What Happened
Elasticsearch Serverless introduced Minimum Search VCUs, a way to provision search capacity before predictable demand arrives. Operators can use the last 10 days of Search VCU usage to choose a whole-number minimum between their observed baseline and peak, then apply it roughly 30–60 minutes before a sale, workday start, or other expected surge. Autoscaling can still add capacity above that floor. The configured minimum remains billable until it is lowered or disabled [1].
This is a search-tier capacity control, not a change to how retrieval-augmented generation (RAG) selects or ranks evidence. The operational lesson applies across RAG stacks: query quality and capacity planning need separate controls.
Why It Matters to Businesses
A RAG application can have well-indexed content and still disappoint users if retrieval slows under load. Search latency consumes the response-time budget before the model begins answering. Prescaling gives teams a way to prepare for known peaks rather than relying only on reactive autoscaling, but it exchanges potential latency protection for a higher capacity floor [1].
Teams choosing among Elasticsearch, Weaviate, Pinecone, Qdrant, Milvus, or Vespa should therefore test more than relevance on a sample dataset. They need production-like measurements of concurrent queries, filtering, ingestion, and peak-hour latency. Frameworks such as LlamaIndex, LangChain, and Haystack help assemble retrieval workflows, but they do not remove the need to measure the underlying search service.
Kimbodo Engineering Perspective
We would treat retrieval as a measurable application component, not a vector-database feature. For many business questions, hybrid retrieval—keyword and vector search followed by reranking—is a better starting point than vector similarity alone: exact names, identifiers, and policy terms may matter as much as semantic matches. That choice should be validated against representative questions and expected evidence, not assumed.
Capacity decisions should follow the same discipline. A scheduled search-capacity floor is useful when demand is predictable and search latency is material to the user experience. For irregular traffic, a persistent floor may cost more than it saves; caching, query tuning, and autoscaling may be preferable. Minimum Search VCUs provide a floor, not a fixed ceiling or a guarantee of end-to-end response time [1].
How We Would Implement It
- Define retrieval tests covering answer evidence, permission boundaries, filtered searches, and no-result cases. Record recall, citation accuracy, and p95 latency separately.
- Build an ingestion pipeline that extracts text, attaches source and access-control metadata, chunks documents consistently, and versions embeddings so indexes can be rebuilt safely.
- Start with hybrid search where supported; add a reranker only if evaluation shows enough quality gain to justify its latency and cost. Return source identifiers with retrieved passages so answers can be traced.
- Load-test the full path—application, search, reranking, and model—at expected concurrency. For Elasticsearch Serverless with predictable peaks, inspect 10 days of Search VCU usage, set an appropriate minimum 30–60 minutes before demand, and schedule its removal afterward [1].
- Monitor search latency, retrieval relevance, empty-result rates, capacity, and spend. Re-run evaluations when documents, embedding models, filters, or ranking rules change.
Risks, Costs and Security
The principal risks are retrieving stale or irrelevant material, exposing content across permission boundaries, and paying for capacity that remains enabled after a peak. Enforce document-level authorization in retrieval rather than asking the model to disregard unauthorized passages. Treat retrieved text as untrusted input, log source identifiers without unnecessarily logging sensitive content, and test deletion and reindexing procedures. For Minimum Search VCUs specifically, assign an owner to lower or disable the floor: billing continues at the configured minimum while it is active, with additional autoscaled capacity charged separately [1].
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our RAG Development Services practice, or Estimate My RAG System.