For the last year, we have been building Evokoa as an operational intelligence layer for multi-location enterprises. We quickly realized that answering complex, multi-hop relationship questions at scale required a graph traversal engine that didn't exist yet. Everything off-the-shelf was either too slow for our agent workloads or required moving our entire system of record into a separate, heavy graph database.
So, we built our own.
Today, we are open-sourcing that engine. We call it pgGraph. It is a high-performance graph traversal extension for PostgreSQL, written in Rust. You can find the repository on GitHub and read the full documentation at pggraph.com.
Graph Superpowers, Postgres Gravity
The core premise of pgGraph is simple: you should not have to migrate your data to get graph capabilities. Postgres remains your single source of truth. Your applications keep writing data the exact same way they always have.
Behind the scenes, pgGraph builds a compact, derived graph index (using compressed sparse row arrays) over your selected relational tables. When your AI agents or applications need to find shortest paths, map relationships, or discover hidden networks, they ask standard SQL questions. Our engine intercepts those queries and walks the integer arrays in microseconds, entirely avoiding the recursive SQL traps that normally bring Postgres to a halt at depth.
Why We Are Open-Sourcing It
We believe that graph capabilities are going to become a fundamental requirement for the agent era. Agents need connected, structural context to reason effectively. If every startup and enterprise has to adopt a separate, heavy graph database just to give their agents context, the market will stall.
By dropping a fast relationship cache directly beside Postgres, we change the unit of adoption. We are building the missing infrastructure for widespread AI adoption, and we are committed to keeping pgGraph 100% free and open-source, forever.
Early Access & Alpha
pgGraph is currently alpha software. It is ready for experimentation, demos, and benchmarks. The core traversal features—like bounded BFS/DFS, shortest path, and connected components—are usable today, though we are still hardening the memory model for extremely large, enterprise-scale graphs.
If you are building AI agents and hitting the execution bottleneck of relational databases, we encourage you to try it out. Spin it up in a Docker container, point it at your schema, and see what microseconds feel like. Star the repo to follow our progress, and let us know what you build.