Vector stores largely interchangeable
AI engineer at Indeed on TurboPuffer vs. Vespa vs. Elasticsearch at scale
This is really a statement about where vector infrastructure competition has moved, which is away from raw retrieval quality and toward the economics and operating model of large scale search. In this workflow, changing chunk size or swapping embedding models did not force a backend change because all three systems could clear the baseline on relevance, filters, and latency. The meaningful split showed up elsewhere, with TurboPuffer winning when a product had lots of rarely touched data and spiky traffic, Vespa winning when ranking logic itself became the product, and Elasticsearch remaining strongest as part of a broader search stack.
-
TurboPuffer is built around compute and storage separation. Durable data lives in object storage, then moves into SSD and memory caches only when queries make it hot. That matters when most tenants or documents are cold most of the time, because the system is not priced like everything must sit in fast storage all day.
-
Vespa stands apart when ranking needs to combine many signals, like BM25, embeddings, freshness, popularity, or a learned model. It exposes ranking expressions, phased reranking, and model inference, so teams can decide exactly why one result beats another. That is a different job from simply storing vectors and returning nearest neighbors.
-
Elasticsearch is interchangeable on basic retrieval, but less so organizationally. Here it sits inside the company's existing search and observability stack, and Elastic also positions hybrid search as a native capability. In practice that makes Elasticsearch attractive when a team already runs it for logs, text search, and operational tooling.
Going forward, the market is likely to split more cleanly into cheap serverless retrieval for generic agent workloads, and highly programmable ranking systems for premium search and recommendation products. As retrieval quality converges, the winners will be decided by who makes cold data cheapest, hot data fastest, and ranking logic easiest to control.