You fine-tune your embedding model on ten million docum...

The index stores embeddings based on the model’s parameters at a specific time. When the model is fine-tuned, the index becomes versioned and inconsistent with the new query encoder. This leads to two primary issues: score inconsistency, where queries and passages are no longer in the same vector space, and a stale candidate set where hard negatives are no longer representative. Periodic re-encoding is required to maintain performance. In tied-encoder cases like Contriever, this is particularly critical as query and passage encoders are updated simultaneously.