Share

Linkedin iconFacebook iconTwitter icon

More deployment options

Most teams building AI agents hit the same problem. Everything isn’t in the same place. The numbers are held in databases and data warehouses. Meanwhile, the real meaning underpinning them is held in contracts, PDFs, and SharePoint. One type of source is the data, the other is the context. An agent that can access only one side gives you half an answer.

For example, take a contract that auto-renews. If you miss the notice window, you are locked in for another term. The renewal date and the spend are in your database. The notice period is in a PDF. Without access to both, the agent can’t give the right answer. It won’t simply fail to answer, it will guess.  

Together, Starburst and NetApp have the answer to this challenge. We have built a way to ask one question and get one answer from both kinds of data, without copying either one into a separate platform. If you are planning your next AI project, it is worth knowing before you start.

This is a problem impacting most enterprises today. 

Does this sound like your next project?

  • You are building or piloting an AI agent, and its answers cover only the database side or only the document side.
  • The questions that matter most, like contract renewals, customer escalations, and compliance reviews, need numbers and documents together.
  • Security or compliance won’t approve copying data into yet another index.
  • Business users ask where an answer came from, and nobody can show them.

If two or more of these fit, the workshop at the end is built for you.

Overview of the approach 

At a high level, the two technologies work together in a specific way. Starburst AIDA plans the question, pulls the numbers through Starburst’s query engine, then pulls the clauses through NetApp AIDE, and joins them. 

The illustration below shows how the handshake works. You can check every figure and clause against its source before you act on it.

Image depicting the high level architectural handshake between Starburst and NetApp.

How the two halves work together

Let’s look at how these two halves work in more detail, and specifically how NetApp handles unstructured data while Starburst handles semi-structured and structured data.  

NetApp for unstructured data

First, NetApp AI Data Engine (AIDE) covers the unstructured side. It does this in several ways: 

  • It turns files into something an agent can search. 
  • It ingests, chunks, embeds, and stores content in a vector index. 
  • It reads contracts, PDFs, SharePoint, NAS, and S3 content where it already sits on NetApp storage, and it exposes the result through an AIDE MCP server. 

Additionally, AIDE also preserves the permissions already set on your files. NetApp AIDE Metadata Services (also called the Metadata Engine) adds to this. It operates as one searchable map of every file and object you own and is the first step for every customer. 

Together, this allows several things to happen. Let’s look at them one by one. 

What NetApp AIDE does

It builds a single searchable index across ONTAP, StorageGRID and third-party NFS, SMB and S3 storage. For each file it records identity, time, ownership, effective permissions, placement and tags. 

How this provides value

It ranks stale, redundant and duplicate data by owner and footprint, so you can see tiering and reclaim options right away. Users can also tag millions of files at once with their own labels, and the tags persist and are searchable. 

How this lowers risk

It reads metadata only, over interfaces you already expose. Nothing is copied or moved, and no GPUs are needed. 

How it this runs in practice

Scale-out scanners read metadata and permissions and resolve owners against the user directory. It runs on a single Linux VM or a multi-node Kubernetes cluster. The catalog is available through NetApp Console, REST API, and MCP for AI agents. 

Who this is designed for 

It targets storage and infrastructure teams. NetApp Professional Services deploys it at no cost, and it works with non-NetApp storage too. It also serves as the foundation for the later preview capabilities, such as Entity Extraction. 

In this setup, Model Context Protocol (MCP) operates as the common interface. The agent doesn’t need to know which system holds what. 

Starburst for structured and semi-structured data 

Turning to Starburst, this architecture leverages both Starburst’s core query engine and Starburst’s AIDA. It covers the structured and semi-structured side. As a federated SQL engine with 50+ connectors, including Oracle, Snowflake, Iceberg, MongoDB, MySQL, PostgreSQL, Hive, and S3 Tables, it queries data in place and applies row-level and column-level access control, masking, and auditing. 

AIDA sits on top of this universal data access, providing conversational analytics support that leverages AI. In this deployment, it plans the question, pulls the numbers through Starburst’s query engine, pulls the clauses through NetApp AIDE, and joins them. 

The image below is an illustrative result. You can check every figure and clause against its source before you act on it.

Image depicting the details of the NetApp and Starburst technology working together.

What AIDA does in this setup

Under the hood, AIDA sits between the people and agents asking (Claude, OpenAI, copilots, BI tools) and the data. 

It works in four steps.

  1. Plan: AIDA splits the question into a structured part (which contracts, how much spend, when) and an unstructured part (what the clauses say).
  2. Federate: the structured part goes to Starburst. The unstructured part goes to AIDE through MCP.
  3. Ground: AIDA joins renewal dates and spend to the matching clauses.
  4. Cite: every figure and clause points back to a source and version.

How this is different from the usual approaches 

Why hasn’t everyone done this before? The Starburst and NetApp approach represents an iteration on standard methods. Traditionally, there are three common ways to solve these kinds of problems, but each one gets stuck. 

Let’s look at how things traditionally work.

Copying everything into one place

Data centralization is often the default answer. For this approach, documents and tables get pulled into a central index so one tool can search them. That means moving data, rebuilding permissions on the copy, and answering from potentially stale data. 

Using our approach, structured data is queried in place using data federation and documents are processed on NetApp storage, so nothing is copied to an external index. At scale, federation and pushdown avoid moving large volumes of objects. 

Pointing an agent at documents, but not data

Another typical problem comes when document searches can find the clause, but it can’t join to spend, so it can’t tell you what is at stake. 

Our approach solves this too. Starburst reaches 50+ sources, and AIDA joins the numbers to the clauses. 

Pointing an agent at production databases only

Often, agents are only able to read production databases. This leads to a situation where SQL can rank vendors by spend, but it can’t read a notice period. 

In contrast, using our approach, AIDE makes contracts, PDFs, SharePoint, and NAS content searchable to the agent through MCP. It handles the document readers, so the agent works with searchable content instead of raw file formats. 

Meanwhile, Starburst already reaches your databases, and if your files sit on NetApp storage, AIDE reads them there. Neither side has to move.  Get started 

Start with one question, not a platform project. Contract renewals, customer escalations, and compliance reviews are good candidates. From there: 

  1. Pick one question that needs both kinds of data. 
  2. Bring it to a joint Starburst and NetApp architecture workshop.
  3. Get one governed, cited answer from your own sources, then expand to more sources and agents. 

    Want to learn more? Read more about Starburst and NetApp.

    Start for Free with Starburst Galaxy

    Try our free trial today and see how you can improve your data performance.
    Start Free