Argue For and Against Generative Retrieval Versus Two-Tower Plus ANN

Your team runs a two-tower retriever with an HNSW index over 50M items and a ranker on top. A principal engineer proposes replacing retrieval with a TIGER-style generative retriever: encode each item's content into a short tuple of semantic IDs with a residual-quantization autoencoder (RQ-VAE), and have a Transformer decode the next item's ID tuple with beam search. You are asked to write a one-page recommendation.

1. Explain how semantic IDs are produced and why similar items share prefixes.
2. Show by arithmetic whether 3 levels with 256 codes each can identify 50M items, and what you would change.
3. List the advantages the proposal claims (especially cold start) and the specific costs and risks (latency, freshness, ID drift, invalid generations, ranking features).
4. Describe the evaluation you would run before any migration and what must be true for a launch.
5. Give a lower-risk hybrid and say what claims in your memo are from memory and not measured.