Multi-Agent RAG
A retrieval-augmented generation setup where multiple cooperating agents share a retrieval layer.
The problem
A single retrieval pipeline struggles when different sub-tasks need different context; coordinating multiple agents over one shared retrieval layer is the design problem this explores.
My role
Prototype — design & build
Multi-Agent RAG
A prototype where several cooperating agents share one retrieval layer, each pulling the context its own sub-task needs. It explores the coordination problem: a single retrieval pipeline serves no sub-task well when the sub-tasks want different context.
Key decisions
One shared retrieval layer, several consumers. Rather than giving every agent its own index, they share one and each asks for the context its sub-task needs. That kept the coordination problem visible instead of hiding it in duplicated pipelines.
The lesson that carried into Alex OS: grounding is a per-task decision, not a global setting. Different questions deserve different context, and the system should know which is which.
Outcome
Working prototype that demonstrated multiple agents coordinating over shared retrieval; deliberately kept at prototype stage. The retrieval lessons carried directly into how I build grounded systems like Alex OS.
Limitations
Prototype-stage: agent coordination was hand-tuned rather than learned, and evaluation was manual.
Stack
Next
A production version would need routing rules for which agent retrieves what, plus systematic evaluation of answer quality.