requirements.txt

# Core Web Framework
fastapi
uvicorn[standard]
python-multipart
orjson  # 2-3x faster JSON encoding/decoding vs stdlib json
pydantic-settings  # Pydantic v2 settings management

# YOLO and Triton
ultralytics<8.4  # Pin to <8.4: v8.4 introduced YOLO26 and breaking changes to end2end property
tritonclient[all]

# Model Export Dependencies
onnx>=1.12.0,<=1.19.1
onnxslim>=0.1.71
onnxsim>=0.4.33
onnxscript
onnxruntime-gpu
onnx-graphsurgeon  # Required for End2End TensorRT NMS graph manipulation
tensorrt-cu12==10.13.3.9  # MUST match Triton 25.10 TensorRT version (10.13.3.9) for engine compatibility

# PyTorch
torch
torchvision

# Image Processing
opencv-python
# Note: Pillow is installed as dependency of ultralytics (no need to specify)

# Scientific Computing
numpy
scipy

# HTTP Client
requests
aiohttp
httpx

# Benchmarking and Visualization
matplotlib
seaborn
pandas
psutil
tqdm
tabulate

# Configuration
pyyaml

# Structured Logging
structlog>=24.1.0  # Structured logging with context and JSON output

# OpenSearch for Visual Search
opensearch-py>=2.3.0  # Async OpenSearch client with k-NN support
transformers>=4.30.0  # CLIP tokenizer for text-to-image search

# MobileCLIP dependencies
timm  # pytorch-image-models - required for MobileCLIP architecture
huggingface_hub  # For downloading MobileCLIP checkpoints
ftfy  # Text cleaning for CLIP tokenizer (required by open_clip)
open-clip-torch>=2.24.0  # OpenCLIP with native MobileCLIP2-S2 support

# FAISS Clustering (GPU-accelerated vector clustering)
faiss-gpu-cu12>=1.10.0  # FAISS GPU for billion-scale clustering (CUDA 12)

# Duplicate Detection
imohash  # Fast constant-time image hashing (48KB sampled, ~1ms for any file size)