RadixView: a live radix-tree view of SGLang KV-cache events

[ClaudeLiang]

Hey everyone,

We are working on an agent that already has active users. While reviewing its cache behavior, aggregate metrics such as hit rate were not enough—we wanted to see which prompts and tool definitions were actually shared, and where sessions started to branch.

That led us to build RadixView. It subscribes to SGLang's ZMQ KV-event stream and reconstructs a continuously updated radix-tree view of prefix sharing and eviction.

On our own traffic, the tree exposed branches much earlier than we expected. Following those branches back to their text helped us find several problems in the agent and tool design that were reducing prefix reuse.

Shared prefixes stay on the same path, sessions branch where their text differs, and the view updates as store and eviction events arrive. You can open any node to inspect its decoded text or search across the observed cache blocks.

RadixView is only a viewer. It does not proxy requests or read the cache in-process, and it can run on a different machine from the GPU server.

Install:

pip install radixview

_{Note: The synthetic demo is not included in the pip package. Clone the repository and run python -m examples.demo to try it without a GPU or model.}

Source and setup instructions:

https://github.com/ClaudeLiang/RadixView

Feedback is welcome, especially from people working on agent or other prefix-heavy workloads.