Your team is selecting a vector store for a RAG system that will serve 10,000 concurrent users, store 50 million document vectors, require sub-100ms query latency at p99, and need to support metadata filtering (date range, department, classification level). The system requires no maintenance downtime. Which criteria should most heavily weight the vector store selection decision?
Select an answer to reveal the explanation.
Short Explanation
You need the managed vector store service with built-in horizontal scaling, native metadata filtering, proven SLAs, and active development. Those operational constraints decide the choice at this scale, so treat BEIR benchmark performance as secondary.
Full Explanation
The correct choice is the managed vector store service with built-in horizontal scaling, native metadata filtering support, proven SLAs at the stated QPS and dimension scale, and active development, with BEIR benchmark performance evaluated as a secondary criterion, because the operational requirements — 10K concurrent users, 50M vectors, sub-100ms p99 latency, metadata filtering, and zero-downtime maintenance — are the binding constraints for vector store selection at this scale. A managed service with horizontal scaling handles concurrent load. Native metadata filtering for date range, department, and classification is a functional requirement; stores without this capability require client-side post-filtering, which degrades performance at scale. Proven SLAs provide contractual guarantees. BEIR benchmark performance, while important, is a secondary criterion given that the operational constraints will rule out most options first. The vector store with the highest cosine similarity accuracy on the BEIR benchmark should not be selected, because BEIR benchmark accuracy, while a useful quality signal, does not address the scale, latency, filtering, or operational requirements that are the actual selection constraints at this scale. The open-source vector store with the largest GitHub star count should not be selected, because GitHub stars measure developer interest, not production readiness — many popular open-source projects require significant operational expertise to run at 50M vectors and 10K concurrent users. The vector store that supports the largest vector dimension, such as 4096 dimensions, should not be chosen, because selecting based on maximum dimension support for future embedding models prioritizes a hypothetical future requirement over present, concrete operational needs.