← All questions
Security & SafetyMedium· Generative Retrieval· Compliance· SLA

Legal needs a retracted document to stop being returned...

The SLA can be met by suppressing the document at the inference layer. By masking the document identifier during constrained decoding, the retriever is prevented from emitting it. This is distinct from unlearning the association from the model weights (theta), which would require a full rebuild. One must clarify if the auditor requires non-retrieval or true unlearning to determine if the 41-minute rebuild is the only viable path.

Code playground (Python)