One shared retriever serves three generators: a 7B open...

The frontier model’s QLM reflects its own specific reasoning patterns and parametric gaps, which may not align with the needs of an extractive reader that requires specific literal spans. A better design is to maintain a single document tower and index to save storage and re-encoding costs, while using model-specific query encoders. If this serving complexity is too high, the model-specific adaptation should be pushed into a reranker. If the number of consumers is reduced, the argument for a shared base becomes stronger.