# Request control

> Limits used by the shared service.

Canonical: https://keta1930.github.io/icml2026-mcp/en/docs/concepts/governance



The server uses persistent per-IP and global request limits, a bounded search queue, separate worker limits for search and SQLite, and semaphores for DashScope and Milvus.

Before a DashScope call, SQLite atomically reserves one request and an upper bound on input tokens for the current UTC minute. The reservation is settled with actual usage after the response. Defaults match the model limits: 24,000 RPM and 1,000,000 TPM.

Successful query embeddings are cached. Concurrent identical misses share one provider call.