So, my co-founder and I are building Relay (https://basegraph.app/) to bring in someone who understands your codebase deeply and asks the right questions to extract the right context out of your head.
- Planning = human judgement: We strongly believe keeping humans in the loop, especially during planning stages. - Auto-routing questions: Relay doesn't just ask questions, it figures out who to ask by analyzing code ownership and ticket history. Product questions go to PMs, architectural decisions to tech leads, implementation details to the assigned dev.
- No vector search: We tried standard RAG for Go, but it wasn't accurate enough. Ex: "Adding a new method to X interface, what might be affected?" led to missing implementors. We built a code graph engine for deterministic search, not semantic similarity.
- Output: A clear technical spec covering decisions, edge cases, and affected services. You can code from it or hand it to your coding agent.
Example:
Ticket: "Add Twilio support"
Relay analyzes and asks:
- @pm: Calls, SMS? What's actually needed?
- @pm: Outbound, Inbound, or both?
- @dev_lead We don't support rate limiting yet. How should we proceed? (Relay knows to tag the dev lead because it found rate limiting gaps in NotificationService via code graph analysis)
Output: a clear technical spec covering decisions, pitfalls, edge cases, and affected services.
This works great when people answer clearly. It breaks when teams ignore questions or when ownership is unclear, that’s an open problem we’re still figuring out.
Current status: - Language Support: We are Go devs, so we built for it first. TS and Python support coming in two weeks.
- Tuning: We're still trying to find the balance between being helpful and noise. Pinging people can get annoying, been there :)
- Privacy concerns: Relay's code graph engine runs in the cloud. We know this is a blocker for some. Self-hosted options coming in the near future.
- Integrations: Relay runs in Linear and Github. Jira and Gitlab coming soon. Also heard a lot of good things about spec-kit.
If this sounds useful, awesome!
I'm also interested if you think this approach won't work, adds friction or misses the real problem. Happy to hear any feedback :)