Generate Embeddings with Hugging Face

Implement embed(sentences, tokenizer, model), returning a (n, d) NumPy array (not a tensor) of normalized sentence embeddings.

1. Tokenize the batch with padding=True, truncation=True, return_tensors="pt".
2. Run the model under torch.no_grad().
3. Masked mean pooling. The model returns one vector per token. Average them into one per sentence, over real tokens only.
4. L2-normalize each row.

The catch: padding positions carry attention mask value 0. Average them in and every short sentence drifts toward the same meaningless point.