This article explains how TAL Education leverages the open‑source Milvus vector database—covering its architecture, features, cloud‑native deployment, monitoring, and real‑world AI applications such as intelligent grading and multimodal search—to handle billions of vectors with millisecond‑level similarity retrieval.
Alibaba Mom’s AI‑driven content risk platform has evolved from simple rule‑matching to a data‑centric, serverless architecture that integrates large‑model acceleration, decision‑tree compilation, high‑throughput vector retrieval and elastic word‑matching, delivering sub‑100 ms text and sub‑1 s image moderation while remaining stable during peak promotional traffic.
Baidu has open‑sourced its high‑performance, trillion‑scale vector search engine Puck—originally built for ultra‑large image‑search workloads, winner of multiple BIGANN categories, now supporting diverse embeddings alongside the medium‑size Tinker algorithm—to accelerate community innovation, improve code quality, and broaden AI retrieval applications across search, recommendation and cloud services.
This article describes the design, algorithmic improvements, and engineering performance optimizations of Bilibili's massive video duplicate detection (collision) system, covering challenges of low‑edit‑degree reposts, two‑stage retrieval, self‑supervised feature extraction, GPU‑accelerated preprocessing, and the resulting gains in accuracy and throughput.
Bilibili built a massive video‑duplicate detection platform that trains a self‑supervised ResNet‑50 feature extractor, removes black borders, and uses a two‑stage ANN‑plus‑segment‑level matching pipeline accelerated by custom GPU decoding and inference, boosting duplicate rejection 7.5×, recall 3.75×, and cutting manual misses from 65 to 5 per day.
The paper reports a year‑long evolution of Alibaba’s search‑advertising keyword recall, replacing the traditional two‑stage rewrite‑and‑score pipeline with a low‑storage binary retrieval model and then a joint recall framework built on a hierarchical bidding graph, delivering near‑full‑precision recall, 16× memory savings, and quota‑free global ranking.
The inaugural BigANN competition, organized by NeurIPS, showcased large‑scale ANN research, and Baidu's self‑developed PUCK algorithm secured top scores across all four tracks by leveraging multi‑layer quantization, two‑level inverted indexing, and extensive system‑level optimizations.
The Dual Vector Foil (DVF) system decouples index construction from model training by building a post‑training HNSW graph, enabling any complex model to score candidates, which yields a 5.7 % recall boost, cuts latency from ~40 ms to 6.5 ms, and raises QPS over tenfold while simplifying maintenance.
This article describes the challenges of operating massive, TB‑scale retrieval clusters at Zuoyebang, and presents a Fluid‑based compute‑storage separation architecture that improves data distribution, update efficiency, scalability, and stability, enabling containerized search services to be managed like regular stateless workloads.
The article describes the design, architecture, and key components of a vector search platform built on Faiss that supports full‑index construction, incremental and distributed indexing, online retrieval, city‑level search, and vector update/delete operations to meet large‑scale AI application needs.
This article details the evolution of WeChat’s Identify product, explaining its end‑to‑end image recognition pipeline—including object detection, multi‑label classification, mobile‑side detection, large‑scale retrieval, unsupervised clustering, and system architecture—while showcasing various application scenarios such as product, plant, and landmark recognition.
This article introduces JTM, a joint optimization framework that simultaneously learns deep scoring models and tree-structured indexes, addressing the limitations of traditional recommendation pipelines and demonstrating significant precision and recall gains on large-scale datasets such as Amazon Books and UserBehavior.
This article details how Alibaba's engineers reduced per‑image feature size from 120 KB to an average of 13 KB and accelerated Euclidean distance calculations from 50 µs to 9 µs through bitmap, offset, and duplicate‑value compression techniques, enabling efficient large‑scale image retrieval.
