179 questions
No questions match those filters.
← All questions
RAGHard· LoRA· Fine-tuning· RAG· Latency
A policy changed last night and must be reflected by 9...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansWhen evaluating the proposal, compare the operational costs of both paths. A LoRA training run involves significant compute, estimated at 1.5 GPU-hours for a 7B model on 50 million tokens. In contrast, re-embedding the changed documents is much faster, taking roughly 2.8 ms per document. Furthermore, a LoRA run requires an evaluation gate for the new checkpoint, whereas an upsert to the vector database does not, making re-embedding the more agile and cost-effective solution for policy updates.