---
title: Modernizing Enterprise Search | Vespa.ai
description: "A manager’s guide to replacing Elasticsearch with Vespa: where Elasticsearch strains, benchmark and customer results, and building the business case."
image: https://content.vespa.ai/hubfs/RGP/vespa-esguide-reader/assets/og.jpg
---

[Skip to the document](https://content.vespa.ai/modernizing-enterprise-search#s01)

Modernizing enterprise searchExecutive summarySix reasons teams switch from Elasticsearch to VespaWhere Elasticsearch strains as requirements growWhat changes with VespaThe evidenceBuilding the business caseWhat a migration involvesWork with a migration partner: SearchplexNext steps

 Contents

Close

## Contents

---

Elasticsearch is a proven search and analytics engine, and for many workloads it remains a sound choice.

[01 Executive summary](https://content.vespa.ai/modernizing-enterprise-search#s01) [02 Where Elasticsearch strains as requirements grow](https://content.vespa.ai/modernizing-enterprise-search#s03) [03 What changes with Vespa](https://content.vespa.ai/modernizing-enterprise-search#s04) [04 The evidence](https://content.vespa.ai/modernizing-enterprise-search#s05) [05 Building the business case](https://content.vespa.ai/modernizing-enterprise-search#s06) [06 What a migration involves](https://content.vespa.ai/modernizing-enterprise-search#s07) [07 Next steps](https://content.vespa.ai/modernizing-enterprise-search#s09) [08 About Vespa](https://content.vespa.ai/modernizing-enterprise-search#about)

![Vespa.ai](https://content.vespa.ai/hubfs/RGP/vespa-esguide-reader/img/VespaAI-logo-white-RGB.svg)

---

# Modernizing enterprise search

A manager’s guide to replacing Elasticsearch with Vespa

one Vespa deploymentsix Elasticsearch clustersVintedone Vespa deploymentsix Elasticsearch clusters

Vespa.ai5 minute read

## Executive summary

---

Elasticsearch is a proven search and analytics engine, and for many workloads it remains a sound choice. But as search becomes central to revenue, and AI applications demand fresher data, hybrid relevance, and richer ranking, many teams find that keeping Elasticsearch performing well takes more infrastructure, more tuning, and more engineering time every year.

This guide is for managers weighing whether to replace Elasticsearch with Vespa. It covers where Elasticsearch strains as requirements grow, what changes with Vespa, the evidence from our [published benchmark](https://content.vespa.ai/vespa-vs-elasticsearch-performance-comparison?hsLang=en) and from customers who have migrated, how to build the business case, and what a migration involves.

- 8.5× higher throughput per CPU core for hybrid queries in our published benchmark (up to 12.9× for vector and 6.5× for lexical queries)
- About 4× more efficient in-place updates
- An estimated 5× lower infrastructure cost for the tested workload
- Vinted halved its search servers and improved search latency 2.5× after migrating

### Six reasons teams switch from Elasticsearch to Vespa

---

- **Performance and cost efficiency:** 6.5× to 12.9× more queries per CPU core in our benchmark, with an estimated 5× lower infrastructure cost.
- **Fresh data in every result:** updates are searchable as soon as they’re written, with no wait for a refresh.
- **Scaling without shard planning:** data redistributes automatically as nodes are added or removed.
- **Built for AI applications:** tensors, multi-vector search, and machine-learned ranking run inside the platform.
- **Cheaper updates:** changing a price or a counter updates that field in place, instead of rewriting the whole document.
- **Less to operate:** Vinted replaced six Elasticsearch clusters with one Vespa deployment and halved its server count.

## Where Elasticsearch strains as requirements grow

---

Elastic has added hybrid search, built-in embeddings, learning to rank, and late-interaction vectors in recent releases. For most teams, the question isn’t whether Elasticsearch can do something, but what it costs to do it at scale.

### Freshness

Elasticsearch makes changes searchable only after a refresh, and every update rewrites the whole document. When prices, stock, or engagement signals change constantly, that means continuous reindexing load and results that lag behind the business. At Vinted, changes took 300 seconds to appear in search; on Vespa, they take 5.

### Hybrid relevance

Keyword and vector results come from separate retrievers and are merged after retrieval, so business signals such as margin, availability, or personalization aren’t part of the same ranking decision. Teams tune the merge weights by hand and retune them as data changes.

### Ranking and personalization

Learning to rank rescores only a window of top documents on each shard and requires a paid subscription. Logic beyond that usually means scripts or a separate reranking service, which adds infrastructure and loses context the search engine already had.

### Scaling and operations

Each index’s shard count is fixed when it’s created, so growth beyond the original plan means splitting, shrinking, or reindexing. Larger deployments often end up running several clusters for different workloads, each to size, monitor, and upgrade.

## What changes with Vespa

---

Vespa runs retrieval, ranking, model inference, and real-time updates in one platform. For the business, that shows up in five ways:

- **Lower infrastructure cost:** more queries per CPU core means fewer servers for the same traffic.
- **More relevant results:** keyword, vector, and business signals, plus your own models, combine in one ranking decision.
- **Results that reflect the business right now:** price, stock, and engagement changes are searchable immediately.
- **Faster experimentation:** teams can test new ranking ideas by selecting a different rank profile per query, without new infrastructure.
- **Less operational overhead:** data distribution is automatic, and one deployment can replace several clusters.

### Where Elasticsearch teams move first

- **Product search and discovery:** semantic search, personalization, and live inventory and pricing in a single query.
- **Recommendations and personalization:** user behavior and preferences shape results as they happen.
- **Retrieval-augmented generation (RAG) and AI agents:** hybrid retrieval and multi-phase ranking give language models more relevant, current context.

## The evidence

---

### Benchmark results

Vespa Engineering ran a reproducible benchmark comparing Vespa 8.427.7 and Elasticsearch 8.15.2 on an e-commerce application with one million products. It covers document ingestion, updates, and lexical, vector, and hybrid queries, all configured to return equivalent results. The [full 80-page report](https://content.vespa.ai/vespa-vs-elasticsearch-performance-comparison?hsLang=en) and a [six-page executive summary](https://content.vespa.ai/hubfs/PDF-reports/Executive-Summary-ElasticSearch-vs-Vespa-Performance-Comparison.pdf?hsLang=en) are available.

Queries per second per CPU core, Vespa against ElasticsearchBar chart in two panels, Vespa against Elasticsearch. Queries, in queries per second per CPU core: Vespa’s bars are taller for hybrid, vector and lexical queries, marked 8.5×, 12.9× and 6.5×. Writes, in throughput per CPU core: Elasticsearch’s bars are taller for feed and refeed, where Vespa’s are marked 0.3× and 0.7×, and Vespa’s bar is taller for update, marked 4.2×.QueriesQPS per CPU Core050100150200250300350hybridvectorlexicalWritesThroughput per CPU Core0500100015002000feedrefeedupdateVespa8.5×12.9×6.5×0.3×0.7×4.2×ElasticsearchQueries per second per CPU core, Vespa against ElasticsearchBar chart in two panels, Vespa against Elasticsearch. Queries, in queries per second per CPU core: Vespa’s bars are taller for hybrid, vector and lexical queries, marked 8.5×, 12.9× and 6.5×. Writes, in throughput per CPU core: Elasticsearch’s bars are taller for feed and refeed, where Vespa’s are marked 0.3× and 0.7×, and Vespa’s bar is taller for update, marked 4.2×.QueriesQPS per CPU Core050100150200250300350hybridvectorlexicalWritesThroughput per CPU Core0500100015002000feedrefeedupdateVespa8.5×12.9×6.5×0.3×0.7×4.2×Elasticsearch

Figure 1. Throughput per CPU core, Vespa against Elasticsearch, for three kinds of query and three kinds of write. Each Vespa bar carries its multiple of the Elasticsearch bar beside it. From Vespa’s published benchmark report.

- Hybrid queries: 8.5× higher throughput per CPU core.
- Vector queries: up to 12.9× higher throughput per CPU core.
- Lexical queries: 6.5× higher throughput per CPU core.
- Updates: about 4× more efficient for in-place updates. Elasticsearch is efficient at the initial bulk load, while Vespa handles steady-state queries and updates more efficiently.

Hybrid query latency and CPU use as load increasesTwo line charts sharing one axis of queries per second, hybrid queries, Vespa against Elasticsearch. Top, average latency in milliseconds: Elasticsearch’s latency climbs steeply as its load rises, while Vespa’s stays nearly flat across a much wider range of load. Bottom, CPU usage: Elasticsearch’s rises steeply, Vespa’s gradually.Average Latency (ms)10203040506070CPU Usage0%20%40%60%80%100%0500100015002000250030003500400045005000Queries per Second (QPS)VespaVespaElasticsearchElasticsearchHybrid query latency and CPU use as load increasesTwo line charts sharing one axis of queries per second, hybrid queries, Vespa against Elasticsearch. Top, average latency in milliseconds: Elasticsearch’s latency climbs steeply as its load rises, while Vespa’s stays nearly flat across a much wider range of load. Bottom, CPU usage: Elasticsearch’s rises steeply, Vespa’s gradually.Average Latency (ms)10203040506070CPU Usage0%20%40%60%80%100%010002000300040005000Queries per Second (QPS)VespaVespaElasticsearchElasticsearch

Figure 2. Hybrid queries as load increases: average latency and CPU use against queries per second, Vespa and Elasticsearch. From Vespa’s published benchmark report.

Vespa’s lower CPU use per query translates directly into infrastructure cost: section 10 of the report estimates up to a 5× reduction for the tested workload. Results depend on workload and configuration, so treat them as a starting point for your own evaluation.

### Customer results

> “The migration was a roaring success. We managed to cut the number of servers we use in half (down to 60). The consistency of search results has improved since we’re now using just one deployment (or cluster, in Vespa terms) to handle all traffic. Search latency has improved by 2.5x and indexing latency by 3x. The time it takes for a change to be visible in search has dropped from 300 seconds (Elasticsearch’s refresh interval) to just 5 seconds.”

[Vinted Engineering](https://vinted.engineering/2024/09/05/goodbye-elasticsearch-hello-vespa/)

- **Kleinanzeigen:** rebuilt its homepage recommendations on Vespa, combining clicks, searches, and preferences, for a substantial uplift in listing clicks.
- **Splore:** moved from Elasticsearch to Vespa with Searchplex, adding hybrid search and cutting reported search latency from over a second to about 20 milliseconds.

## Building the business case

---

A successful migration starts with a clear view of why it’s needed. That view drives the business case, the ROI assessment, and a well-informed go or no-go decision. Priorities vary, but these drivers come up most often:

- **Cost savings:** higher query throughput per CPU core, fewer clusters to run, and automatic data distribution instead of shard planning. Vespa Cloud removes cluster operations entirely.
- **Customer engagement:** more relevant results from a single ranking decision, personalization that responds immediately, and lexical, vector, and structured search in one query.
- **Operational resilience:** fewer moving parts, no reindexing to change shard counts, and results that stay current.

### Costs and risks to plan for

- **Learning curve:** Vespa’s application packages, schemas, and rank profiles work differently from Elasticsearch, so budget time for your team to learn them.
- **Migration effort:** schema and query translation, a proof of concept, benchmarking, and a phased cutover. Effort depends on the size and complexity of your deployment.
- **Running both systems:** Elasticsearch stays in service during the cutover, so plan for a period of parallel running.

Three things reduce that risk: a proof of concept on a representative sample of your production data before you commit, Vespa Cloud to take on operations, scaling, and upgrades, and an experienced migration partner.

Benchmark on your own data before you decide. Vendor numbers, including ours, are a starting point; the decision should rest on your own relevance, latency, throughput, and cost requirements.

## What a migration involves

---

A proven migration runs in seven stages: audit your current setup, form a migration team, build a proof of concept, run scaled benchmarks, test with shadow traffic, A/B test relevance, and cut over gradually with Elasticsearch running in parallel. The [Elasticsearch to Vespa migration brief](https://vespa.ai/competitors/elasticsearch-alternative/elasticsearchmigration/) covers each stage in detail for your engineering team.

A migration in seven stagesSeven stages in order, each leading to the next: audit your current setup, form a migration team, build a proof of concept, run scaled benchmarks, test with shadow traffic, A/B test relevance, and cut over gradually, with Elasticsearch running in parallel during that last stage.1audit your current setup2form a migration team3build a proof of concept4run scaled benchmarks5test with shadow traffic6A/B test relevance7cut over graduallywith Elasticsearch running in parallelA migration in seven stagesSeven stages in order, each leading to the next: audit your current setup, form a migration team, build a proof of concept, run scaled benchmarks, test with shadow traffic, A/B test relevance, and cut over gradually, with Elasticsearch running in parallel during that last stage.1audit your current setup2form a migration team3build a proof of concept4run scaled benchmarks5test with shadow traffic6A/B test relevance7cut over graduallywith Elasticsearch running in parallel

Figure 3. A migration in seven stages, in the order they run. Elasticsearch keeps running in parallel during the gradual cutover.

### Work with a migration partner: Searchplex

---

Searchplex, our migration partner, offers audit and migration services for teams moving from Elasticsearch to Vespa, from the first assessment through production cutover. Searchplex helped Splore complete its move to Vespa.

## Next steps

1. Share the [migration brief](https://vespa.ai/competitors/elasticsearch-alternative/elasticsearchmigration/) with your engineering team.
2. Read the [benchmark report](https://content.vespa.ai/vespa-vs-elasticsearch-performance-comparison?hsLang=en).
3. [Talk to our team about scoping a proof of concept](https://vespa.ai/contact-sales/?source=Resources%20-%20Modernizing%20Enterprise%20Search&t1_source=Marketing&utm_source=vespa-resources&utm_medium=resource-page&utm_campaign=modernizing-enterprise-search&utm_content=next-steps).

## About Vespa

Vespa.ai develops the Vespa AI Search Platform, which companies such as Perplexity, Spotify, Yahoo, and Vinted use to run AI search, recommendations, and retrieval-augmented generation in production. Vespa brings retrieval, ranking, and model inference together in one platform, so teams can deliver relevant, current results at any scale.

[Talk to sales](https://vespa.ai/contact-sales/?source=Resources%20-%20Modernizing%20Enterprise%20Search&t1_source=Marketing&utm_source=vespa-resources&utm_medium=resource-page&utm_campaign=modernizing-enterprise-search&utm_content=body)

Read next

[Built for Change: Why Vespa Outpaces Elasticsearch in Modern AI and Search Workloads](https://content.vespa.ai/built-for-change-why-vespa-outpaces-elasticsearch-in-modern-ai-and-search-workloads-document?hsLang=en)

© Vespa.ai 5 minute read

```json
{
  "@context" : "https://schema.org",
  "@type" : "Article",
  "articleSection" : [ "Executive summary", "Where Elasticsearch strains as requirements grow", "What changes with Vespa", "The evidence", "Building the business case", "What a migration involves", "Next steps", "About Vespa" ],
  "dateModified" : "2026-11-11",
  "description" : "A manager’s guide to replacing Elasticsearch with Vespa: where Elasticsearch strains, benchmark and customer results, and building the business case.",
  "headline" : "Modernizing enterprise search",
  "inLanguage" : "en",
  "publisher" : {
    "@type" : "Organization",
    "name" : "Vespa.ai",
    "url" : "https://vespa.ai"
  },
  "url" : "https://content.vespa.ai/modernizing-enterprise-search",
  "wordCount" : 1335
}
```