--- title: Becoming AI Native | Vespa.ai description: How organizations are putting their proprietary data to work with AI image: https://content.vespa.ai/hubfs/RGP/vespa-ainative-reader/assets/og.jpg --- [Skip to the document](https://content.vespa.ai/becoming-ai-native-document#s02) Becoming AI NativeAbout this eBookData is (Still) the New OilRetrieval at WorkPerplexity — Building AI Search at Web ScaleRavenPack — AI Research Across Billions of Financial DocumentsThomson Reuters: Bringing Legal Search from Research to ProductionElicit — Better RAG Starts with Better SearchMetal — Retrieval Built for AI AgentsYahoo — AI Applications at Internet ScaleSpotify — Helping Listeners Discover Content Beyond KeywordsDeviantArt — Personalized Discovery Across a Billion DocumentsKleinanzeigen — Reinventing Personalized Marketplace DiscoveryVinted — Better Product Discovery, Measurable Business ImpactSummaryAbout Vespa.ai Contents Close ## Contents --- How organizations are putting their proprietary data to work with AI [01 About this eBook About](https://content.vespa.ai/becoming-ai-native-document#s02) [02 Data is (Still) the New Oil The premise](https://content.vespa.ai/becoming-ai-native-document#s03) [03 Retrieval at Work Use cases](https://content.vespa.ai/becoming-ai-native-document#s04) [Perplexity — Building AI Search at Web Scale](https://content.vespa.ai/becoming-ai-native-document#case-perplexity) [RavenPack — AI Research Across Billions of Financial Documents](https://content.vespa.ai/becoming-ai-native-document#case-ravenpack) [Thomson Reuters: Bringing Legal Search from Research to Production](https://content.vespa.ai/becoming-ai-native-document#case-thomson-reuters) [Elicit — Better RAG Starts with Better Search](https://content.vespa.ai/becoming-ai-native-document#case-elicit) [Metal — Retrieval Built for AI Agents](https://content.vespa.ai/becoming-ai-native-document#case-metal) [Yahoo — AI Applications at Internet Scale](https://content.vespa.ai/becoming-ai-native-document#case-yahoo) [Spotify — Helping Listeners Discover Content Beyond Keywords](https://content.vespa.ai/becoming-ai-native-document#case-spotify) [DeviantArt — Personalized Discovery Across a Billion Documents](https://content.vespa.ai/becoming-ai-native-document#case-deviantart) [Kleinanzeigen — Reinventing Personalized Marketplace Discovery](https://content.vespa.ai/becoming-ai-native-document#case-kleinanzeigen) [Vinted — Better Product Discovery, Measurable Business Impact](https://content.vespa.ai/becoming-ai-native-document#case-vinted) [04 Summary The question](https://content.vespa.ai/becoming-ai-native-document#s06) ![Vespa.ai](https://content.vespa.ai/hubfs/RGP/vespa-ainative-reader/img/VespaAI-logo-white-RGB.svg) --- # Becoming AI Native How organizations are putting their proprietary data to work with AI 9 minute read Vespa.ai 01 / About ## About this eBook --- Becoming AI native is about more than adding large language models to existing applications. As foundation models become increasingly capable and accessible, differentiation shifts toward proprietary information and how effectively organizations can retrieve, rank and use it. This eBook explores that shift through real-world examples of retrieval at work. Drawing on organizations using the Vespa AI Search Platform across RAG, AI agents, search, recommendations, personalization and product discovery, it shows how retrieval is being applied to very different business problems, from improving commercial outcomes and customer experiences to supporting AI applications operating across billions of documents. Together, these examples illustrate an important point: retrieval isn't simply another component of the AI stack. It is how organizations put their information to work and create business value. 02 / The premise ## Data is (Still) the New Oil --- Long before generative AI, organizations competed on proprietary information and how effectively they used it. As AI becomes embedded in core business processes, that advantage is taking new forms. So far, becoming AI native has focused largely on adding LLMs to applications. But as foundation models become increasingly capable and accessible, the model itself becomes less of a differentiator. What increasingly sets organizations apart is their information and how effectively they put it to work. The challenge is making that information available to AI applications. LLMs need the right information, at the right time and in the right context, without delays, stale data or inaccuracies. That makes retrieving and ranking proprietary information critical to the quality of the AI experience. Putting proprietary data to work means reliably retrieving and ranking the right information for the right user or application, at the moment it matters and at production scale. 03 / Use cases ## Retrieval at Work --- Organizations use Vespa to solve a remarkably diverse range of AI problems. These examples show how the same platform for retrieving and ranking the right information supports both emerging generative AI experiences and established search, recommendation and discovery applications. While RAG, AI search and increasingly AI agents are creating new retrieval requirements, the same underlying capabilities also power recommendations, personalization, product discovery, advertising and other AI applications. The Vespa AI Search Platform brings together hybrid retrieval, advanced ranking and machine learning inference in a single platform. It combines text, vector and structured data to find relevant information, then ranks results using the context of each request, from user preferences to business priorities. With real-time updates and low-latency serving at scale, it provides a common foundation for AI experiences that depend on accurate, timely and relevant information. More than 100 companies run applications on Vespa Cloud, alongside a global open-source community numbering in the thousands. Together, they are applying retrieval to an extraordinarily diverse range of business problems. And these aren't necessarily separate worlds. Organizations can use Vespa.ai for the applications they need today and extend the same retrieval foundation as new AI experiences emerge. A retailer might start with Product Discovery, for example, then add conversational or agentic commerce over time. The adjacent table lists examples of real customer use cases. They may look very different, but they share a common requirement: retrieving and ranking the right information, in the right context, at the moment it is needed. The customer examples that follow show how that plays out in practice — from RAG across billions of financial documents and retrieval driven by AI agents, to personalized discovery, measurable commercial impact and trustworthy, evidence-grounded AI. They also show that this isn't a capability that emerged with generative AI. The same retrieval foundation has been supporting large-scale AI applications for decades and is now evolving to meet a new generation of AI experiences. ### Generative AI and RAG Examples - Smarter investment research Agents retrieve and connect relevant information across complex institutional knowledge. - Better AI answers Finding and ranking trustworthy information for AI-generated responses. - AI-powered research Helping researchers find and synthesize relevant information across large information collections. - Meeting intelligence Making conversations, decisions and commitments searchable and available to AI applications. - Better-informed public services Retrieving relevant government knowledge for AI-powered services and applications. ### Search, Recommendation & Discovery Examples - Faster financial research Retrieving and ranking relevant intelligence across enormous datasets. - More relevant marketplaces Matching shoppers with products across rapidly changing catalogs. - Personalized content discovery Matching people with content, creators and communities based on their interests. - Smarter job matching Combining multiple signals to connect candidates with relevant opportunities. - Evidence-based advocacy Making verified information easier to discover and use. 03 / Customer stories - [![Perplexity](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-perplexity.jpg?width=105&height=27&name=logo-perplexity.jpg)](https://content.vespa.ai/becoming-ai-native-document#case-perplexity) - [![RavenPack](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-ravenpack.png?width=149&height=19&name=logo-ravenpack.png)](https://content.vespa.ai/becoming-ai-native-document#case-ravenpack) - [![Thomson Reuters](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-thomson-reuters.png?width=90&height=31&name=logo-thomson-reuters.png)](https://content.vespa.ai/becoming-ai-native-document#case-thomson-reuters) - [![Elicit](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-elicit.png?width=99&height=28&name=logo-elicit.png)](https://content.vespa.ai/becoming-ai-native-document#case-elicit) - [![Metal](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-metal.png?width=99&height=28&name=logo-metal.png)](https://content.vespa.ai/becoming-ai-native-document#case-metal) - [![Yahoo](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-yahoo.png?width=97&height=29&name=logo-yahoo.png)](https://content.vespa.ai/becoming-ai-native-document#case-yahoo) - [![Spotify](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-spotify.jpg?width=93&height=30&name=logo-spotify.jpg)](https://content.vespa.ai/becoming-ai-native-document#case-spotify) - [![DeviantArt](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-deviantart.png?width=102&height=28&name=logo-deviantart.png)](https://content.vespa.ai/becoming-ai-native-document#case-deviantart) - [![Kleinanzeigen](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-kleinanzeigen.png?width=62&height=45&name=logo-kleinanzeigen.png)](https://content.vespa.ai/becoming-ai-native-document#case-kleinanzeigen) - [![Vinted](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-vinted.jpg?width=89&height=31&name=logo-vinted.jpg)](https://content.vespa.ai/becoming-ai-native-document#case-vinted) ### Perplexity — Building AI Search at Web Scale ![Perplexity](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-perplexity.jpg?width=208&height=53&name=logo-perplexity.jpg) [Read the Perplexity case study](https://vespa.ai/case-studies/why-perplexity-chose-vespa-for-ai-search/) Perplexity is an AI-powered answer engine that combines web search with large language models to provide current, cited answers. Before it can generate an answer, Perplexity needs to find the right information from a huge, continuously changing web index, identify the most relevant passages, and rank them well enough to give the LLM high-quality context. Perplexity uses Vespa for retrieval at the scale and speed required by a global consumer application. Vespa enables hybrid retrieval, filtering, passage-level ranking and continuous indexing within a single distributed serving platform. The same retrieval foundation also supports Perplexity's Search API, demonstrating how a sophisticated retrieval layer supports multiple AI experiences as applications evolve. 780M+ queries/month · 200B+ URLs indexed · 200M+ queries/day · 358ms median Search API response time ### RavenPack — AI Research Across Billions of Financial Documents ![RavenPack](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-ravenpack.png?width=295&height=37&name=logo-ravenpack.png) [Read the Vespa/RavenPack announcement](https://vespa.ai/ravenpack-launches-bigdata-com-with-vespa-ai-to-revolutionize-billion-scale-vector-search-for-financial-research/) RavenPack helps financial professionals turn vast amounts of information into actionable intelligence. Its Bigdata.com platform combines RAG with RavenPack's financial knowledge graph, allowing users to research and analyze billions of financial documents through a real-time AI research assistant. In financial markets, where information changes continuously, the challenge is not simply finding relevant content but retrieving it quickly and accurately enough to support timely decisions, while keeping answers traceable to their sources. RavenPack uses Vespa as the search and retrieval platform behind Bigdata.com, supporting vector search across billions of documents. After several years using Vespa Open Source, RavenPack moved to Vespa Cloud to support its enterprise workflows and simplify the infrastructure required as the service scales. Vespa enables Bigdata.com to retrieve and rank relevant financial information at billion-document scale, providing the evidence needed by its RAG applications for real-time research and analysis. Billions of financial documents · Real-time AI research · OSS → Vespa Cloud ### Thomson Reuters: Bringing Legal Search from Research to Production ![Thomson Reuters](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-thomson-reuters.png?width=178&height=62&name=logo-thomson-reuters.png) [Read the Thomson Reuters Labs blog](https://medium.com/tr-labs-ml-engineering-blog/building-legal-search-at-tr-4e4b597d595e) Legal search requires more than matching meaning. Results must reflect the right jurisdiction, legal authority and whether a ruling remains good law. For Thomson Reuters Labs, addressing this complexity also meant finding a simpler way to experiment, improve ranking and bring new search capabilities into production. Using Vespa, the team consolidated chunking, embedding, content enrichment and custom ranking within a single search application. A proof of concept covering more than two million active U.S. statute documents progressed into a production system serving customers, without rebuilding the research in a separate technology stack. **2M+** active documents in initial corpus · Production deployment · Unified retrieval and ranking ### Elicit — Better RAG Starts with Better Search ![Elicit](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-elicit.png?width=196&height=56&name=logo-elicit.png) [Read the Elicit blog](https://elicit.com/blog/search-vs-vector-db) Elicit uses AI to help researchers find and work with scientific literature. Its engineering team argues that effective RAG requires more than putting embeddings into a vector database: if the underlying search can't reliably retrieve the right information, the language model has little chance of producing a good answer. Elicit uses Vespa to combine vector and keyword retrieval as part of that search foundation. Rather than treating vector search as an external memory for an LLM, Elicit approaches retrieval as a search problem — combining different retrieval techniques, structured queries and reranking to improve the quality of the information ultimately available to the AI application. > “If you want to build a RAG-based tool, first build search.” > > — Adrian “Panda” Smith, Infrastructure Engineer, Elicit ### Metal — Retrieval Built for AI Agents ![Metal](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-metal.png?width=197&height=56&name=logo-metal.png) [Read the Metal case study](https://vespa.ai/metal-ai/) Metal provides an institutional intelligence platform for private equity firms, connecting information across deal documents, expert calls, financial data, CRM records and other sources. Rather than treating this information as a collection of documents, Metal models companies, people, activities and financial data — and the relationships between them — so its AI applications can retrieve information in the context of an investment workflow. Metal uses Vespa as the core retrieval layer for an increasingly agent-driven platform. AI agents determine what information they need, query different entity types, apply filters such as recency and business rules, and iteratively retrieve additional information before constructing the context required for an answer. Today, 95% of Metal's retrieval is performed by AI agents, with Vespa supporting the structured, multi-entity retrieval and ranking underneath them. > “95% of our retrieval is done by AI agents.” > > — Sergio Prada, Co-Founder & CTO, Metal ### Yahoo — AI Applications at Internet Scale ![Yahoo](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-yahoo.png?width=192&height=57&name=logo-yahoo.png) [Read the TechCrunch article](https://techcrunch.com/2023/10/04/yahoo-spins-out-vespa-its-search-tech-into-an-independent-company/) Yahoo has used Vespa for more than two decades across search, recommendations, personalization and advertising. What began as search technology evolved into a general-purpose platform for combining large-scale data, machine learning and real-time decisioning across Yahoo's consumer services. Today, Vespa powers around 150 applications across properties including Yahoo Finance, Yahoo News and Yahoo Sports. Collectively, these applications serve approximately one billion users and process around 800,000 queries per second demonstrating how the same retrieval and ranking foundation can support many different AI-powered experiences at sustained internet scale. ~150 applications · ~1 billion users · ~800,000 queries/sec ### Spotify — Helping Listeners Discover Content Beyond Keywords ![Spotify](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-spotify.jpg?width=185&height=59&name=logo-spotify.jpg) [Read the Spotify Engineering blog](https://engineering.atspotify.com/2022/03/introducing-natural-language-search-for-podcast-episodes) Spotify wanted podcast search to understand what listeners meant, rather than requiring words from episode metadata. Traditional term matching, even with fuzzy matching, normalization and aliases, couldn't capture all the ways people express the same idea in natural language. Spotify therefore introduced Natural Language Search, using semantic retrieval to find relevant podcast episodes even when their titles and descriptions don't contain the user's search terms. Spotify uses Vespa to index episode vectors and perform approximate nearest-neighbor retrieval across tens of millions of podcast episodes, with Vespa reranking results using signals such as episode popularity. Rather than replacing its existing retrieval methods, Spotify combines semantic retrieval with other sources before final ranking. The approach produced a significant increase in podcast engagement in A/B testing and was rolled out to most Spotify users. Significant increase in podcast engagement - Following A/B testing of Natural Language Search ### DeviantArt — Personalized Discovery Across a Billion Documents ![DeviantArt](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-deviantart.png?width=201&height=55&name=logo-deviantart.png) [Read the DeviantArt case study](https://vespa.ai/case-studies/wix-case-study/) DeviantArt, a Wix subsidiary, is one of the world's largest online art communities, with more than 100 million registered members and over 650 million works. With tens of thousands of new artworks uploaded every day and highly diverse user interests, discovery depends on continuously retrieving and ranking the right content while adapting to changing content, trends and user behavior. DeviantArt uses Vespa to power personalized search and recommendations, combining machine learning, user context, custom query processing and real-time ranking within a single serving platform. The deployment manages more than one billion documents while delivering extremely low query latency, enabling DeviantArt to provide fast, personalized discovery experiences across its global community. > “Vespa gives us the flexibility to integrate custom ranking models, user context, and domain-specific processing within a single high-performance system.” 1B+ documents · Nearly 4,000 QPS · 12ms average query latency ### Kleinanzeigen — Reinventing Personalized Marketplace Discovery ![Kleinanzeigen](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-kleinanzeigen.png?width=123&height=90&name=logo-kleinanzeigen.png) [Read the Kleinanzeigen tech blog](https://medium.com/berlin-tech-blog/rebuilding-the-kleinanzeigen-homepage-feed-a-deep-dive-into-fashion-796ff379481a) Kleinanzeigen is one of Germany's largest online marketplaces, where discovering relevant listings means understanding multiple, sometimes competing signals of intent. Rebuilding its homepage recommendations, Kleinanzeigen moved from Elasticsearch to Vespa, bringing user profiles, retrieval and ranking together within one serving platform. The new approach combines click history, explicit preferences, recent searches and listing freshness to create more personalized recommendations. In Fashion, Kleinanzeigen found that simply combining these signals during ranking wasn't enough: relevant products first had to make it into the candidate set. By giving different signals their own retrieval paths and then combining and ranking the results, the team was able to better reflect the diversity of individual user interests. A subsequent A/B test produced a “very substantial uplift in listing clicks” compared with the previous approach. Very substantial uplift in Fashion listing clicks in A/B testing ### Vinted — Better Product Discovery, Measurable Business Impact ![Vinted](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/logo-vinted.jpg?width=177&height=62&name=logo-vinted.jpg) [Read the Vinted case study](https://vespa.ai/vinted/) Vinted operates one of the world's largest second-hand marketplaces, with more than 120 million registered users and around one billion searchable items, handling up to 25,000 item searches per second. Search, recommendations and personalization help buyers discover relevant products across a huge, constantly changing inventory. As the marketplace grew, Vinted's Elasticsearch infrastructure became increasingly costly and difficult to scale, prompting the company to move its search and discovery workloads to Vespa. The move delivered improvements across performance, efficiency and relevance. Search latency improved by 2.5×, the infrastructure footprint was cut from 120 to 60 servers, and Vinted was able to increase ranking depth more than threefold. More importantly, better search translated directly into business results: purchased transactions increased by 1.1% and Gross Merchandise Value by 0.6%, representing more than €3.5 million in additional merchandise sold through the platform. +1.1% transactions · +0.6% GMV · €3.5M+ additional merchandise sold 04 / The question ## Summary --- Across AI search, research, recommendations and product discovery, these stories show how organizations use Vespa to turn complex data into relevant experiences. ![White water rushing through a narrow gorge of dark, wet rock](https://content.vespa.ai/hs-fs/hubfs/RGP/vespa-ainative-reader/img/summary-photo.jpg?width=1400&height=801&name=summary-photo.jpg) The challenges vary, but common needs emerge: retrieve the right candidates, rank them for the task, and serve results as data and demand change. Vespa brings these capabilities together, helping teams simplify their systems, move ideas into production and improve the experiences they deliver. From Thomson Reuters’ legal search to Vinted’s measurable business gains, the examples show what better retrieval and ranking can enable. ### What could better retrieval and ranking unlock for you? Explore how the Vespa AI Search Platform can help you build relevant AI experiences with retrieval, ranking and real-time data in one platform. [Talk to sales](https://vespa.ai/contact-sales/?source=Resources%20-%20Becoming%20AI%20Native%20Document&t1_source=Marketing&utm_source=vespa-resources&utm_medium=resource-page&utm_campaign=becoming-ai-native-document&utm_content=body) Tell us what you’re building and the challenges you need to solve. ## About Vespa.ai --- Vespa.ai develops the Vespa AI Search Platform, enabling organizations to build high-performance AI search, recommendations, conversational AI, and AI agents on a single platform. Trusted by companies including Yahoo, Spotify, Perplexity, and AlphaSense, Vespa combines retrieval, ranking, machine learning inference, and real-time serving to deliver high relevance, predictable latency, and massive scalability without the complexity of multiple specialized systems. Want to learn more? We have many different resources and information available through our social platforms [GitHub](https://github.com/vespa-engine)X[LinkedIn](https://www.linkedin.com/company/vespa-ai/posts/?feedView=all)[YouTube](https://www.youtube.com/channel/UCVXw_f6UHff8-V9FA1LMIiw) Read next [What Is Retrieval Engineering?](https://content.vespa.ai/what-is-retrieval-engineering?hsLang=en) © Vespa.ai Norway AI. Oct 2026 9 minute read ```json { "@context" : "https://schema.org", "@type" : "Article", "articleSection" : [ "About this eBook", "Data is (Still) the New Oil", "Retrieval at Work", "Summary" ], "author" : { "@type" : "Organization", "name" : "Vespa.ai", "url" : "https://vespa.ai" }, "datePublished" : "2026-10", "description" : "How organizations are putting their proprietary data to work with AI", "headline" : "Becoming AI Native", "inLanguage" : "en", "publisher" : { "@type" : "Organization", "name" : "Vespa.ai", "url" : "https://vespa.ai" }, "url" : "https://content.vespa.ai/becoming-ai-native-document", "wordCount" : 2433 } ```