PuppyGraph Raised 10 Million Dollars to Let AI Agents Query Your Warehouse as a Graph
Data Engineering

PuppyGraph Raised 10 Million Dollars to Let AI Agents Query Your Warehouse as a Graph

PuppyGraph's new Series A bets that enterprises want graph queries across thousands of warehouse tables without migrating any data, a pitch aimed squarely at AI agents that choke on narrow proof-of-concept schemas.

PublishedNovember 18, 2026
Read time5 min read
Share

The funding and what it buys

PuppyGraph, a Santa Clara startup, raised a 10 million dollar Series A led by Particle Future and defy.vc, SiliconANGLE reported on November 10. The round follows the general release of PuppyGraph 1.0 and brings the company's total funding to 15 million dollars, including a 5 million dollar seed round from November 2024 that defy.vc also led at the time. The company says most of the new capital goes toward enterprise go-to-market efforts, with the remainder directed toward continued engineering work on the core query engine itself.

The product is a graph query engine that sits on top of data a company already has in a warehouse or lakehouse, translating it into a graph representation on the fly rather than requiring a separate graph database and the migration project that would normally come with standing one up. For teams already running Snowflake or Databricks, that is essentially the entire pitch distilled into one sentence: get graph-shaped answers to graph-shaped questions without taking on a new system to operate, secure, and keep continuously in sync with the source of truth.

The problem it says it solves

PuppyGraph's framing of the market is specific and worth sitting with rather than skimming past. Proofs of concept for agentic data access typically cover somewhere between 30 and 50 tables, and at that scale they tend to work fine, which is exactly why so many pilots look promising in a demo setting. The trouble starts once a project moves toward production, where schema descriptions for thousands of tables crowd out an agent's available context window, and multi-hop queries that trace relationships across many tables at once start timing out or returning incomplete, unreliable answers.

That is a real and genuinely underappreciated constraint on moving agentic pilots into production, and it is separate entirely from whatever the underlying language model can or cannot reason about on its own. An agent that performs well against a tidy demo schema is not strong evidence that the same agent will perform well once it has to navigate the actual scale and messiness of enterprise metadata, and PuppyGraph is explicitly selling into that specific gap rather than into the model layer itself, where most of the current competitive attention is concentrated.

The claims, and the caveats that come with them

The company says agents running on its platform can work across 3,000 enterprise tables and return graph query results in under a second, a figure meant to contrast sharply with the 30-to-50-table ceiling it describes for typical pilots elsewhere. PuppyGraph 1.0 adds AI-drafted schemas that a team reviews and approves rather than writing entirely by hand, plain-English querying, automatic rejection of relationships the schema does not explicitly define, and permission checks enforced at each hop of a traversal. A free Docker install lowers the barrier for a team to simply try it before committing to anything larger.

None of the performance figures have been independently verified by a third party, and the company has not disclosed the dollar revenue behind its sevenfold annual recurring revenue growth claim over the past year. Named customers, including AMD, Palo Alto Networks, Coinbase, Datadog, and Trend Micro, lend the claim meaningful credibility on their own, but a CTO evaluating this seriously should still ask for a proof of concept run directly against their own schema complexity, not against the vendor's own curated benchmark environment built to showcase best-case performance.

Why the permissioning detail matters most

The feature likely to matter most in day-to-day practice is the permission check enforced at every single hop of a graph traversal, more than the headline speed claim that tends to grab attention first. Coinbase staff engineer Xinyu Liu described standing up a new graph on PuppyGraph as "a schema definition, not a data migration," language that captures why security teams in particular have taken interest: the use case the company highlights most is tracing which systems an attacker could reach starting from a single compromised account, a question that depends entirely on relationships being queried correctly and access being checked rigorously at every single step along the way.

defy.vc partner Amy Yin made a related point about where the genuinely hard engineering work actually sits in a system like this. Writing a graph query is now something a user can reasonably hand off to an AI tool, she noted, but executing that query correctly across billions of relationships, with permissions respected consistently throughout the entire traversal, still "takes serious engineering" underneath the friendly interface. That distinction is the part of the pitch hardest to fake convincingly, and it is the part most worth testing directly rather than taking on faith from a sales deck.

The decision in front of the reader

If your organization is already running agentic pilots against Snowflake or Databricks and hitting the exact wall PuppyGraph describes, agents that work fine in a demo but stall once they hit your real table count, this is worth a scoped evaluation now, run specifically against your own schema rather than a vendor sandbox tuned for a good first impression. CEO Weimo Liu's line, "the insights of enterprise data will soon be generated by agents," is unmistakably marketing, but the underlying problem he is describing, context window limits and multi-hop query timeouts at real production scale, is one most teams running serious agent pilots will eventually run into regardless of which vendor ultimately solves it for them.

If your organization is earlier in its agentic rollout and still operating comfortably at the 30-to-50-table proof-of-concept stage, treat PuppyGraph's existence mainly as a preview of the next obstacle rather than as an immediate purchasing decision to make this quarter. Which specific vendor you eventually choose to solve this with matters considerably less right now than building your current pilots with enough awareness of table-scale limits from the start that the eventual transition to production does not quietly require starting the whole project over from scratch.

Tagged#news#data#data-engineering#databases#analytics#lakehouse#streaming#puppygraph#graph-database#series-a-funding#ai-agents#enterprise-data-access#lakehouse-tooling