BGE-M3 on Azure AI Foundry appears to expose dense embeddings only; need guidance for sparse lexical_weights support

We are evaluating BAAI/bge-m3 on Azure AI Foundry / Azure ML for a production hybrid retrieval workload.

BGE-M3 natively supports dense retrieval, sparse retrieval, and multi-vector retrieval through FlagEmbedding. For example:

```python
from FlagEmbedding import BGEM3FlagModel

model = BGEM3FlagModel("BAAI/bge-m3", use_fp16=True)

output = model.encode(
texts,
return_dense=True,
return_sparse=True,
return_colbert_vecs=False,
)

dense_vecs = output["dense_vecs"]
lexical_weights = output["lexical_weights"]
```

However, the Azure AI Foundry-hosted deployment appears to expose only dense embedding output through the standard embedding endpoint. We have not found a documented way to request lexical_weights / sparse output.

**Questions:**
1. Does the Azure AI Foundry deployment path for BGE-M3 support sparse embeddings / lexical weights?
2. If yes, what request schema or configuration enables return_sparse=True behavior?
3. If no, is the recommended Azure pattern to deploy BGE-M3 via Azure ML managed online endpoint with a custom score.py?
4. Is there an official sample showing BGE-M3 dense + sparse output on Azure ML / Foundry managed compute?

Start by checking the standard Azure AI Foundry embedding endpoint and the Azure ML managed online endpoint with the BGE-M3 FlagEmbedding example in the issue. Determine whether sparse lexical_weights output is supported and whether custom score.py is required. Done means documenting the supported request schema or configuration, the recommended deployment path, and an official dense-plus-sparse sample if one exists.

Written by the indexing model from the issue text.