The exponential growth of high-dimensional data has driven distributed Approximate Nearest Neighbor (ANN) search systems to confront dual challenges: escalating query throughput demands and complex network bottlenecks. While conventional solutions scale machine clusters for load balancing, they incur significant latency penalties and resource inefficiencies. We present AINetD, a paradigm-shifting framework that embeds computational intelligence directly into network infrastructure through programmable switches. Our approach introduces intelligent query awareness and in-network candidate pruning. Implemented on P4-programmable intelligent switches, AINetD demonstrates performance improvements compared with typical distributed search systems: a 2.4x throughput enhancement for tail latency-sensitive queries, 25% queries reduction in redundant computations and 80% lower network communication overhead through network-level optimizations.