Text-to-SQL agents look great in demos with ten tables. Point them at a real enterprise with thousands of tables and they start guessing how things connect, and guessing wrong. PuppyGraph thinks a graph is the fix. On November 10 the company announced a $10 million Series A led by Particle Future and defy.vc, bringing total funding to $15 million, following general availability of PuppyGraph 1.0.
Graph queries without a graph database
PuppyGraph is a graph query engine that runs directly on data where it already lives: Apache Iceberg, Databricks, Snowflake, ClickHouse, DuckDB, PostgreSQL, Microsoft Fabric, Apache Hudi and others. You define entities and relationships once as a schema. There’s no copying data into a separate graph store and no pipeline to keep in sync.
For AI agents that schema works as an enforced ontology. An agent asks which products depend on a given supplier, and the engine follows the defined relationships across tables. If the agent refers to a relationship that doesn’t exist, the query is rejected with feedback instead of returning confident nonsense. Agents connect through MCP or the built-in assistant.
Performance is the other selling point. The company says it completed a 10-hop query across billions of edges in 2.26 seconds on a four-machine cluster. Version 1.0 also enforces row-level permissions at every step of a traversal, which matters when an agent shouldn’t see everything a database holds.
Who’s using it
The customer list is long for a company this size: AMD, Palo Alto Networks, Datadog, Netskope, Coinbase, Trend Micro and Baker Hughes among them. Security firms are a natural fit, since attack-path analysis is a graph problem sitting on top of huge tables of logs and permissions.
The caveat analysts raise is that complexity moves into designing and governing the graph model, and heavy graph workloads still cost money on the underlying warehouse. But for agent builders the trade looks good. Defining relationships once beats hoping a model infers them from column names.