Around 2016, deep learning drove unstructured data such as text, images and audio to be encoded as high-dimensional vectors, while approximate nearest-neighbor algorithms like HNSW balanced retrieval speed, memory overhead and recall. In 2017 Faiss was open-sourced, pushing billion-scale vector similarity search and GPU acceleration from research into engineering practice. Dedicated vector databases such as Milvus then emerged, extending vector technology from a single algorithm library to data persistence, index management, distributed deployment, real-time updates, fault recovery and elastic scaling.
RAG was proposed in 2020, combining generative models with external vector indexes. After 2022, with the rapid rise of LLMs and generative AI, vector databases became key infrastructure for enterprise knowledge bases, semantic search, recommendation, multimodal retrieval and AI assistants. The industry focus also shifted from pure dense-vector search to dense-sparse fusion, full-text search, metadata filtering, hybrid ranking and reranking.
In 2025, Faiss began integrating NVIDIA cuVS, using GPUs to further accelerate index building and online queries. By 2026, vector systems are evolving toward lakehouse-native architectures, zero-copy access, multi-GPU computing, storage-compute separation and unified retrieval engines. Looking ahead, vector retrieval will deeply integrate with full-text search, graph databases, agent long-term memory and multimodal reasoning, and the competition will upgrade from “faster lookups” to “more accurate understanding, more real-time updates, lower cost, better governance”, gradually becoming the foundational data layer of intelligent applications.