What can degrade silently
- Cache hit rate drops — pipeline regression, cache key change, cache wiped
- Top-k score distribution shifts — corpus drift, embedding model regression
- Empty result rate climbs — filter too aggressive, score floor too high, or the corpus has gaps
- Latency creeps up — index getting bigger, autovacuum behind, GPU contention
- LLM 'I don't know' rate climbs — retrieval quality decay reaches the LLM
The minimum dashboard
Six counters, plotted over time, will catch most silent failures before users do:
- Queries per minute
- Median + p95 retrieval latency (per stage)
- Median + p95 top-1 score
- Empty-result rate (no candidate above the floor)
- Embedding cache hit rate
- LLM 'I don't know' rate