Skip to the cards
Filed
Source
James Smith
Loans
501

PuppyGraph 1.0 Adds AI Schema Builder and MCP Support for Enterprise Data

PuppyGraph 1.0 maps tables into graph schemas and connects MCP compatible agents, with query validation and distributed execution. Costs and model upkeep still need review.

PuppyGraph 1.0 Adds AI Schema Builder and MCP Support for Enterprise Data

PuppyGraph 1.0 adds an AI assisted schema builder and Model Context Protocol (MCP) integrations to its distributed graph query engine. For IT teams connecting AI agents to data spread across databases, warehouses and lakehouses, the change offers a way to describe relationships once and make them available to compatible agents without copying the data into a separate graph store.

The practical gain depends on how well the generated model reflects the business and how much compute graph queries consume on the underlying platforms. Schema generation can reduce initial engineering work, but it doesn’t remove the need to review relationships, govern the model and monitor costs as workloads grow.

From tables to a graph model

The schema builder maps existing tables into graph entities, relationships and properties. Developers can inspect and approve the proposed schema, rather than defining every connection manually. In the graph model, entities become nodes, relationships become edges, and attributes become properties.

That structure gives agents a map of how information across tables and systems connects. Without it, an agent may need to identify relevant tables and generate complex SQL joins for each question. A predefined graph model can simplify query generation, and may reduce errors that cause agents to retry.

PuppyGraph also checks agent generated queries against the graph schema. If a query refers to a relationship that hasn’t been defined, the engine rejects it and returns feedback the agent can use to revise its query. Ashish Chaturvedi, executive research leader at HFS Research, described this as an enforced ontology: agents are kept within the entities and relationships established in the model. That constraint can reduce faulty joins and unreliable results, though it can’t correct inaccurate source data or a poorly designed schema.

For queries that span several relationships, PuppyGraph’s execution engine breaks the work into smaller operations that run in parallel across compute nodes. The company says this can speed up traversal of large datasets. Actual performance will depend on the data platform, workload and available compute resources.

MCP connects compatible agents

PuppyGraph 1.0 adds MCP integrations so agents and tools that support the protocol can call the engine’s graph query capabilities as part of their workflows. The stated aim is to avoid building a separate custom connection for each supported agent. The engine also continues to support conventional SQL queries.

The MCP addition builds on PuppyGraph’s existing approach: query relational databases, data warehouses and lakehouses as a unified graph without moving or duplicating data into a dedicated graph database. The engine maps tables into a Labeled Property Graph model, then runs queries across the underlying sources through its distributed compute cluster.

That design may suit organizations trying to make connected data available to agents while keeping it in existing data environments. Devin Pratt, research director at IDC, said developers can model entities and relationships once, then use graph languages for connected questions instead of maintaining long chains of joins. He said that multi step questions, such as tracing connections through accounts, are more natural to express as graph queries.

Potential savings depend on workload

Pratt said a predefined graph model could reduce the schema information agents need to process and the number of retries required to form a query. Avoiding a separate graph store can also sidestep duplicate storage and the pipelines needed to keep copied data synchronized. Those advantages may help with data management costs, but they don’t guarantee lower total spending: the engine still uses compute on the platforms where the data resides.

Pratt also pointed to IDC research showing that 96% of organizations are actively pursuing data management consolidation and that most manage more than 10 database technologies. He said the cost of preparing, storing and serving data is among the barriers organizations cite when scaling generative and agentic AI.

For CIOs, the key question is where PuppyGraph’s query model fits. Ashish Chaturvedi cautioned that graph intensive workloads can increase compute consumption in warehouses and lakehouses, so teams should estimate those costs before expanding deployment. The engine is aimed primarily at analytical queries. Applications requiring low latency graph writes or frequent transactional updates may be better served by a native graph database.

Model quality and supplier capacity matter

Automatic schema generation can shorten the setup process, but the graph model still needs human oversight. Michael Leone, principal analyst at Moor Strategy and Insights, said developers must decide which tables represent entities and which keys define relationships. The engineering work shifts from writing and maintaining complex joins to designing and governing that model.

Source quality is another dependency. Pratt said query quality will reflect the consistency of the underlying schemas and the reliability of the source data. Teams should check whether tables use consistent identifiers and whether the proposed relationships match how the business actually works before exposing the model to agents.

Organizations should also weigh PuppyGraph’s maturity and support capacity before making it a critical part of data infrastructure, Chaturvedi said. The company says its technology is in use at AMD, Palo Alto Networks, Datadog, Netskope, Coinbase, Trend Micro, Baker Hughes, Sola Security, HighStreet Insurance Partners and Prevalent AI.

PuppyGraph lists support for Apache Iceberg, Databricks, Snowflake, ClickHouse, DuckDB, PostgreSQL, Microsoft Fabric and Apache Hudi, among other platforms. Its installation documentation describes deployment through Docker or Kubernetes, as well as cloud marketplaces operated by AWS, Microsoft Azure and Google Cloud. For administrators assessing the release, the next steps are to validate the generated graph model, test representative analytical queries on the intended data platforms, and estimate compute use before expanding access to agents.

Source: James Smith · www.infoworld.com

Filed alongside