← All questions
RAGMedium· Model Editing· Retrieval· Fine-tuning

A product fact in your assistant is wrong. You have a 2...

Full retraining is computationally expensive, requiring 108 GB for the forward pass at fp32, and is not feasible for quick turnaround. Model editing, such as ROME, can update specific facts but requires re-benchmarking and risks damaging unrelated model behavior. Retrieval is the superior choice for product facts because it allows for simple, surgical updates by changing a single row in a datastore, which is essential for data that changes repeatedly over time.

Code playground (Python)