numpy
pandas
tqdm
pillow
python-dotenv

# Video / frames
decord
opencv-python-headless

# Closed-API backbones
google-genai          # Gemini 3 via Vertex AI
openai                # GPT-5-mini
anthropic             # Claude Haiku 4.5

# Local VLMs -- please check the original repo esp. when it comes to flashattention
torch
transformers>=4.45
qwen-vl-utils         # required by Qwen3-VL frame processing
accelerate
sentencepiece

# Agentic retrieval
# Used by experiments/agentic-retrieval/videoagent for CLIP-based frame retrieval.
# CLIP weights download on first use.
# transformers (above) already provides CLIPModel / CLIPProcessor.

# Notebook demos
jupyter
ipykernel