How LanceDB Supercharged Our Knowledge Graph

Quick context, LanceDB just published a deep dive on how Dosu moved from pgvector to LanceDB to keep pace with thousands of GitHub events and an ever-growing body of code. If you’re short on time, here’s the essence and why it matters.

One stubborn bottleneck

Our first production search stack ran on pgvector. It was fine until every model tweak forced a migration script. Development velocity ground to a halt, and query performance at scale was shaky.

Why LanceDB clicked

• File-based, local-first: point Dosu at a directory and start prototyping with zero migrations.
• Hybrid vector + text search out of the box, tuned for code semantics and developer intent.
• Time-travel versioning allows our agents to reason over any historical state of a repository.

Impact so far

Within weeks, we swapped in LanceDB and scaled to thousands of customers. Engineers spend time shipping features, not babysitting migrations. Teams using Dosu now enjoy millisecond search across millions of vectors and 80% automated issue labeling in days.

Explore the new LanceDB's Public Space

We’ve made LanceDB one of our first Public Spaces, a free, read-only knowledge base where anyone can chat with Dosu and get answers pulled from the project’s code, commits, discussions, and docs.

No setup, no log-in, just ask away and let Dosu fetch context for you. To try it, head over to LanceDB’s Public Space. For the full story behind Dosu Public Spaces and how they fit into our mission of making engineering knowledge accessible to everyone, check out July’s Dosu Drop announcement.

Looking ahead

Versioning and richer domain hooks are on the roadmap, turning “docs drift” into a solved problem and letting AI agents pull accurate context on demand.

Read the complete case study. LanceDB’s write-up provides an in-depth look at the architecture, benchmarks, and lessons learned. If you’re evaluating vector databases or wrestling with migration fatigue, grab a coffee and give it a read.

Do you have questions about how we implemented LanceDB or would like to see it live in Dosu? Ping me and let’s chat.