A step-by-step guide to deploying a custom real-time inference endpoint on Amazon SageMaker using Hugging Face's Inference DLCs and Sentence Transformers for generating document embeddings. Covers writing a custom inference.py script to override default model loading and prediction methods, packaging model artifacts into a model.tar.gz archive, uploading to S3, and deploying a HuggingFaceModel endpoint. Uses the sentence-transformers/all-MiniLM-L6-v2 model with mean pooling to produce 384-dimensional dense vector embeddings suitable for semantic search and clustering.
