elevated camera angle viewing a road, "
        "intersection, or highway\n"
        "- Visible vehicles (cars, trucks, motorcycles, buses) using the "
        "roadway\n"
        "- Road infrastructure (lane markings, traffic signals, signs, "
        "crosswalks, medians)\n\n"
        "Negative examples (NOT traffic CCTV):\n"
        "- Dashcam or in-vehicle POV footage\n"
        "- Indoor surveillance (warehouses, offices, retail stores)\n"
        "- Footage of parking lots with no active traffic flow\n"
        "- Drone footage with rapidly changing altitude or angle\n\n"
        "Answer with ONLY \"Yes\" or \"No\"."
    ),

    # =========================================================================
    # ANOMALY CLASSIFICATION
    # =========================================================================
    "video_anomaly_classification": (
        "You are an expert traffic analyst. Watch this video and classify "
        "whether it contains an anomalous event or only normal traffic "
        "activity.\n\n"
        "An anomaly in traffic CCTV is any event involving a collision, "
        "near-miss, stalled or disabled vehicle, or traffic rule violation "
        "(running red lights, illegal turns, wrong-way driving, failure to "
        "yield).\n\n"
        "Examples of anomalies:\n"
        "- Vehicle runs a red light and collides with cross-traffic\n"
        "- Vehicle stalls or breaks down in a travel lane, creating a hazard\n"
        "- Vehicle makes an illegal U-turn or drives the wrong way\n"
        "- Near-miss where vehicles swerve or brake hard to avoid collision\n\n"
        "Examples of normal activity:\n"
        "- Vehicles waiting at a red light, then proceeding on green\n"
        "- Steady traffic flow with lane discipline and safe following "
        "distances\n"
        "- Pedestrians crossing at marked crosswalks during walk signals\n\n"
        "Near-misses count as anomalies even without contact. Minor "
        "violations count even without consequences. 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 traffic analyst. Watch the video carefully.\n\n"
        "Describe this video in detail, addressing the following points in a "
        "cohesive narrative:\n"
        "- **Traffic Signal State and Timing**: What is the current signal "
        "phase (red/green/yellow/flashing)? Which directions have "
        "right-of-way? Do signal transitions occur during the video?\n"
        "- **Vehicle Movements and Positions**: For each distinguishable "
        "vehicle, describe: color, type (sedan/SUV/truck/motorcycle/bus), "
        "direction of travel, lane position, and speed "
        "(stopped/slow/moderate/fast). Note turns, lane changes, and "
        "approach to the intersection.\n"
        "- **The Collision / Incident**: Describe the anomalous event in "
        "detail. What exactly happened? Identify all vehicles involved and "
        "describe their specific actions leading up to, during, and "
        "immediately after the event. Where was the point of impact or "
        "violation?\n"
        "- **Cause and Right-of-Way Analysis**: Which vehicle had "
        "right-of-way? What traffic rule was violated? What was the "
        "apparent root cause? What were the immediate consequences "
        "(damage, disruption, other vehicles' reactions)?\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 traffic analyst. Provide a dense, event-level "
        "video caption with precise timestamps for a video containing a "
        "traffic anomaly.\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. **Anomalous Actions**: Describe in detail the actions of "
        "vehicles involved in the anomaly \u2014 approach speeds, signal "
        "compliance, evasive maneuvers, point of impact, post-collision "
        "movements.\n"
        "3. **Vehicle Identification**: For every key vehicle, specify "
        "color, type (sedan/SUV/truck/motorcycle/bus), lane, and direction "
        "of travel. Track vehicles through the scene.\n"
        "4. **Intersection Layout**: Note lane configuration, signal "
        "positions, turn lanes, crosswalks, medians, and which approaches "
        "are visible.\n"
        "5. **Environmental Factors**: Mention lighting (day/night/dusk), "
        "weather (clear/rain/fog), road surface condition, and visibility "
        "when relevant.\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 four-way signalized intersection viewed "
        "from an overhead camera; eastbound traffic has a green signal and "
        "vehicles are flowing through normally.\n"
        "<00:00:03><00:00:05> A dark sedan approaches from the south at "
        "moderate speed in the left lane; the southbound signal is red.\n"
        "<00:00:05><00:00:07> The dark sedan enters the intersection "
        "against the red signal; a white SUV traveling eastbound on green "
        "is in the collision path.\n"
        "<00:00:07><00:00:12> The sedan strikes the SUV on its passenger "
        "side at the center of the intersection. Both vehicles rotate and "
        "come to rest. Debris scatters across the eastbound lanes. "
        "Surrounding vehicles brake and stop."
    ),

    "anomaly_chunk_caption": (
        "You are an expert traffic analyst. You are viewing a strict "
        "{chunk_duration}-second window of a video that contains a traffic "
        "anomaly 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 anomaly \u2014 "
        "it covers only a fixed time window. Describe only what you observe "
        "in this clip; do not assume or fabricate anomalous content if it "
        "is not visible.\n\n"
        "Please describe the events in this {chunk_duration}-second window, "
        "focusing on:\n\n"
        "1. **Vehicle Dynamics**:\n"
        "   - Acceleration, deceleration, braking, lane changes, turns\n"
        "   - Approach to intersection, signal compliance, yielding\n"
        "   - For each key vehicle: color, type, lane, direction, speed\n\n"
        "2. **Traffic Signal and Infrastructure**:\n"
        "   - Current signal phase and any transitions\n"
        "   - Lane markings, turn arrows, crosswalk signals\n"
        "   - Visibility conditions (lighting, weather, obstructions)\n\n"
        "3. **Anomaly or Interaction Details**:\n"
        "   - If a collision, near-miss, or violation occurs in this chunk, "
        "describe the sequence precisely: approach, contact/violation, "
        "immediate aftermath.\n"
        "   - If the aftermath is visible (stopped vehicles, debris, "
        "emergency response), describe the post-event state.\n"
        "   - If no anomaly is visible, describe the normal traffic flow.\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 traffic analyst. Watch the video carefully.\n\n"
        "Describe this video in detail, addressing the following points in a "
        "cohesive narrative:\n"
        "- **Traffic Signal State and Timing**: What is the current signal "
        "phase? Which directions have right-of-way? Do signal transitions "
        "occur during the video?\n"
        "- **Vehicle Movements and Positions**: For each distinguishable "
        "vehicle, describe: color, type, direction of travel, lane position, "
        "and speed. Note turns, lane changes, and signal compliance.\n"
        "- **Activities and Traffic Flow**: Describe the overall traffic "
        "pattern \u2014 flow direction, density (light/moderate/heavy), "
        "pedestrian activity, and any notable behaviors.\n"
        "- **Overall Context**: Summarize the intersection or road segment, "
        "time of day, weather, and general traffic conditions.\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 traffic analyst. Provide a dense, event-level "
        "video caption with precise timestamps for this video showing normal "
        "traffic 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. **Traffic Behaviors**: Describe vehicle movements \u2014 lane "
        "changes, turns, stops at signals, yielding, merging, pedestrian "
        "crossings.\n"
        "3. **Vehicle Identification**: For every key vehicle, specify "
        "color, type, lane, and direction. Track vehicles through the "
        "scene.\n"
        "4. **Intersection Layout**: Note lane configuration, signal "
        "positions, and spatial relationships.\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 T-intersection viewed from an elevated "
        "angle; the northbound signal is red and three vehicles are "
        "waiting in the left-turn lane.\n"
        "<00:00:04><00:00:08> The signal turns green; the first vehicle "
        "(white sedan) begins a left turn while yielding to oncoming "
        "traffic.\n"
        "<00:00:08><00:00:12> The white sedan completes the turn; the "
        "second vehicle (red pickup) follows, also turning left. A "
        "pedestrian waits at the crosswalk.\n"
        "<00:00:12><00:00:16> The pedestrian walk signal activates; "
        "the pedestrian crosses while remaining vehicles wait."
    ),

    "normal_chunk_caption": (
        "You are an expert traffic analyst. You are viewing a strict "
        "{chunk_duration}-second window of a video showing normal traffic "
        "activity with no anomalous events.\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. **Vehicle Dynamics**:\n"
        "   - Movements: acceleration, braking, turns, lane changes, "
        "merging\n"
        "   - Signal compliance: stopping at red, proceeding on green, "
        "yielding\n"
        "   - For each key vehicle: color, type, lane, direction\n\n"
        "2. **Traffic Signal and Infrastructure**:\n"
        "   - Current signal phase and any transitions\n"
        "   - Pedestrian signals and crosswalk activity\n"
        "   - Visibility conditions\n\n"
        "3. **Traffic Flow Patterns**:\n"
        "   - Density and speed of traffic\n"
        "   - Queuing behavior at signals\n"
        "   - Pedestrian and cyclist interactions with vehicles\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 traffic analyst. Given the captions below for a "
        "video containing a traffic anomaly, identify the EXACT timestamp "
        "(in seconds) when the primary anomaly occurs \u2014 the critical moment "
        "of impact, violation, or dangerous event 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 traffic analyst. You are viewing a critical "
        "{duration}-second highlight clip extracted from a longer video, "
        "centered on the moment a traffic anomaly occurs.\n\n"
        "This clip spans from {start_time}s to {end_time}s of the "
        "original video. The anomaly 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, lanes, and speeds of all key vehicles\n"
        "   - Traffic signal state and right-of-way\n"
        "   - Any last-second actions (braking, accelerating, swerving)\n\n"
        "2. **The Critical Moment** (the collision, violation, or "
        "near-miss):\n"
        "   - Exactly what happens \u2014 point of impact, type of violation, "
        "evasive maneuver\n"
        "   - Which vehicles are involved, their colors, types, and "
        "approach directions\n"
        "   - Which vehicle had right-of-way and which violated the rule\n\n"
        "3. **Immediate Aftermath** (last 1\u20133 seconds):\n"
        "   - Where vehicles end up (spun, stopped, continued driving)\n"
        "   - Debris, damage visible\n"
        "   - Other vehicles' reactions (braking, stopping, changing lanes)\n\n"
        "Be extremely precise about identifying vehicles (color, type, "
        "direction) 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 traffic analyst. You are provided with captions "
        "for a video containing a traffic anomaly:\n"
        "1. Global Caption: General description of the intersection 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 intersection or road "
        "segment: layout (number of lanes, turn lanes, medians), signal "
        "configuration, lighting, weather, road surface, and general "
        "traffic conditions.\n"
        "2. Temporal and Spatial Localization of Anomaly 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 "
        "intersection] what happened\n"
        "3. Description of the Anomaly \u2014 Covering:\n"
        "   - Category of anomaly (collision, near-miss, stalled vehicle, "
        "rule violation)\n"
        "   - Detailed description of the event\n"
        "   - Time and duration\n"
        "   - Spatial location within the intersection\n"
        "   - Right-of-way analysis: which vehicle had the right-of-way, "
        "which signal phase was active, what rule was violated\n"
        "   - Root cause (red-light running, failure to yield, distraction, "
        "mechanical failure, etc.)\n"
        "   - Consequences (damage, traffic 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 anomaly.\n"
        "- If a [Highlight Chunk Caption] is provided, treat it as the "
        "most reliable source for the anomaly 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 vehicles by their visual attributes (color, type, "
        "direction), 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 traffic analyst. You are provided with captions "
        "for a video showing normal traffic 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 intersection or road "
        "segment: layout, signal configuration, lighting, weather, and "
        "general traffic conditions.\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 "
        "intersection] what happened\n"
        "3. Event Description \u2014 An activity summary. Structure this "
        "section as bullet points, each in \"Key: value\" format. Cover "
        "the traffic flow patterns, signal compliance, pedestrian "
        "activity, notable vehicle behaviors, and overall conditions.\n\n"
        "Instructions:\n"
        "- Use the [Global Caption] for general scene and traffic "
        "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 traffic analyst. You are provided with captions "
        "and a description for a video containing a traffic anomaly.\n\n"
        "Task:\n"
        "1. Design a multiple-choice question about the traffic anomaly "
        "that requires perception and reasoning to answer.\n"
        "2. The question should focus on one of: the root cause of the "
        "anomaly, right-of-way violation, direct consequence, sequence of "
        "events, which vehicle was at fault, or the type of violation.\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 (signal state, vehicle "
        "positions, movements) 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 was the primary cause of the "
        "collision at the intersection?\n"
        "A. The dark sedan ran a red light while cross-traffic had a "
        "green signal\n"
        "B. The white SUV made an illegal left turn across oncoming "
        "traffic\n"
        "C. A pedestrian stepped into the roadway causing the sedan to "
        "swerve\n"
        "D. The traffic signal malfunctioned, showing green in both "
        "directions\n"
        "======\n"
        "Answer: A\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows the dark sedan approaching the intersection at "
        "speed from the east while the signal for eastbound traffic is "
        "clearly red. Meanwhile, the white SUV is proceeding through the "
        "intersection on a green light traveling northbound. The sedan "
        "enters the intersection without stopping, striking the SUV on "
        "its passenger side. The traffic signals are functioning normally "
        "\u2014 the eastbound signal is red while northbound has green. There "
        "is no pedestrian involvement and no signal malfunction visible. "
        "The root cause is the sedan's failure to stop at the red light.\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 traffic analyst. You are provided with captions "
        "for a video with normal traffic 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: traffic flow patterns, signal "
        "compliance, vehicle behaviors, pedestrian interactions, spatial "
        "relationships, or environmental conditions.\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 happens when the traffic signal "
        "turns green for northbound traffic?\n"
        "A. The white sedan in the left-turn lane yields to oncoming "
        "traffic before completing its turn\n"
        "B. All northbound vehicles proceed straight through the "
        "intersection simultaneously\n"
        "C. A pedestrian begins crossing against the signal, causing "
        "vehicles to stop\n"
        "D. The northbound vehicles remain stopped due to a vehicle "
        "blocking the intersection\n"
        "======\n"
        "Answer: A\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows the northbound signal transition from red to "
        "green at approximately 00:00:06. In the left-turn lane, a white "
        "sedan activates its turn signal and begins to enter the "
        "intersection but pauses to yield to oncoming southbound traffic. "
        "Once the oncoming lane clears, the sedan completes its left "
        "turn at 00:00:12. The straight-through vehicles proceed, but "
        "not simultaneously \u2014 they stagger naturally. No pedestrian "
        "crosses against the signal, and no vehicle blocks the "
        "intersection.\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 traffic analyst. You are provided with captions "
        "and a description for a video containing a traffic anomaly.\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 traffic scene and the anomalous event, 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: Did the dark sedan have a green light when it entered "
        "the intersection?\n"
        "======\n"
        "Answer: No. The dark sedan entered the intersection against a "
        "red signal while cross-traffic had the green light.\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows a signalized intersection. At the time the dark "
        "sedan enters the intersection (approximately 00:00:06), the "
        "signal for its approach direction is clearly red. Meanwhile, "
        "cross-traffic vehicles are proceeding on green. The sedan does "
        "not stop or slow down for the red signal, entering the "
        "intersection illegally and causing the collision.\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 traffic analyst. You are provided with captions "
        "for a video with normal traffic 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 "
        "signal compliance, vehicle behaviors, pedestrian activity, "
        "or traffic flow patterns.\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: Do all vehicles in the video obey the traffic signals?\n"
        "======\n"
        "Answer: Yes. Every vehicle in the video stops at the red light "
        "and proceeds only after the signal turns green.\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows a standard intersection over one full signal "
        "cycle. When the light is red for northbound traffic, three "
        "vehicles wait behind the stop line. When the signal turns green "
        "at 00:00:08, they proceed in order. Eastbound traffic similarly "
        "stops at its red phase and proceeds on green. No vehicle enters "
        "the intersection against a red signal throughout the video.\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 traffic analyst. You are provided with captions "
        "and a description for a video containing a traffic anomaly.\n\n"
        "Task:\n"
        "1. Design an open-ended question about the traffic anomaly that "
        "requires perception and reasoning to answer.\n"
        "2. The question should ask about the root cause, right-of-way "
        "analysis, consequence, 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 (signal state, vehicle "
        "positions, movements) 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 is the root cause of the collision "
        "and how could it have been prevented?\n"
        "======\n"
        "Answer: The root cause of the collision is the dark sedan "
        "running a red light at approximately 00:00:06 while cross-traffic "
        "had a green signal and right-of-way. The sedan failed to stop or "
        "even decelerate as it approached the intersection, entering at "
        "speed and striking the white SUV that was lawfully proceeding "
        "through on green. The collision could have been prevented if "
        "the sedan driver had obeyed the red signal and stopped behind "
        "the stop line, as the signal was clearly visible with adequate "
        "approach distance.\n"
        "=====\n"
        "Reasoning:\n"
        "The video shows a standard signalized intersection. Before the "
        "collision, the eastbound signal is red while the northbound "
        "signal is green. The white SUV begins moving northbound on the "
        "fresh green signal. Simultaneously, the dark sedan approaches "
        "from the east at moderate-to-high speed without any sign of "
        "braking or deceleration. The sedan enters the intersection "
        "directly against the red signal and strikes the SUV on the "
        "passenger side.\n\n"
        "The root cause is the sedan's failure to comply with the red "
        "signal. The signal was clearly visible, the road was straight "
        "with good visibility, and there were no obstructions. The sedan "
        "had sufficient distance to stop. The collision was entirely "
        "preventable through basic signal compliance.\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 traffic analyst. You are provided with captions "
        "for a video with normal traffic 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 traffic flow patterns, signal "
        "timing effects, how vehicles interact at the intersection, "
        "pedestrian behavior, or what the observable patterns reveal "
        "about the traffic conditions.\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 does the traffic signal cycle affect "
        "the flow of vehicles through the intersection in this video?\n"
        "======\n"
        "Answer: The signal cycle creates a clear pattern of queuing and "
        "release for each direction. When the northbound signal is red, "
        "vehicles accumulate behind the stop line, forming a queue of "
        "three to four vehicles. When the signal turns green, the queue "
        "clears within approximately 10 seconds as vehicles proceed in "
        "order. During the northbound green phase, eastbound traffic is "
        "held at red, accumulating its own queue. This alternating "
        "pattern ensures orderly flow without conflicts, with each "
        "direction getting approximately equal green time.\n"
        "=====\n"
        "Reasoning:\n"
        "The video covers one full signal cycle at the intersection. "
        "Starting with the northbound red phase, we observe three "
        "vehicles queued. At 00:00:08, the signal turns green and the "
        "queue begins to clear \u2014 the first vehicle moves at 00:00:09, "
        "followed by the second at 00:00:11, and the third at 00:00:13. "
        "By 00:00:18, the intersection is clear. Meanwhile, eastbound "
        "vehicles have been accumulating at their red signal, with four "
        "vehicles visible in the queue by the time their signal turns "
        "green. This demonstrates a balanced signal cycle managing "
        "moderate traffic volume effectively.\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 traffic safety analyst. You are provided with "
        "a holistic and detailed description for a traffic 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, road layout, or environmental "
        "context of the video (e.g., \"What does the scene look like in "
        "this traffic video?\" or \"Describe the road layout and "
        "surroundings visible in the video.\").\n"
        "2. Write an **answer** — a concise yet thorough scene "
        "description covering:\n"
        "   - The camera perspective (overhead, dashboard, fixed CCTV, "
        "etc.).\n"
        "   - Weather, lighting, and approximate time of day.\n"
        "   - Road and lane layout: number of lanes, lane directions (use "
        "frame-relative terms such as \"top to bottom\", \"left to right\"), "
        "median/divider presence, road markings, and lane types "
        "(bus-only, turn-only, shoulder, etc.).\n"
        "   - Traffic infrastructure: traffic signals, signboards, street "
        "lights, crosswalks, sidewalks.\n"
        "   - Stationary objects and environment: buildings, trees, "
        "fences, parked vehicles, text overlays.\n"
        "   - People/objects present and their attributes (color, type, "
        "size) when discernible.\n"
        "   Do NOT describe events, vehicle movements, or collisions — "
        "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., \"Lane markings and a median divider indicate a divided "
        "four-lane road …\"). 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 video recording from an overhead traffic "
        "camera during the daytime, showing a four-lane roadway with one "
        "split lane. Traffic in lanes 1 and 2 flows from top to bottom of "
        "the frame, while traffic in lanes 3 and 4 flows from the bottom "
        "to the top of the frame. The split lane contains stopped "
        "traffic, and one yellow bus moves ahead towards the top of the "
        "frame in the split lane. There is a text overlay on the top "
        "right of the frame that reads \"FOX TV STATIONS\".\n"
        "</example>\n"
        "<example>\n"
        "Answer: The video captures an overhead view from a traffic "
        "signal CCTV camera at a four-way intersection. A white text "
        "overlay shows \"2024-11-04\" with a running timer. Streetlights "
        "are present along the whole intersection, starting from left to "
        "right. Sidewalks are present towards the right and left of the "
        "top lane. Traffic lights are present to the left, right, and top "
        "of the frame. There are three signboards placed on each of the "
        "traffic signal poles, respectively. The road has white markings "
        "showing the intersection clearly. There are trees to the left, "
        "in the top right corner, and in the bottom right corner of the "
        "frame. Distinctly marked road lanes are present, signifying a "
        "signal-regulated junction. There are two houses on the right "
        "side of the frame.\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 traffic safety analyst. You are provided with "
        "a holistic and detailed description for a traffic video. Your "
        "task is to produce an **event summary caption** that describes "
        "the key events and vehicle 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 traffic video?\" or \"Summarize the events and any "
        "anomalies 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 traffic: describe the traffic flow, signal "
        "changes, and any notable vehicle maneuvers.\n"
        "   - For anomalous traffic: briefly describe normal traffic, "
        "then describe the anomaly in detail, including the color and "
        "type of the involved vehicles, their driving directions and "
        "lanes, the sequence of events leading to the anomaly, and the "
        "physical aftermath (vehicles stalled, dragged, flipped, etc.).\n"
        "   - **Root cause**: For any anomalous video, explicitly state "
        "the root cause of the incident (e.g., running a red light, "
        "failing to yield, distracted driving).\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: The video shows a blue sedan entering lane 4 at the "
        "bottom of the frame and moving toward the top of the frame. It "
        "collides with a black SUV, which is moving ahead of it. After "
        "the collision, the blue sedan changed lanes to the left and "
        "entered lane 3 without turning on its blinker and became stalled "
        "in lane 3 while the black SUV moved straight in lane 4 toward "
        "the top of the frame and entered the shoulder lane, which is "
        "beside the split lane.\n"
        "</example>\n"
        "<example>\n"
        "Answer: A black sedan enters the frame from the bottom and is "
        "hit on the right side by a rider on a scooter. After the "
        "collision, the black sedan drives ahead towards the right of the "
        "frame, towards a white divider, and the scooter skids, spins, "
        "and moves forward while the rider is flung in the air and lands "
        "near a divider with bushes on it. The bus, cars and motorcycles "
        "behind this collision have come to a halt.\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 traffic safety analyst and video annotator. "
        "You are provided with a holistic and detailed description for a "
        "traffic video containing an anomaly.\n\n"
        "Task:\n"
        "1. Identify a single anomalous event (e.g., a collision, a "
        "traffic violation, a near-miss) 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 vehicles, 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 anomaly 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 baseline state that makes the "
        "anomaly identifiable) and post-event context after t2 (the "
        "resulting state 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 vehicles, 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 00:00:02 and "
        "00:00:06?\n"
        "======\n"
        "Answer: A silver sedan turns left across oncoming traffic "
        "without yielding, causing a T-bone collision with a "
        "straight-moving vehicle. The sedan spins to a halt and a dark "
        "pickup truck is simultaneously struck by a FedEx truck in the "
        "adjacent lane, with all surrounding traffic coming to a full "
        "stop by 00:00:06. The root cause is the sedan driver's failure "
        "to yield, a misjudgment that made the collision unavoidable once "
        "the turn was committed.\n"
        "=====\n"
        "Reasoning:\n"
        "Before 00:00:02, the intersection is operating normally — "
        "oncoming vehicles are flowing straight through with the "
        "right-of-way and the silver sedan is stationary in the left-turn "
        "lane, signaling intent to turn. This baseline of orderly, "
        "rule-compliant traffic is what makes the subsequent deviation "
        "identifiable as an anomaly.\n\n"
        "At 00:00:02, the sedan commits to the left turn without clearing "
        "oncoming traffic, creating an unavoidable conflict. The "
        "straight-moving vehicle has no time to brake, resulting in a "
        "T-bone impact that also displaces the pickup into the path of "
        "the FedEx truck.\n\n"
        "After 00:00:06, all vehicles have come to rest and surrounding "
        "traffic has stopped completely. The intersection, which was "
        "flowing normally moments before, is now fully obstructed. The "
        "root cause is the sedan driver's misjudgment of the available "
        "gap — a failure that made the collision inevitable once the turn "
        "was committed.\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 traffic analyst. You are "
        "provided with a holistic and detailed description for a normal "
        "traffic video.\n\n"
        "Task:\n"
        "1. Identify a single, self-contained traffic event (e.g., a "
        "vehicle completing a turn, a pedestrian crossing, a lane change, "
        "a signal phase change affecting flow) 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 vehicles, 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 (traffic rules, right-of-way, signal phases, vehicle "
        "interactions — 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 vehicles, 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 00:00:08 and "
        "00:00:12?\n"
        "======\n"
        "Answer: A white sedan waiting in the left-turn lane initiates "
        "its turn as a gap opens in oncoming traffic and smoothly "
        "completes the maneuver by 00:00:12. The driver yields until the "
        "last crossing vehicle clears, then commits to the arc and "
        "straightens into the new lane — normal left-turn yielding "
        "behavior executed correctly.\n"
        "=====\n"
        "Reasoning:\n"
        "Before 00:00:08, the white sedan has been stationary in the "
        "left-turn lane with its turn signal active, waiting as a steady "
        "stream of oncoming vehicles passes through the intersection. "
        "This waiting state is the precondition that explains why the "
        "turn begins precisely when it does.\n\n"
        "At 00:00:08, the final oncoming vehicle clears the intersection, "
        "creating a sufficient gap. The sedan enters and follows a "
        "controlled arc across the opposing lanes with no conflicting "
        "traffic, the driver committing only once right-of-way was "
        "available.\n\n"
        "After 00:00:12, the sedan is traveling straight in the new lane "
        "with no further interaction. The turn is fully resolved, and "
        "normal flow resumes — a textbook example of a driver correctly "
        "reading the traffic conditions before acting.\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 traffic safety analyst. You are provided with "
        "a holistic and detailed description for a traffic 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 traffic "
        "anomaly 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 vehicles, actions, "
        "or outcomes in the question.\n"
        "   - Please design the questions that require perception and "
        "cognition (reasoning) to answer.\n"
        "   - Note: The relationship between the two moments can be "
        "causal (cause/effect), regulatory (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 legal 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 (traffic lights, signs, road markings) that "
        "define the rules of the scene.\n"
        "   - Trace the \"Logical Bridge\" and relevant events between A "
        "and B. Explain the mechanics (e.g., momentum, reaction time, "
        "obstructed sightlines) 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 00:00:01 "
        "and the situation at 00:00:05.\n"
        "======\n"
        "Answer: At 00:00:01, a black pickup truck accelerates from "
        "behind a stopped queue to enter the intersection, while at "
        "00:00:05, a dark grey SUV is resting on its roof after a "
        "rollover. The relationship is a direct causal chain where the "
        "pickup truck's high-speed signal violation at 00:00:01 created "
        "an unavoidable T-bone collision with the SUV. This impact "
        "transferred massive kinetic energy to the SUV, causing it to "
        "flip and slide into the far-left corner of the frame, resulting "
        "in the stationary wreckage and roadway obstruction observed at "
        "00:00:05.\n"
        "=====\n"
        "Reasoning:\n"
        "The video at 00:00:01 shows a black pickup truck veering out of "
        "a stationary queue to bypass a red light, while the situation at "
        "00:00:05 shows a post-collision scene with an overturned SUV and "
        "debris. Looking at the wider context of the video, we can see "
        "that a red semi-truck and a white sedan remained stationary at "
        "the off-ramp during this entire window, which serves as an "
        "environmental anchor confirming that the vertical signal phase "
        "was red. This baseline establishes that the pickup's entry at "
        "00:00:01 was a significant departure from safe, predictable "
        "traffic behavior.\n\n"
        "The logical bridge between these timestamps is the physical "
        "impact that occurs at 00:00:02. Because the pickup truck entered "
        "the junction at such a high rate of speed, a \"point of no "
        "return\" was reached where neither driver could perform an "
        "evasive maneuver. The front end of the pickup struck the side of "
        "the SUV, initiating a rollover sequence. Between 00:00:02 and "
        "00:00:05, the SUV's momentum carried it across the asphalt while "
        "it inverted, eventually coming to a stop as its kinetic energy "
        "dissipated.\n\n"
        "We can conclude that the situation at 00:00:05 is the terminal "
        "physical resolution of the violation initiated at 00:00:01. By "
        "tracing the chain of events from the truck's initial "
        "acceleration to the SUV's final resting position, it is clear "
        "that the obstruction at 00:00:05 is the direct result of the "
        "illegal bypass of the traffic queue. The relationship is a "
        "complete causal sequence where the initial decision at 00:00:01 "
        "made the catastrophic outcome at 00:00:05 physically inevitable.\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 traffic analyst. You are provided with a "
        "holistic and detailed description for a normal traffic 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 traffic signal phase change, a vehicle "
        "beginning a lane change or turn, a pedestrian entering the "
        "crosswalk, a vehicle merging, or the start of a parking "
        "maneuver).\n"
        "   - [Timestamp B] (Focus Point): The subsequent outcome (e.g., "
        "vehicles moving through after the signal change, the vehicle "
        "completing the turn, the pedestrian reaching the other side, "
        "traffic flow resuming).\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 vehicles, 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 regulatory (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 regulatory 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 "
        "(traffic lights, signs, road markings).\n"
        "   - Trace the \"Logical Bridge\" and relevant events between A "
        "and B (e.g., gap in traffic, signal phase, vehicle motion). "
        "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 00:00:08 "
        "and the situation at 00:00:12.\n"
        "======\n"
        "Answer: At 00:00:08, a white sedan begins its left turn from the "
        "turn lane after a gap appears in oncoming traffic, and at "
        "00:00:12 the same sedan has fully completed the turn and is "
        "traveling straight in the perpendicular lane. The relationship "
        "is sequential and regulatory: the event at 00:00:08 is the "
        "driver's decision to initiate the turn once it was safe (gap in "
        "traffic, possible signal or yield rule), and the situation at "
        "00:00:12 is the natural completion of that maneuver—the vehicle "
        "has traversed the intersection, straightened its wheels, and "
        "resumed normal flow in the new lane.\n"
        "=====\n"
        "Reasoning:\n"
        "First, from observation of the video: the scene is an "
        "intersection with a left-turn lane. Before 00:00:08, the white "
        "sedan is stationary in that lane with its turn signal on while "
        "oncoming traffic passes. The two timestamps to focus on are "
        "00:00:08 and 00:00:12. At 00:00:08, the key event is the white "
        "sedan beginning its turning motion into the intersection as a "
        "gap appears in the oncoming stream. At 00:00:12, the key "
        "situation is the sedan having completed the turn, with wheels "
        "straightened and the vehicle moving in the perpendicular lane.\n\n"
        "The logical bridge between these two moments is the sedan's arc "
        "through the intersection and the absence of conflicting traffic. "
        "The driver used the gap to cross the path of the former oncoming "
        "flow; the turn was executed in one continuous motion. By "
        "00:00:12 the vehicle has left the intersection and is in a "
        "stable, straight trajectory.\n\n"
        "We can conclude that the situation at 00:00:12 is the direct "
        "result of the decision and action at 00:00:08. The relationship "
        "is one of action and completion: the event at 00:00:08 "
        "(initiating the turn when safe) leads to the situation at "
        "00:00:12 (turn completed, vehicle in new lane). There is no "
        "anomaly; the sequence follows normal traffic behavior and "
        "right-of-way.\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 traffic 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 traffic rules or flow "
        "patterns that were being followed to establish why the event had "
        "NOT yet begun.\n"
        "   - Pinpoint the exact visual 'state-change' at the start "
        "timestamp. Explain the root cause (e.g., signal violation, "
        "misjudgment, mechanical failure) that initiated the anomaly.\n"
        "   - Justify the end timestamp by describing the transition into "
        "a new, stable, or static state (e.g., vehicles reaching a full "
        "stop, clearing the intersection, or the camera view changing).\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 side-impact collision between the black "
        "pickup truck and the crossing SUV occur?\n"
        "======\n"
        "Answer: Start_Time: 00:00:01, End_Time: 00:00:05\n"
        "=====\n"
        "Reasoning:\n"
        "The video begins with a stable baseline of urban traffic flow. "
        "Between 00:00:00 and 00:00:01, a dark grey SUV is seen crossing "
        "the intersection from right to left with the right-of-way, while "
        "a queue of vehicles (including a red semi-truck) remains "
        "correctly stationary at the off-ramp signal. The event is "
        "triggered at 00:00:01 when a black pickup truck, bypassing the "
        "stationary queue at high speed, violates the red light and "
        "enters the intersection. This root-cause violation results in an "
        "immediate T-bone impact at 00:00:01, where the pickup strikes "
        "the passenger side of the SUV. The force of the impact causes "
        "the SUV to roll onto its side and slide toward the bottom-left "
        "corner of the frame. The event reaches a resolution at 00:00:05; "
        "by this time, the SUV has slid completely out of the camera's "
        "view, and the black pickup truck has come to a full, static halt "
        "in the center of the intersection with visible front-end damage, "
        "transitioning the scene from an active collision to a stationary "
        "obstruction.\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 traffic video.\n\n"
        "Task:\n"
        "1. Identify a significant traffic event that requires logical "
        "reasoning to localize. Focus on regular traffic events such as:\n"
        "   - A vehicle completing a lane change or turn\n"
        "   - A traffic signal phase change affecting vehicle flow\n"
        "   - A pedestrian crossing the street\n"
        "   - A vehicle merging into traffic\n"
        "   - A parking maneuver\n"
        "   - Traffic flow pattern changes (e.g., congestion forming or "
        "clearing)\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 traffic flow or vehicle position "
        "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., vehicle completing the maneuver, "
        "pedestrian reaching the other side, traffic resuming normal "
        "flow).\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 white sedan complete its left turn at "
        "the intersection?\n"
        "======\n"
        "Answer: Start_Time: 00:00:08, End_Time: 00:00:12\n"
        "=====\n"
        "Reasoning:\n"
        "Prior to 00:00:08, the white sedan is stationary in the "
        "left-turn lane, waiting for oncoming traffic to clear. The "
        "intersection shows a steady flow of vehicles traveling straight "
        "through from the opposite direction, and the white sedan's turn "
        "signal is visible, indicating intent to turn. At 00:00:08, a gap "
        "appears in the oncoming traffic as the last vehicle passes, and "
        "the white sedan begins its turning motion, entering the "
        "intersection and rotating leftward. The vehicle continues its "
        "arc through the intersection, crossing the path where oncoming "
        "traffic was previously flowing. By 00:00:12, the white sedan has "
        "fully completed its turn, now traveling in the perpendicular "
        "lane with its wheels straightened and maintaining a consistent "
        "forward trajectory, marking the conclusion of the turning "
        "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