Post

Injecting RAG similarity scores into the context alongside content works better when uaing broad labels like "very relevant". Letting the model eval the scores directly tends to break down with long contexts.