Do we need a separate vector database like Pinecone or Weaviate?

No, and this is the biggest architectural change in MERN AI over the last two years. MongoDB Atlas Vector Search stores embeddings alongside operational data in the same cluster, under the same query interface, backup policy, and access controls. This removes the sync layer that used to fail silently between a separate vector database and your primary MongoDB. The exception is if you need sub-50ms multi-region vector search latency, which currently favours Pinecone.