(Architecture / Design Question - No specific error logs)
I love Weaviate’s hybrid search, but I’m struggling with the architecture around context rot.
If I have a highly relevant 3-year-old chunk and a moderately relevant chunk from yesterday, standard hybrid search still heavily favors the old one because the vector similarity is higher.
Is anyone doing mathematical freshness scoring (like a temporal decay curve) before embedding, or are we all just relying on post-retrieval metadata filters to drop old data? Looking for best practices on how to actually weight recency without destroying relevance.