A long-lost conversation with old friends and a wonderful collision of AI exploration.
Recently, Milvus Arch Meetup concluded successfully in Shanghai. This Meetup had many highlights. It not only received strong support from the KubeBlocks community, but also invited senior experts from NetEase Fuxi and Ant Group to share their thoughts on cloud native and vector databases in the AI era.
[Swipe to see more]
Next, let’s quickly review the key takeaways from the event:
Zilliz senior engineer Xia Congqi : Whether it is architecture, new features, performance or maintainability, Milvus 2.3.x is a leader among vector databases and is definitely worth a try
Chen Jinglai, senior AI R&D engineer at NetEase Fuxi : Milvus has been shown in the practice of NetEase multi-modal graphic and text scenes that it effectively supports the implementation of NetEase Fuxi's billion-level graphic and text data and applications.
Guo Ziang, senior engineer at Yunsheng Data : Use KubeBlocks to easily manage AIGC data infrastructure such as vector database + LLM
Ant Group Engineer Xu Pengfei : Use KCL declarative configuration language and tools to address engineering configuration strategy challenges
The following is a detailed explanation, you can enjoy it on demand:
01.
Interpretation of new functions of Milvus 2.3.x
Xia Congqi gave a comprehensive interpretation of Milvus 2.3.x from the aspects of architecture, new functions, performance and maintainability.
He first mentioned that Milvus 2.3.x has been upgraded in terms of architecture, including support for heterogeneous hardware (GPU Index: RAFT; ARM) and upgraded QueryNode (QueryNodeV2) . Xia Congqi focused on introducing QueryNodeV2. QueryNode is responsible for the most important retrieval service in the entire Milvus system. Its stability, performance, and scalability are crucial to Milvus. However, QueryNodeV1 has problems such as complex status, repeated message queues, unclear code structure, and unintuitive error content. In the new design of QueryNodeV2, the team reorganized the code structure, changed the complex state to a stateless design, and removed the message queue of delete data to reduce resource waste. In the subsequent continuous stability testing, the performance of QueryNodeV2 More excellent.
In terms of new features, these features launched in Milvus 2.3.x deserve special attention:
Upsert : Since Milvus does not currently support the update operation, users need to delete the old records and re-insert them when they need to update the vector. In version 2.3, the Upsert interface provided by Milvus guarantees an atomic "modification" operation;
SCANN Index : Knowhere 2.0, supports SCANN index;
Iterator (pymilvus only) : Both Search and Query of Milvus have data online. When users need to query a large amount or even a full amount of data, the existing interface cannot fully meet this demand. After Milvus supports range search, pymilvus simulates a set of Iterator interfaces by dynamically adjusting the range, which can return large batches of data required by users;
Delete By Expression : Delete interface. Before 2.3, deletion could only be done through primary key expression (ID in [1, 2, 3, …]). When the user wants to delete some data that meets the conditions, he needs to query its primary key first, and then perform the delete operation. Delete by expression provides the "syntactic sugar" of the Milvus server to complete this operation within the system.
Not only that, Milvus 2.3.x also supports MMap, Growing Index, dynamic configuration modification, CDC, etc., which greatly improves the overall performance and operability of Milvus. Interested students can check out the Milvus 2.3.x series of articles for details.
02.
Milvus’ practice in NetEase’s graphic and text multi-modal scenes
Chen Jinglai shared " Milvus's practice in multi-modal scenes of NetEase graphics and text ". He said that increasing model size and improving data quality are important means to obtain better artificial intelligence results.
NetEase Fuxi has been engaged in large model research for 5 years, accumulated rich algorithm and engineering experience, and has created dozens of text and multi-modal pre-training models. Vectors play an important role in large-scale language models, such as: Embedding - AI-based tools and algorithms that can map unstructured data, such as text, images, audio and video, to low-dimensional space and represent it as embedding.
Today, Fuxi's graphic data has accumulated over 1 billion Internet data and NetEase's own copyright data. There is a large demand for graphic retrieval, which also brings various challenges:
High resource usage : taking up a lot of computing and storage resources
Heterogeneous resources : GPU, CPU, SSD and other different types of resources
Complex business types : multi-modal graphics and text, NLP, user profiling and other businesses. Different business data sizes, delays, service quality, recall accuracy, etc. are different.
Stability and reliability
In this case, how to build a high-performance, highly reliable, heterogeneous graphic and text vector engine is crucial. The Milvus architecture has the characteristics of cloud native, separation of storage and computing, distribution, redundancy and high availability. With its help, NetEase Fuxi has achieved the creation of a billion-level Milvus cluster.
In addition, Chen Jinglai also mentioned that under the research and development framework such as tackling difficult tasks, Fuxi initiated the accumulation and application exploration of capabilities in the multi-modal field of graphics and text. It has self-developed a picture and text generation model "Danqing" that supports Chinese scenes. Based on this, it has launched an AI painting platform "Danqingyue", and LangChain + Milvus can build a Danqingyue painting agent.
For the future, Chen Jinglai looks forward to using Milvus to explore and retrieve enhanced generation (RAG) to improve the capabilities of image and text multi-modal models, and to use more Milvus+ capabilities to improve the application of image and text multi-modal scenes.
03.
KubeBlocks: Easily manage AIGC data infrastructure
With the theme of " KubeBlocks: Easily Manage AIGC Data Infrastructure ", Guo Ziang explained KubeBlocks' AIGC data infrastructure solution in the context of the AI era: KubeBlocks helps users build their own AI applications by providing vector database hosting and LLM hosting capabilities. Greatly reduces the burden on application developers.
Database hosting capabilities of KubeBlocks
KubeBlocks 作为开源管控平台，可运行和管理 K8s 上的数据库、消息队列及其他数据基础设施。基于这一特点，KubeBlocks 的解决方案采用托管向量数据库（如 Milvus）和图数据库（如NebulaGraph）的方式，实现多云和线下部署，在实现快速 day-1 集成的同时，也提供了丰富的 day-2 运维操作。
KubeBlocks 依靠其强大的集成和抽象能力，可快速实现数据库集成。郭子昂以 Milvus 为例，展示了根据 KubeBlocks 的 API 在 YAML 文件中定义 Milvus 各种特性、运维配置，轻松实现向量数据库全生命周期管理。
KubeBlocks 的 LLM 托管能力
KubeBlocks 具有强大的 LLMOps 能力，支持托管 LLM 及多种大模型。基于 KubeBlocks，开发者可实现 LLM 私有化部署，同时支持定制化大模型，实现行业数据的精细调整。此外，KubeBlocks 的 LLMOps 能力还支持 LLM 开发环境私有化部署、分布式部署、高性能 batching，充分适配本地开发环境和生产环境，提升 GPU 利用率。
KubeChat：KubeBlocks AIGC 解决方案落地
基于上述解决方案，KubeBlocks 已成功落地 AI 应用，在 10 天时间开发出 AI 知识库应用 KubeChat，轻松应对 Embedding、向量数据库和大模型在开发 AI 应用过程中带来的各类挑战。
点击下方链接，查看 KubeChat 演示视频：
04.
KCL 在 AI 工程配置策略场景的探索和落地使用
徐鹏飞分享了《KCL 在 AI 工程配置策略场景的探索和落地使用》。平台工程和 AI 工程的发展日益迅猛，但这也带来了问题和挑战，比如认知负担、配置/数据种类繁杂、配置/数据清洗过程易出错、效率可靠性低等。KCL 作为专用配置策略语言为配置和自动化提供了解决方案，以收敛的语言和工具集合解决领域问题近乎无限的变化和复杂性，同时兼顾表达力和易用性。
此外，KCL 以数据和模型为中心，采用开发者可以理解的声明式 Schema/配置/策略模型用于 AI 工程、云原生工程等场景。KCL 为开发人员提供了通过记录和函数语言设计将配置（config）、建模抽象（schema）、逻辑（lambda）和策略（rule）作为核心能力，具有可复用可扩展、抽象和组合能力、稳定性、高性能等特点。
KCL 可以广泛用于表格数据集验证和转换、云原生配置验证和转换、通过抽象进行应用交付、IaC & GitOps等场景。KCL 也注重开发者体验，提供完备的 Language + Tools + IDEs + SDKs + Plugins 工具链支持，还支持模型 Registry。
彩蛋：看看模型 Registry 里出现了谁？
回复关键词【老友汇上海】获取现场嘉宾 PPT。
本文分享自微信公众号 - ZILLIZ（Zilliztech）。
如有侵权，请联系 [email protected] 删除。
本文参与“OSC源创计划”，欢迎正在阅读的你也加入，一起分享。