Embedding Portability Ledger

The Embedding Portability Ledger treats every vector store decision as two separate bets: the query layer and the embedding layer. Agencies routinely price the first and ignore the second. A managed platform such as Pinecone or Zilliz removes indexing and rebalancing work, but the embeddings your client's corpus was vectorized with often cannot move without a full re-embed and re-index pass. That pass is the real switching cost, and it scales with corpus size, not seat count. Qdrant and Weaviate let a delivery team keep the embedding model and the store under one roof, which lowers exit cost at the price of running infrastructure. Before signing a retainer that depends on semantic search, log three numbers: corpus size, embedding model version, and the hours a full re-embed would take. Forrester's September 2026 argument that private AI deployments outperform shared public tooling applies directly here, because a portable embedding layer is what makes a private retrieval stack defensible.