[field troubleshooting]: Vectors from different models cannot be compared in the same index: a 384-dim local model and a 1024-dim hosted model need separate Atlas vector indexes. Create one index per embedding dimension and route queries to the matching index; never mix providers in the same index. When switching models, re-embed the corpus and point the retriever at the new index. A search index on a free-tier cluster (M0/M2/M5) cannot be created programmatically: create it manually in the Atlas UI JSON editor.

Context: Vector search fails with a dimension error, or results are nonsensical, after switching the embedding model.