AWS’s Brad Bebee is big on graph databases. He pretty much invented one: Blazegraph, which Amazon bought in 2018. Bebee stuck around after the acquisition, and now owns the hyperscaler’s managed graph database service Neptune, as well as its timeseries database and in-memory caching services. He’s kept a passion for open-source too: stalking the halls of AWS re:Invent, Bebee can be seen rubbing friendly shoulders with CEOs of firms that, years earlier, would have been fierce rivals and which are now close partners – many of whom are quick to praise his New York-based team for their collegiate approach to collaborating on features and go-to-market.

(InfluxData CEO Evan Kaplan is a case in point, telling The Stack that it’s “a very personal and ‘say - do’ thing to build a relationship of trust… Brad and his team have very consistently done exactly what they said they’d do.”)

Amazon Neptune: An AI accuracy game-changer?

Bebee’s depth of expertise has never been more important: Amazon Neptune, AWS’s managed graph database service, has seen growing adoption by firms focused on improving AI accuracy and tracking relationships between entities – two areas where graph databases excel. (Customers like cloud security firm Wiz rely on Amazon Neptune; Trend Micro is also among the organizations using Neptune – by doing so it boosted the accuracy of a security chatbot from 70% to 90%, Bebee says.)

As he puts it: “The thing that makes graphs special compared to other databases is in a graph, the data model allows you to explicitly express relationships, so I don't have to do foreign keys, I don't have to do other techniques in the database; I can just simply store in the database the fact that we're related, and so that allows us to create applications where we query, store, run algorithms over those relationships between the data.”

Improving AI accuracy with GraphRAG

A key area of interest for many is graph’s use in AWS’s managed “GraphRAG” service – which takes the heavy lifting out of the “context engineering” organizations are doing as they work to improve AI accuracy. 

“Amazon Bedrock Knowledge Bases”, for example, builds on Neptune to manage the creation and maintenance of chunking strategies or complex RAG integrations with LLMs and vector stores in a way that combines graph analytics and vector search to enhance the accuracy and explainability of AI responses via “multi-hop connections between related entities or topics.”

As Bebee explains: “If you think about vector similarity search as compared to graph search, in a vector, similarity is defined by mathematical distance in a really high dimensional space, and it's super cool; it can find things that you wouldn't expect, but it's really opaque. So if you do a vector search it's hard for you to explain, as a developer, why is this result higher than that?

“But in a graph search, you can say ‘this is the most relevant result because this is related to this’. You can take the speed and the ability to look at across different domains and across different languages that you can get from vector search, and when you combine it with a graph that you've automatically created of your data, you can do a hybrid search, where you can explain findings using the graph, and that can be very powerful…”

"You don't have to know anything about a graph"

As Bebee explains to The Stack “You create a knowledge base using Amazon Bedrock, and you simply give it S3 location with documents; it does the chunking and builds the retrieval augmented generation workflow. 

“It also creates a graph for you underneath – so you don't have to know anything about a graph – it's looking at documents, chunks of documents, topics and entities that might occur in them, as extracted by LMS or foundation models, and then it uses that graph automatically for you in the re-ranking process. So when you do a vector search, you get a top-k result, and based on the graph, it will effectively boost other things up or down. 

“Sometimes the most relevant answer might not be in the top-k results: suppose you have a multi part question, or you're trying to answer something like ‘the impact of military action on the Suez Canal on the toy supply chain for the UK Christmas market in 2026!’ Those kinds of questions need you to aggregate lots of different data sources and the graph can help you boost some of those other more relevant things to the top…” 

To make AI application development easier, Neptune offers fully managed GraphRAG with Amazon Bedrock Knowledge Bases, and integrations with Strands AI Agents SDK and popular agentic memory tools. It also easily analyzes tens of billions of relationships across structured and unstructured data within seconds delivering strategic insights. Learn more here.

Delivered in partnership with AWS.

The link has been copied!