As per my understanding, Use RAG for retrieving relevant documents. We can pass options to LLM when predicting. So, instead of predicting answer, predicting option would be a better option. Make sure that you aren't passing options when retrieving relevant documents using RAG. This doesn't violate guidelines of the competition. Hope this helps : )

When answering multiple-choice questions, the RAG system might generate a response that doesn’t match the provided options exactly. To address this, optimizing the selection process involves using similarity measures (e.g., cosine similarity) between the generated answer and the options bitlife. This way, even if the response is more verbose or slightly different, it selects the closest matching option based on content similarity.

• Use Similarity Metrics: Calculate cosine similarity between the generated answer and each option, choosing the closest match.
• Prompt Optimization: Adjust prompts to encourage concise answers that align better with options.
• Confidence Threshold: Set a threshold to avoid guessing if no option closely matches. snow rider
• Re-Ranking with QA Models: Use a separate QA model to rank options based on their similarity to the RAG output.