.. _fwflag_add_sentence_embedding:
==========================
--add_sentence_embedding
==========================
Switch
======

--add_sentence_embedding

Description
===========

Adds sentence-embedding (e.g. sBERT) mean features, one row per group. Aliased as ``--add_sent_emb_feat``, ``--add_sbert``, ``--add_cosine_emb``, ``--add_msg_emb``.

Argument and Default Value
==========================

No argument. Boolean switch, default: ``False``. Uses the same :doc:`fwflag_embedding_model`/:doc:`fwflag_emb_class`/:doc:`fwflag_tokenizer_model`/:doc:`fwflag_batch_size` switches as :doc:`fwflag_add_bert`.

Details
=======

**Known issue — requires a CUDA GPU, unconditionally, even on CPU-only or Apple Silicon (MPS) machines.** ``FeatureExtractor.tokenizeWithLengthWarning()`` (used only by this switch's ``addSentenceEmbTable()``, not by the regular :doc:`fwflag_add_bert`/:doc:`fwflag_add_emb` path, which has its own separate, CPU-safe tokenization) calls ``tokenizer(text, ...).to('cuda')`` with no check for ``torch.cuda.is_available()`` and no CPU/MPS fallback. Verified: on a machine without CUDA, this crashes immediately with ``AssertionError: Torch not compiled with CUDA enabled``, regardless of model choice. Not usable on non-CUDA hardware as of this writing.

Other Switches
==============

Required Switches:

* :doc:`fwflag_d`, :doc:`fwflag_c`, :doc:`fwflag_t`

Example Commands
================

Requires a CUDA-enabled GPU — not usable on CPU-only/Apple Silicon machines, see Known issue above:

.. code-block:: bash

	dlatkInterface.py -d dla_tutorial -t msgs -c user_id --add_sentence_embedding \
	--embedding_model sentence-transformers/all-MiniLM-L6-v2 --emb_class auto \
	--tokenizer_model sentence-transformers/all-MiniLM-L6-v2