galthran-wq / clip-web-service / 23550511119

push

github

Add CLIP zero-shot classification endpoint (#1)

* Add CLIP zero-shot classification endpoint using onnx_clip

Replace template placeholder with a working CLIP inference service.
Uses onnx_clip (ONNX Runtime) instead of PyTorch for lightweight
CPU inference (~5-10ms/image, ~70MB deps vs ~1.5GB for torch).

- POST /classify/batch: accepts base64 images + text prompts, returns
  softmax probabilities
- CLIPService wraps OnnxClip with base64 decoding and batch inference
- Readiness probe gates on model loaded state (503 until ready)
- 18 tests, 90% coverage, mypy strict clean

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Fix per-image softmax normalization

Applied via @cursor push command

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>

75 of 89 new or added lines in 8 files covered. (84.27%)

166 of 186 relevant lines covered (89.25%)

0.89 hits per line