config.yaml

app:
  name: Smart Classroom
  cleanup_on_exit: true
  use_ov_genai: True # True or False, applicable for provider: openvino

monitoring:
  logs_dir: ./monitoring/logs
  interval: 2
  socwatch_path: monitoring/tools/socwatch_windows_v2026.0/64/socwatch.exe
  npu_exe_path: tools/level-zero/build/bin/Release/npu_utilization.exe
  execution_logs: monitoring/executionlogs
models:
  asr:
    provider: openai # openvino, funasr and openai supported
    name: whisper-small  # can be (whisper-base, whisper-small etc) or paraformer-zh
    device: CPU # CPU Recommended
    temperature: 0.0
    diarization: False
    hf_token: None   # needed only if diarization=true
    models_base_path: "models"
    threads_limit: Null # applied only if > 0 (else defaults are used); value can be tuned based on CPU specifications
    no_speech_threshold: 0.6
    logprob_threshold: -1.0
    min_duration_sec: 0.25
    min_words: 2

  summarizer:
    provider: openvino 
    name: Qwen/Qwen2.5-7B-Instruct # Qwen/Qwen2-7B-Instruct or Qwen/Qwen2.5-7B-Instruct
    device: GPU # GPU or CPU
    weight_format: int8 # supports fp16, int4, int8 (Recommended)
    max_new_tokens: 5120
    temperature: 0.3 # 0.5 default
    use_cache: True
    models_base_path: "models"
    language: en # en or zh
    mode: dialog   # dialog | teacher | hybrid
    system_prompt: 
      en:

        Dialog: |
          Summarize the classroom transcript.

          Format:
          TEACHER = instructor
          STUDENT_XX = students

          Instructions:
          - Provide a full dialog summary.
          - Preserve the flow of discussion.
          - Include both teacher explanations and questions.
          - Ignore filler or repetition.

          Output valid Markdown using exactly these sections:

          ## Teacher Summary
          - ...

          ## Questions
          - ...

          ## Key Takeaways
          - ...

          ## Session Outline
          - ...

          Rules:
          - Bullet points only.
          - Do not quote transcript text.
          - Do not add extra headings.


        Teacher: |
          Summarize only the TEACHER explanations from the transcript.

          Instructions:
          - Ignore student dialog completely.
          - Focus strictly on teacher explanations.
          - Do not invent missing content.
          - If a section has no content, write "None".

          Output valid Markdown using exactly these sections:

          ## Teacher Summary
          - ...

          ## Questions
          - None

          ## Key Takeaways
          - ...

          ## Session Outline
          - ...

          Rules:
          - Bullet points only.
          - Do not quote transcript text.
          - Do not add extra headings.


        Hybrid: |
          Summarize with focus on TEACHER explanations,
          but use STUDENT questions as supporting context.

          Instructions:
          - Teacher explanations are primary.
          - Include student questions only when they clarify context.
          - Do not invent content.
          - Ignore filler.

          Output valid Markdown using exactly these sections:

          ## Teacher Summary
          - ...

          ## Questions
          - ...

          ## Key Takeaways
          - ...

          ## Session Outline
          - ...

          Rules:
          - Bullet points only.
          - Do not quote transcript text.
          - Do not add extra headings.
      zh:
        Dialog: |
          总结课堂转录内容。

          转录格式:
          教师 = 教师发言
          学生_XX = 学生发言

          指示:
          - 提供完整的对话摘要。
          - 保持讨论的整体流程。
          - 包含教师讲解和学生提问。
          - 忽略无意义的填充语或重复内容。

          使用以下固定章节标题输出有效 Markdown:

          ## 教师总结
          - ...

          ## 问题
          - ...

          ## 关键要点
          - ...

          ## 课程结构
          - ...

          规则:
          - 仅使用项目符号列表。
          - 不要直接引用原始转录文本。
          - 不要添加额外标题。

        Teacher: |
          仅总结转录中 教师 的讲解内容。

          指示:
          - 完全忽略学生对话。
          - 严格聚焦教师的解释内容。
          - 不要编造缺失的内容。
          - 如果某个章节没有内容，写 "None"。

          使用以下固定章节标题输出有效 Markdown:

          ## 教师总结
          - ...

          ## 问题
          - None

          ## 关键要点
          - ...

          ## 课程结构
          - ...

          规则:
          - 仅使用项目符号列表。
          - 不要直接引用原始转录文本。
          - 不要添加额外标题。


        Hybrid: |
          以 教师 讲解为重点进行总结，
          并在必要时使用 学生 提问作为上下文补充。

          指示:
          - 教师讲解为主要内容。
          - 仅在学生问题能帮助理解上下文时才包含。
          - 不要编造内容。
          - 忽略填充语或重复内容。

          使用以下固定章节标题输出有效 Markdown:

          ## 教师总结
          - ...

          ## 问题
          - ...

          ## 关键要点
          - ...

          ## 课程结构
          - ...

          规则:
          - 仅使用项目符号列表。
          - 不要直接引用原始转录文本。
          - 不要添加额外标题。

    model_hub: huggingface # huggingface or modelscope

  va:
    models_base_path: "models"
    threshold: 0.5 # confidence threshold for YOLO detections
  
  embedding:
    name: "BAAI/bge-large-en-v1.5"
    min_score: 0.65


mindmap:
  system_prompt:
    en: |
      Generate a comprehensive jsMind JSON structure that accurately captures ALL key concepts from the summary. Follow these strict rules:
 
      1. The output must be a valid JSON object with "meta", "format", and "data" properties
      2. Set format to "node_tree"
      3. There must be exactly ONE root node with id "root"
      4. If the summary contains multiple top-level topics, create a synthetic root topic that reflects the main theme:
         - Use the actual title if provided 
         - Otherwise use descriptive titles like "Class Overview", "Lecture Summary", "Main Topics"
      5. Each node must have "id" and "topic" properties
      6. Child nodes are placed in a "children" array
      7. IMPORTANT: Extract ALL key details and concepts from the summary:
         - Include specific examples, definitions, and explanations mentioned
         - Capture contrasts, comparisons, and relationships between concepts
         - Include technical terms, equations, or scientific principles mentioned
         - Don't oversimplify - preserve the depth and nuance of the content
      8. Use descriptive and specific topic names that reflect the actual content:
         - Instead of "Topic A" use "Discontinuous electron transitions between orbits"
         - Instead of "Definition" use "Areas without atomic particles (ether/empty space)"
         - Capture the essence and specificity of each concept
      9. Create 2-4 levels of hierarchy to properly organize information:
         - Main topics (from headings)
         - Sub-concepts (key points within each topic)
         - Specific details (examples, explanations, technical details)
      10. Ensure conceptual accuracy - each node should represent a distinct, meaningful concept
      11. Do NOT include any markdown formatting, code fences, or explanatory text
      12. Output only the raw JSON structure
 
      Example of GOOD specificity:
      {
        "meta": {
          "name": "physics_lecture",
          "author": "ai_assistant", 
          "version": "1.0"
        },
        "format": "node_tree",
        "data": {
          "id": "root",
          "topic": "Quantum Mechanics and Wave Theory",
          "children": [
            {
              "id": "wave_particle_duality",
              "topic": "Wave-Particle Duality",
              "children": [
                {"id": "photon_behavior", "topic": "Photons exhibit both wave and particle properties"},
                {"id": "double_slit", "topic": "Double-slit experiment demonstrates interference"},
                {"id": "measurement_effect", "topic": "Measurement collapses wave function"}
              ]
            },
            {
              "id": "uncertainty_principle",
              "topic": "Heisenberg Uncertainty Principle",
              "children": [
                {"id": "position_momentum", "topic": "Cannot precisely know both position and momentum"},
                {"id": "mathematical_formulation", "topic": "Δx·Δp ≥ ℏ/2 mathematical relationship"},
                {"id": "quantum_limitations", "topic": "Fundamental limit of quantum measurements"}
              ]
            }
          ]
        }
      }

    zh: |
      生成全面的jsMind JSON结构，准确捕获摘要中的所有关键概念。严格遵循以下规则：
 
      1. 输出必须是包含"meta"、"format"和"data"属性的有效JSON对象
      2. 将format设置为"node_tree"
      3. 必须且只能有一个id为"root"的根节点
      4. 如果总结内容包含多个顶级主题，创建反映主要主题的合成根主题：
         - 如果提供了实际标题则使用（如"相对论与原子世界"）
         - 否则使用描述性标题如"课堂概览"、"课程总结"、"主要内容"
      5. 每个节点必须有"id"和"topic"属性
      6. 子节点放在"children"数组中
      7. 重要：从摘要中提取所有关键细节和概念：
         - 包括提到的具体例子、定义和解释
         - 捕获概念之间的对比、比较和关系
         - 包括提到的技术术语、方程或科学原理
         - 不要过度简化 - 保持内容的深度和细致差别
      8. 使用反映实际内容的描述性和具体的主题名称：
         - 不用"主题A"而用"电子在轨道间的不连续跃迁"
         - 不用"定义"而用"无原子粒子区域（以太/空间）"
         - 捕获每个概念的本质和特殊性
      9. 创建2-4层级结构来正确组织信息：
         - 主要主题（来自标题）
         - 子概念（每个主题内的关键点）
         - 具体细节（例子、解释、技术细节）
      10. 确保概念准确性 - 每个节点应代表一个独特、有意义的概念
      11. 不要包含任何markdown格式、代码块或解释性文本
      12. 只输出原始JSON结构
 
      良好特异性示例：
      {
        "meta": {
          "name": "physics_lecture",
          "author": "ai_assistant",
          "version": "1.0"
        },
        "format": "node_tree", 
        "data": {
          "id": "root",
          "topic": "量子力学与波动理论",
          "children": [
            {
              "id": "wave_particle_duality",
              "topic": "波粒二象性",
              "children": [
                {"id": "photon_behavior", "topic": "光子同时表现出波动和粒子特性"},
                {"id": "double_slit", "topic": "双缝实验证明干涉现象"},
                {"id": "measurement_effect", "topic": "测量导致波函数坍缩"}
              ]
            },
            {
              "id": "uncertainty_principle",
              "topic": "海森堡不确定性原理",
              "children": [
                {"id": "position_momentum", "topic": "无法同时精确知道位置和动量"},
                {"id": "mathematical_formulation", "topic": "Δx·Δp ≥ ℏ/2 数学关系"},
                {"id": "quantum_limitations", "topic": "量子测量的基本限制"}
              ]
            }
          ]
        }
      }
  min_token: 20

audio_preprocessing:
  chunk_duration_sec: 15
  silence_threshold: -35  # in dB
  silence_duration: 0.3   # minimum silence length in seconds
  search_window_sec: 1    # how far to look for silence if no silence exactly at chunk boundary
  chunk_output_path: chunks/

audio_util:
  max_size_mb: 200
  allowed_extensions:
    - .wav
    - .mp3
    - .m4a
  chunk_size: 52428800   # 1024 * 1024 * 50 = 50MB

pipeline:
  delete_chunks_after_use: true

va_pipeline:
  mediamtx_path: components/va/bin/mediamtx
  plugin_path: components/va
  rtsp_codec: h264 # h264 or h265
  output_rtsp_url: rtsp://127.0.0.1:8554
  hls_base_url: http://127.0.0.1:8888