← All questions
RAGHard· Diagnostics· Few-shot Learning· Embeddings

Your demonstration selector uses embedding top-k and sc...

If embedding-based selection performs no better than random, the issue is likely that the demonstrations are not providing the necessary signal. Check if the top-k retrieval has collapsed to a single label or if the scoring metric is not providing a meaningful signal. It is also possible that the demonstrations are only providing format and label space information, rather than specific task-solving logic.

Code playground (Python)