Chroma Cloud's hosted embedding functions

Chroma Cloud ran two embedding models on its own servers. They stop when Chroma Cloud stops.

Embedding function	Open model	Kind
chroma-cloud-qwen	Qwen/Qwen3-Embedding-0.6B	dense, 1024 dimensions
chroma-cloud-splade	prithivida/Splade_PP_en_v1	sparse

Your stored embeddings don't change. pgchroma migrate copies them as they are. Only new documents and query text need a model.

Let the migration switch them

pip install sentence-transformers
pgchroma migrate --from-cloud --api-key "$CHROMA_API_KEY" --database-url "$DATABASE_URL" \
    --local-embedding-functions

Collections that used a hosted function are configured to use the same open model through sentence-transformers, wherever the function appears in the collection's configuration or schema. The chromadb client loads it from the collection, so query_texts= and add(documents=...) keep working with no code change. Without the flag, the migration warns about every collection that uses a hosted function.

Qwen is configured in float32; in bfloat16 its vectors drift slightly. If your collection used a Qwen task (such as nl_to_code), queries need an instruction prefix: examples/local_qwen_embeddings.py shows how.

Check before December 22

Run a few of your own queries on Chroma Cloud and on pgchroma while Chroma Cloud still answers, and compare the top results. The models are the same, but the runtime differs, so confirm it on your data.