requirements.txt

# ============================================================================
# Video Analysis Pro - Core Dependencies (Python 3.10 - 3.13)
# 分层说明:
#   core  = 本文件（应用启动与主功能的最小依赖集）
#   ocr   = pip install -r requirements-ocr.txt   (PaddleOCR 文字识别，体积大)
#   全量  = ocr 已含其余可选依赖
# ============================================================================

# --- Video / CV core ---
opencv-python-headless>=4.8.0
numpy>=1.24,<2.3
scenedetect>=0.6.2
ultralytics>=8.3.0               # YOLOv11 需要 8.3+
moviepy>=2.0                     # 代码已迁移到 moviepy 2.x API (subclipped/resized)
imageio-ffmpeg                   # 自带 FFmpeg 二进制，实现 FFmpeg 自给自足
pillow>=11.3,<12                 # moviepy 2.2 上限 pillow<12；11.3 已含已知图像解析 CVE 修复
matplotlib>=3.7                  # plot_metrics / get_advanced_video_metrics
seaborn>=0.13                    # _detect_advanced_features 必需项：缺失时 Phase 3 摘要媒体全链路被禁用
pandas<3                         # seaborn 依赖；pandas 3.x 要求 numpy>=2.3 与上面 numpy<2.3 约束冲突

# --- AI core ---
torch>=2.2
faster-whisper>=1.0.0
sentence-transformers>=2.3.0,<7.0  # 2.2.2 与新版 huggingface_hub 不兼容
transformers>=4.41                # CI 曾误锁 <5.17:真实根因是 torch/torchvision 分源配对,
                                  # 与 transformers 版本无关(5.16.1/5.17.0 均可)
keyring>=24.0.0                  # API Key OS 密钥环存储 (DPAPI/Keychain)，否则降级明文 ini

# --- Persistence / utils ---
chromadb>=1.5.9                  # CVE-2026-45830/31/33 已修复
markdown2
requests
psutil
pymediainfo
pydub                            # 静音检测 + 波形提取（缺失时会静默降级）
nvidia-ml-py                     # NVIDIA GPU 监控；无 GPU 时自动跳过

# --- Agent tools (web search) ---
duckduckgo-search

# --- PDF export ---
weasyprint