PE-AV inference server for KODAPA

Contract (KODAPA app expects)

• GET /health — 200 when model is loaded.
• POST /embed_image — Body {"url": "https://..."} → {"embedding": [1024 floats]}.
• POST /embed_audio — Body {"url": "https://..."} → same.

Running on Hugging Face

1. Create the repo on HF (once): huggingface.co/new → e.g. OnurKenci/pe-av-server, Private or Public.
2. Push this folder via SSH (from this directory):
   bash

   1git init 2git remote add origin git@hf.co:OnurKenci/pe-av-server.git 3git add . 4git commit -m "PE-AV inference server for KODAPA" 5git push -u origin main
3. Build and push Docker image to Docker Hub (or another registry HF can pull from):
   bash

   1docker build --platform linux/amd64 -t YOUR_DOCKERHUB_USER/pe-av-server:latest . 2docker push YOUR_DOCKERHUB_USER/pe-av-server:latest
4. Create Inference Endpoint on HF: Inference Endpoints → New → Custom container. Model: facebook/pe-av-base. Container image: YOUR_DOCKERHUB_USER/pe-av-server:latest. Port 8000. Hardware: T4 (or A10G). Enable scale-to-zero.
5. Set PE_AV_ENDPOINT_URL and HF_TOKEN on KODAPA API servers (see PE_AV_HUGGINGFACE_BATCH_SETUP.md).

Local run (with model on disk)

Notes

• On HF Inference Endpoints, the selected Hub model is mounted at /repository; the app uses HF_HUB_MODEL_PATH=/repository so no weights are in the image.
• For audio, the processor may expect a specific sample rate; if embed_audio fails, resample to 16 kHz before calling the processor.