tree:
  max_paragraphs_per_l2: 5
  l1_segment_duration: 600.0           # L1: 每段 10 分钟（长视频适配）
  l2_clip_duration: 60.0               # L2: 每 clip 60 秒
  l3_fps: 0.5                          # L3: 0.5 帧/秒（每 2 秒一帧）
  l2_representative_frames: 6          # L2 VLM 描述用的代表帧数（从10降到6以提速）
  cache_dir: "data/videomme/trees"     # 树索引缓存目录（相对项目根目录）
  concurrency: 16                      # asyncio Semaphore 上限：每视频最多 16 路同时在途 VLM 请求

embed:
  backend: "local"                     # CPU 本地运行，无需远程嵌入服务
  model_name: "BAAI/bge-base-zh-v1.5"
  embed_dim: 768                       # bge-base-zh-v1.5 输出维度
  device: "cpu"                        # 本地 CPU 推理
  api_key: ""
  api_url: ""

llm:
  backend: "openai"                    # GPUStack 兼容 OpenAI API
  model: "gemma-4-31B"
  api_url: "http://100.83.164.94:11904/v1"
  max_tokens: 256
  temperature: 0.1
  # api_key: 从 .env 加载

vlm:
  backend: "openai"                    # GPUStack 兼容 OpenAI API
  model: "gemma-4-31B"
  api_url: "http://100.83.164.94:11904/v1"
  max_tokens: 512                      # 5帧描述 ~300 tokens，256 会截断 JSON 触发 fallback
  temperature: 0.1
  # api_key: 从 .env 加载

retriever:
  embed_dim: 768                       # 与 bge-base-zh-v1.5 维度一致
  num_heads: 4
  L_layers: 2
  L_cycles: 4
  max_rounds: 5
  ffn_expansion: 2.0
  checkpoint: null
  k_l1: 1
  k_l2: 1
  k_l3: 1
  max_paths: 5

train:
  lr: 1.0e-4
  weight_decay: 1.0e-5
  batch_size: 1
  max_epochs_phase1: 30
  max_epochs_phase2: 20
  nav_loss_weight: 1.0
  act_loss_weight: 0.1
  margin_loss_weight: 0.5
  act_lambda_step: 0.1
  act_gamma: 0.9
  eval_interval: 5
  save_dir: "data/videomme/checkpoints"
  dataset: "videomme"
  dataset_path: "data/videomme/splits/train.jsonl"