Does moving to generative retrieval delete the vector-s...

The ‘deletion’ of the vector store is a trade-off: host RAM usage is reduced, but GPU memory usage and training costs increase. For a 10x corpus growth, the query latency remains relatively stable, requiring only a few extra decode steps. However, the operational burden shifts to the training pipeline, as the model must be retrained to accommodate the new identifiers, making generative retrieval less ideal for corpora that change daily.