In the 1st quarter of 2026, something changed at the enterprise RAG. VB Pulse data covering January through March tells a consistent story: the market has stopped adding layers of search and has begun to adjust the ones it already has. Call it search reengineering.
The survey covered three consecutive monthly waves from organizations with 100 or more employees, with 45-58 qualified respondents per month on platform adoption, buyer intent, architecture outlook and evaluation criteria. The information should be taken as a guide.
Enterprises’ intent to implement a hybrid search system tripled in one quarter, from 10.3% to 33.3%, as 22% of qualified enterprise respondents said they had no production RAG systems at all. For data engineers and enterprise architects building agency AI infrastructure, the data reveals a market in active transition: the RAG architecture most enterprises building at scale is not one they expect to be running by the end of the year.
Hybrid search has become a consensus enterprise strategy. Unlike single-method RAG pipelines based solely on vector similarity, hybrid search combines dense inputs with sparse keyword search and reordering layers, trading simplicity for search accuracy and access control required by production agent workloads.
The independent vector database category is under pressure. Weaviate, Milvus, Pinecone and Qdrant lost adoption share in VB Pulse data each quarter. Individual stacks and provider local lookups absorb their transferred share.
A growing minority of firms are retreating from RAG entirely—a sign that the market’s maturity story has meaningful exceptions.
In 2025, RAG-wide organizations are hitting the same point of failure: the architecture built for document retrieval doesn’t scale to agent scale.
Businesses that rapidly expanded RAG are now paying to rebuild it
The two biggest intent movements in the first quarter are directly related—enterprises facing search quality issues at scale and hybrid search emerging as the consensus answer.
Investment priorities have changed in parallel. The assessment and compliance test was 32.8% of the budget intention in January and fell to 15.6% by March. Search optimization moved in the opposite direction, from 19.0% to 28.9% – surpassing valuation as the highest growth investment area for the first time.
In an interview with VentureBeat in March, Stephen Dickens, vice president and head of practice at HyperFRAME Research, described the operational burden facing enterprise data teams. Oracle’s agent AI data stack. "Data teams are exhausted from fragmentation fatigue," Dickens said. "Managing a separate vector store, graph database, and relational system to power just one agent is a DevOps nightmare."
This fatigue is directly reflected in the platform data. The increase in custom stack to 35.6% is not a rejection of managed search – many organizations manage both. This is the consolidation response of engineering teams that have reached the limit of stacking multiple components.
Not every enterprise has gone that far. VB Pulse data includes a signal that complicates the market’s overall growth story: 22.2% of qualified respondents reported no RAG production by March, compared to 8.6% in January. The report refers to organizations with this cohort "have not yet committed to any search infrastructure or discontinued programs" — Concentrated in Health, Education and Government, showing the highest proportions of fixed budgets in the same sectors.
Independent vector databases lose the acceptance argument but gain the validity argument
Recent reports from VentureBeat show why a custom search layer is still important in production.
Two enterprises built on Qdrant show why purpose-built vector infrastructure is still winning in production.
&AI builds patent litigation infrastructure and performs semantic searches on hundreds of millions of documents. It is optional that each conclusion be based on a real source document – patent attorneys will not act on AI-generated text. This requirement makes the choice of architecture clear.
"An agent is an interface," AI founder and CTO Herbie Turner told VentureBeat in March. "A vector database is ground truth."
GlassDollar, a startup that helps Siemens and Mahle value startups, implements an agent search model on a corpus approaching 10 million indexed documents. A single user sends fans to multiple parallel polls, each getting candidates from a different angle before the results are combined and re-ranked. It is this query volume and accuracy requirement that led to the selection of a purpose-built vector infrastructure.
"We measure success by remembering," Kamen Kanev, head of product at GlassDollar, told VentureBeat in March. "If the top companies aren’t in the results, nothing else matters. The user loses trust."
VB Pulse data shows that framing—search as a ground truth, not a feature—is gaining traction in the broader enterprise market, even as stand-alone vector database adoption declines.
Businesses say why they need a custom vector layer changed significantly during the first quarter. In January, the main reasons were access control complexity (20.7%) and search accuracy (19.0%). By March, operating reliability had risen to 31.1% at scale – more than doubling, surpassing everything else. Enterprises no longer maintain vector infrastructure primarily for accuracy. They keep it because it’s part of the stack they can rely on when scaling query volumes.
How businesses are redefining what good search means
How businesses judge search engines changed significantly during the first quarter – and the direction of this change points to a more sophisticated market about what good search actually means.
In January, answer accuracy dominated the evaluation criteria at 67.2% – more than anything else. By March, response accuracy (53.3%), search accuracy (53.3%), and response relevance (53.3%) had all converged. If the error came from the document or missed the context of the question, it is no longer enough to get the correct answer.
Response relevance was the only criterion to rise during the quarter, gaining five percentage points. It’s also the hardest to measure—whether the retrieved context is actually the right context for that particular question requires a purpose-built evaluation infrastructure, not just pass or fail checks. Its rise indicates that a significant number of enterprise buyers completely pass the basic RAG test.
Market verdict: RAG is not dead. It is an original architecture
The "RAG is dead" the narrative had real momentum in 2026. It was based on two claims. First: long context windows—models that can process hundreds of thousands of tokens in a single query—will make custom search redundant. Second: instead of retrieving what the agent learns anew each time, agent memory systems that store between sessions will completely master the knowledge access problem.
VB Pulse data is the enterprise market’s answer to the first claim. The long-term context in the dominant architectural position fell from 15.5% in January to 3.5% in February before partially recovering to 6.7% in March. The January pattern was heavily weighted for respondents in Technology and Software – the segment most exposed to long context model announcements in late 2025. As the pattern diversified, the position evaporated.
Jonathan Frankle, Chief AI Scientist at Databricks, clearly described the architecture in the memory question. March interview Via VentureBeat: A vector database with millions of entries sits at the base of the agent memory heap, too large to fit into context. The LLM context window is located at the top. Between them, new caching and compression layers appear, but none of them replace the search layer in the base. New agent memory systems such as Hindsight developed by Vectorize and observational memory approaches such as those in the Mastra framework address session persistence and agent context over time—a different challenge than high-recall searching among millions of changing enterprise documents.
The most powerful signal: the share of respondents who do not expect large-scale RAG deployments by the end of the year increased from 3.4% to 15.6% – almost 5 times. This is not a judgment against refund. This is a judgment call against the search architecture that most enterprises built in the first place.
Rebuilding the search is optional
Refactoring search is the cost of scaling a RAG without first deciding which architecture can support it.
If your organization is one of the 43.1% planning to expand RAG to more workflows, VB Pulse data shows that the plan has already changed for many of your peers and may need to change for you. Hybrid search is the goal of consensus. Custom stack growth of 35.6% reflects teams building search infrastructure around requirements that off-the-shelf products don’t fully address.
RAG is not dead. It is the architecture that most enterprises use to implement it. The data suggests that restructuring is not a future decision. For 33% of enterprises, restructuring is already a declared priority.
