Quantifying retrieval degradation in RAG over time — drift detection without labeled data

Our RAG pipeline serves legal document QA and we've noticed answer quality degrading over months. The corpus grows (~500 docs/week), but the retrieval component (DensePRF + BM25 hybrid) hasn't been retrained or recalibrated. Without ground truth labels (impossible for our use case), what metrics do you use to detect retrieval drift? We're considering: - Query-result overlap decay (Jaccard over time windows) - LLM-as-judge scoring on historical query samples - Embedding space density shifts (PCA + cluster analysis) What's actually worked in production vs. what sounds good in papers? Context: LangChain, OpenAI embeddings, ~50K docs, FAISS index.