PuppyGraph Raises $10M to Let AI Agents Query Warehouse Data as a Graph

PuppyGraph has raised $10 million to expand its enterprise sales operation, and the money follows general availability of version 1.0. The product is more specific than the usual graph database pitch. It doesn't ask you to copy data anywhere. It sits on top of tables that already live in your warehouse or lakehouse and treats them as a single graph. My read is that this is the more interesting bet here, because it targets the place where agent projects tend to stall.

Why the pilot-to-production gap is a graph problem

The company's central example is a question that sounds simple: which critical systems could an attacker reach from one compromised account? Answering it means following how identities, roles and permissions connect across thousands of tables spread over several systems. Each hop in that chain is a join, and the chain can run several relationships deep.

A proof of concept can get away with 30 to 50 tables. Past that, two things break. The first is context. Schema descriptions that once fit comfortably in an agent's window start consuming a meaningful share of it, which leaves less room for reasoning. The second is latency. PuppyGraph says deep multi-hop queries can stall until they time out. Neither problem is unusual for teams that have tried to hand an agent a large relational estate, but they rarely show up in a demo.

PuppyGraph claims its agents can work across 3,000 enterprise tables and return graph query results in under a second. Those are the company's figures, and I'd treat them as a vendor benchmark until someone reproduces them against their own schema. The number that matters more for your planning is how the system behaves as the schema grows and the traversal depth increases, and the announcement doesn't break that down.

How the schema and guardrails work

Setup starts with a schema that an AI-assisted tool drafts from your existing tables. A data team reviews and approves it, and only then do users start asking questions in plain English. That approval step is the important design choice. The graph is defined by people who know what the relationships mean, and the agent is not left to infer how the data fits together.

That constraint shows up in how failures are handled. If an agent refers to a relationship nobody defined, the query is rejected, and the error message is written so the agent can correct itself on the next attempt. Teams building agents against loosely governed data often end up debugging hallucinated joins after the fact. Rejecting undefined relationships at the query layer shifts that failure earlier, where it's cheaper to catch.

PuppyGraph includes its own agent, and it supports the Model Context Protocol, so companies can connect the AI tools they already run. That matters for adoption. Few engineering teams want to standardize on a single assistant, and an open integration point lowers the cost of trying the query engine without retooling everything else.

Access control and credentials

Access is where enterprise deployments usually get hard. When PuppyGraph connects to Snowflake or Databricks, it can request short-lived credentials tied to the individual user's identity. Nobody has to distribute long-lived keys to a service that an agent can reach. Row-level permissions already configured on the underlying data still apply, and PuppyGraph says it checks them at every hop of a traversal, not just at the entry point.

The per-hop check is the detail I'd examine closely in a real evaluation. A permission model that only validates the starting table can leak data several joins away. Checking at each step is more expensive, and whether it holds up under heavy traversal workloads is the sort of thing a proof of concept should measure directly.

Who's using it, and what the company is signaling

PuppyGraph names Advanced Micro Devices and Palo Alto Networks among its customers. The announcement also lists Coinbase, Datadog and Trend Micro. Xinyu Liu, a staff software engineer at Coinbase, described the appeal in practical terms: because the exchange connects straight to tables it already keeps in Snowflake and Databricks, "standing up a new graph is a schema definition, not a data migration." For teams that have already invested in a warehouse, that framing captures most of the value. The alternative is an extract-and-load pipeline into a separate graph store, with its own sync problems and a second copy of sensitive data to govern.

Revenue traction is reported only in relative terms. The company says annual recurring revenue grew sevenfold over the past year. No dollar figure was disclosed, so there's no way to judge scale from this announcement alone.

The round was led by Particle Future and defy.vc, and it's the second time defy.vc has led a round for PuppyGraph, following a $5 million seed in November 2024. The company says part of the new capital will go to engineering. Co-founder and CEO Weimo Liu, a former software engineer at Google, framed the direction plainly: "The insights of enterprise data will soon be generated by agents." He also argued that users will expect to ask questions that cross departments and systems, beyond what someone hard-coded into a dashboard.

What to watch

Amy Yin, a partner at defy.vc, made a point that's worth keeping in view. Writing the query is now something a user can hand to a model such as Anthropic's Claude. Running it across billions of relationships, in her words, still "takes serious engineering." That gap between generating a query and executing it reliably at scale is where graph-over-warehouse products will be tested.

For evaluation, a few questions matter more than the headline benchmark. How does latency behave as traversal depth increases on your own schema? Does the per-hop permission check hold when many agents run concurrently? How much of the schema review burden falls on your data team as table counts grow? And how often do agents hit rejected relationships, and does the error feedback actually get them to a correct query?

Developers can try the free version with a single Docker command and no sign-up form. The company says it takes about 10 minutes to move from deployment to a first query. That's a low-cost way to test the approach against a real warehouse before committing to anything larger. Whether graph queries become a standard layer under enterprise agents will depend less on the funding than on those numbers holding up in production.