Vector search drives AI use cases like RAG and semantic search, but as index size climbs, keeping data in RAM becomes costly and a不卡 bottleneck. ANN algorithms trade exactness for speed, and choosing the right one depends on scale and latency needs.

In‑memory ANN, exemplified by HNSW, offers ultra‑low latency by storing the entire graph in RAM. It excels for small to medium workloads, but when indexes exceed RAM or become prohibitively expensive, performance drops sharply due to non‑sequential disk I/O.

On‑disk ANN, such as SPANN and Disk ANN, reduces memory footprints by keeping vectors on disk while keeping routing or graph layers in RAM. SPANN uses centroids for clustered, sequential reads, whereas Disk ANN’s Vamana graph minimizes hops and random access, yielding acceptable latency even for billion‑vector+ indexes.

Vector databases like Qdrant, Milvus, and Chroma DB implement these algorithms, letting teams balance cost, speed, and scale by selecting the appropriate ANN strategy for their workload.