Your assistant retrieves the right guideline and answer...

The problem is identified as a behavior deficit because the model successfully reaches the correct context and answer, meaning recall and reranking are not the primary issues. The performance is poor under a rubric scoring format and citation, not accuracy. A strong approach prioritizes low-cost interventions like system prompts or few-shot templates. Fine-tuning is only recommended if the behavior must be strictly enforced on every query without consuming context window space.