Qdrant and Minima Deliver 2.92x More Agentic RAG Tasks per GPU-Hour - Qdrant

This post details a benchmark demonstrating a 2.92x increase in successful agentic RAG tasks per GPU-hour by combining Qdrant's hybrid search, payload filtering, and reranking with Minima's LLM weight and KV cache compression. The joint solution reduced median task latency from 21.3s to 7.7s and cut GPU cost per 1,000 successful tasks by 65%. Key technical contributions include: Qdrant's hybrid search (dense + BM25 + RRF fusion) improving first-pass evidence sufficiency from 72% to 87% and reducing context tokens by 56%; Minima's NVFP4 W4A4 weight compression and FP8 KV cache reducing model size and increasing resident sessions by 8.7x; and the combined effect leading to 3,158 successful tasks/GPU-hour compared to 1,081 baseline.