← All questions
LLMOps & ProductionHard· Indexing· Performance· Metrics

How would you pick the refresh interval for your index?

The refresh interval is a trade-off between compute costs and retrieval quality. One should derive the ratio Cr/Cs based on corpus size and encoder complexity. The capacity floor Kmin = Cr/(ρCs) defines the lower bound, while the measured top-k overlap provides the staleness ceiling. To determine if the index is truly stale, implement a probe (e.g., a 105-chunk, 26-step, 1% probe) to measure the degradation of retrieval quality rather than relying on fixed step counts.

Code playground (Python)