← All questions
EvaluationHard· Diagnostics· RAG· In-Context Learning

Your RAG assistant scores well on the offline set and d...

Start by checking the production logs to see if the gold passage was retrieved. If not, the issue is retrieval. If the gold passage is present, keep the context fixed and permute the order or template of the demonstrations. If the performance varies significantly, the issue lies in the ICL/demonstration strategy. Also, consider that production queries may be longer, pushing the gold passage into the middle of the context window, which can degrade performance.

Code playground (Python)