Never mind...
I am using Pinecone Inference’s multilingual-e5-large embedding model (1024‑dim dense vectors).
Calls like self.pc.inference.embed(model=self.inference_model, parameters={"input_type": "passage"}) generate single dense vectors per chunk/query.
There’s no multi‑vector or sparse component—each chunk/query produces one dense embedding, and retrieval happens with standard top_k cosine search.