Exploring contributions in docs-agent & Katib guidance on next steps

12 views

Skip to first unread message

Ayush Kathil

unread,

Apr 19, 2026, 8:52:48 PM (5 days ago) Apr 19

to kubeflow-discuss

Hi everyone,

I’m Ayush, a student currently exploring Kubeflow and contributing as part of my GSoC preparation.

Over the past few days, I’ve started working on a couple of areas to better understand the system end-to-end.

In docs-agent, I opened a PR focused on improving the retrieval pipeline by:

avoiding repeated SentenceTransformer initialization to reduce per-request latency
cleaning up duplicate Milvus search logic
experimenting with a lightweight reranking step to improve relevance of retrieved context

In Katib, I’ve also been working on a PR to better understand how hyperparameter tuning components are structured and how they interact with training workflows.

Through this, I’m trying to build a clearer picture of how different parts of Kubeflow retrieval (docs-agent), training, and tuning (Katib),fit together in practice.

I’d appreciate some guidance on:

whether it makes sense to continue going deeper into docs-agent (retrieval/RAG side), or shift more toward Trainer/Katib-related components
areas where contributors are currently most needed
any gaps where small but meaningful contributions would be helpful

GitHub: github.com/Ayush-kathil

Thanks, I have been finding the discussions here really helpful for understanding the design decisions.