Imagine choosing a site for a clinical trial. The obvious task is to find hospitals and investigators who have worked on something similar. The less obvious task is to work out which investigators succeeded under which protocols, at which institutions, with which patients, and under which local rules. The answer lives in the connections. A search result can hand you a document. It cannot, on its own, explain the web around it.
PSI, a clinical research organization, says its site-selection process once took as long as six weeks. It built an AI-assisted knowledge engine called SYNETIC on Arango's platform, connecting historical projects, institutions, investigators, protocols and outcomes. PSI now says that search can take minutes. The result is a customer-reported outcome, and the savings claim depends on the trial. But the mechanism is plain enough: make relationships available before a researcher asks the question.
- Arango combines graph, vector, document, key-value and search capabilities in one data platform.
- Its 2026 release adds tools to build knowledge graphs and retrieve context for AI agents.
- PSI and HPE Aruba offer concrete examples: faster trial-site research and fewer network-data silos.
- Its commercial offer spans enterprise software, managed cloud and embedded licensing; prices are quoted.
The database had a head start
Arango did not begin with agents. Founders Claudius Weinberger and Frank Celler built a database for an older irritation: applications often need several ways to represent the same world. A customer can be a document, an account balance a key-value pair, and a referral a link between people. Conventional architecture might put each in a different system and ask developers to maintain the joins. ArangoDB's answer was a multimodel engine and one query language, AQL, across those models.
There is a pleasingly odd name in the early history: the software was once called AvocadoDB, and Arango refers to an avocado variety. By 2016, the company had released an enterprise edition and raised €2.2 million in a round led by Target Partners. It expanded from its Cologne roots toward the United States. A $10 million Series A followed in 2019; a $27.8 million Series B arrived in 2021, led by Iris Capital. In 2022, Shekhar Iyer became chief executive, while Weinberger moved into product leadership. This is the long path behind the new AI vocabulary, rather than a database assembled for the latest conference banner.
A match is not a relationship
A vector search system is good at finding text that resembles a question. A graph is good at following explicit links: this user owns that device, this device depends on that service, this service changed after that incident. Full-text search and documents have their own jobs. Arango's competitive argument is that applications frequently need all of them, while the surrounding permissions and operational state have to agree.
The company calls its current offering the Contextual Data Platform. At the bottom sits ArangoDB. An operations layer adds deployment, access controls and observability. On top, its Agentic AI Suite supplies tools such as AutoGraph, which constructs knowledge graphs from enterprise material; AutoRAG and Deep Search, which apply graph, vector or hybrid retrieval; AQLizer, which turns natural-language requests into database queries; and Ada, a developer assistant. Version 4.0 was announced in March 2026, with later release notes recording further updates. Arango says the platform is in more than 200 production environments worldwide.
The point is to preserve useful connections as data changes, then reuse them across questions.
An enterprise could assemble an alternative from Neo4j or TigerGraph for graph work, Pinecone or Weaviate for vectors, MongoDB for documents, and Elasticsearch or OpenSearch for text. Some teams will prefer the freedom of separate specialists. Arango asks a different purchasing question: how much engineering effort goes into keeping those systems synchronized, secured and available when a new application arrives? Its claimed advantage is fewer joins between vendors and a shared query path. That is a product claim, best tested against an organization's actual workload.
Where the joins hurt
HPE Aruba Networking describes another kind of pain. Its network information was split among six databases. The company says it consolidated the data on Arango, retired those six databases and simplified global network operations. The case study also mentions retiring eight legacy systems, a broader count that should not be confused with the database tally. Either way, the telling detail is operational: connected devices and incidents are easier to investigate when their records inhabit a common model.
“We retired six databases and now run global networking on one platform - simpler operations, fewer incidents, and faster answers.”HPE Aruba Engineering, in an Arango customer account
The same logic appears in Arango's semiconductor example. Bug reports can number in the millions; useful clues may hide in dependencies among components and earlier failures. A similar-looking ticket is only a lead. Engineers need to trace the system around it. This is where graph traversal earns its place beside semantic search. Arango's work with NVIDIA on graph analytics and video search fits the same theme, though a partnership or a demonstration should not be mistaken for a universal performance result.
The costs deserve equal attention. Arango does not publish a single price for its enterprise platform. It quotes by deployment type, environment size and features. A free Community Edition gives developers a route in; paid options include enterprise software, a managed platform and OEM licensing for software vendors embedding the database. Buyers therefore have to compare the platform's quote with the engineering and cloud bills for their own collection of tools. A neat diagram cannot settle that calculation.
The test worth copying
There is a practical lesson here for anyone building an AI product. Before choosing a database, write down one question your users cannot answer with a document search. Name the entities involved. Draw the links the answer must cross. Check which source owns each fact, how quickly it changes, and who is permitted to see it. Then build a small retrieval test and inspect the answer's trail. If the result depends on several relationships, Arango's approach becomes relevant. If your workload is only a simple semantic lookup, the case for a whole contextual platform is thinner.
That discipline also clarifies what can fail first. In PSI's account, the blocker was fragmented research history rather than a shortage of trial data. In HPE Aruba's, it was the number of separate operational databases. Arango's product tries to move those connections into a durable layer, rather than remake them for every agent. The company cannot supply clean source data or sound access policy by magic. Those are still the customer's responsibilities.
Arango sits in a crowded market where every database vendor has found an AI story. Its more persuasive story is older: data has always had relationships, and those relationships have always been expensive to lose. The arrival of AI agents has simply made the loss more visible. When an answer needs to explain not only what it found but why the pieces belong together, the graph stops being an illustration and becomes part of the evidence.