You have deployed a DSI-style retriever. The product te...

When using a DSI-style retriever, the model is constrained by its decoding process, which cannot emit an identifier outside of its trie. Adding new documents daily without retraining means the model has no learned association between queries and these new identifiers. A common mistake is to suggest nightly fine-tuning, which is operationally expensive and can destroy the existing index. Instead, one must calculate the retrain cost and define a staleness window that the product team can manage.