Milvus Configuration for NVIDIA RAG Blueprint

Why it matters

GPU acceleration is two separate decisions in Milvus: where the index is built and where queries run. This guide makes you set each one, and then puts access control on the same collections.

What to practise

Move Milvus between the -gpu and CPU images, and set APP_VECTORSTORE_ENABLEGPUINDEX and APP_VECTORSTORE_ENABLEGPUSEARCH for GPU, CPU, or GPU indexing with CPU search. In the mixed mode, note that the client sets adapt_for_cpu and that search requests must then supply ef. With security.authorizationEnabled: true, pass reader and writer credentials in the Authorization header and watch a user without Search or Insert privileges fail. Note what recovery from a GPU_CAGRA error after a switch to CPU mode costs: deleting the Milvus volumes and all vector data.

How it fits

The image and flags put 6.2.1’s CAGRA index and 6.2.2’s GPU execution into practice; compare the mixed mode against 6.2.2’s recall and latency target. The reader and writer tokens are one form of 6.2.3’s tenant isolation.