9 月 23 日上午，在 2026 云栖大会 AI 搜索分论坛上，阿里云智能集团高级技术专家杨孔仕发表《面向 Agent 时代的搜索基础设施：亿级租户、千亿向量的阿里云 Elasticsearch Agent 引擎版》主题演讲，系统介绍阿里云如何以对象存储、租户级检索和无状态计算，应对 Agent 应用带来的海量独立检索空间、冷热负载悬殊与持续增长的数据规模。
演讲披露，阿里云 Elasticsearch Agent 引擎版按亿级租户、千亿级向量设计架构容量；在百亿文档、万级知识库的参考负载下，热查询 P99 约 60 毫秒，向量召回率 95% 以上，整体成本比自建 Elasticsearch 降低约 70%。
Agent 时代，搜索基础设施先要应对检索空间“海量而低频”
企业知识库、代码工作区、行业资料库和个人记忆，正在成为 Agent 反复调用的外部上下文。它们给基础设施带来一个新矛盾：需要独立检索的数据空间越来越多，数据要长期保存，但同一时刻真正活跃的只是其中一部分。
“数据需要长期保存，资源却不能为每一份数据始终待命。”杨孔仕表示。传统 Elasticsearch 部署中，按租户建索引容易形成海量分片，共享索引又需要额外的租户过滤；向量索引和多副本 SSD 长期占用资源；存算绑定还意味着，扩缩容和故障恢复都要搬迁数据。总数据量不断增长时，如果一次查询仍要触达大量其他租户的数据，无效的读取和计算也会随之增加。
因此，面向 Agent 的搜索基础设施不仅要扩大容量，还要兼顾检索体验、资源弹性和成本。演讲列出的用户目标是：检索 P95 低于 100 毫秒、召回率不低于 95%；而在现有方案的案例中，检索耗时在 1 秒以上（未注明分位）。
OSS 承载持久数据，计算与缓存只服务活跃工作集
Agent 引擎版把索引数据、预写日志（WAL）和元数据统一持久化到 OSS；写入层（索引节点）负责构建索引，查询层（查询节点）负责检索，两层资源可以分别配置。节点本地的内存和 SSD 主要用作缓存和临时空间，扩缩容或替换节点时，无需先搬迁完整数据集。
杨孔仕将这一机制比作“办公桌和资料库”：桌面只放当前使用的资料，长期资料保存在库中。活跃租户的数据按需加载进缓存，冷数据留在 OSS；业务也可以提前预热目标工作区，减少首次访问的等待。让资源投入取决于活跃负载、而不是全量数据规模，是降低长尾租户常驻开销的关键。
管控系统会分别观测写入层和查询层的 CPU、内存、磁盘和 I/O 水位。计算资源自动扩容计划于 2026 年 10 月上线，届时可在高水位时补充节点、低水位时回收容量；演讲中的节点数量与增长曲线为机制示意。
以 Slice 划定租户边界，让查询直达自己的数据
在数据组织层，一个 Slice 对应一个租户的检索空间，同一租户的数据在存储上连续存放；Slice Collection 则把大量 Slice 汇聚到一个统一的逻辑入口下，后端物理索引按需扩展。业务无需因为底层索引变化反复修改入口，查询、缓存和预热也能围绕明确的租户边界执行。
向量检索方面，DiskBBQ 先用分层 K-Means 把相近的向量聚成簇，再把向量量化为 1 bit 编码，按簇顺序写入磁盘，形成可以整块读取的数据。1 bit 编码的位宽是 32 位浮点数的 1/32，但这不等于整体存储压缩了 32 倍。由于不需要建图，DiskBBQ 的构建明显更快：在 100 万向量的基准上，DiskBBQ 用时 94 秒，HNSW 用时 1054 秒，快了约 10 倍。
多租户场景下，Sliced DiskBBQ 进一步把租户边界带入向量索引：查询时直接定位到当前租户的数据区间，其他租户的数据零 I/O、零打分。“一个知识库的一次查询，不应该因为其他知识库的增长，而承担越来越多无关的读取和计算。” 杨孔仕说。
同一负载下，成本降低约 70%，查询更快、召回更全
先看成本。以自建 Elasticsearch 的各项成本为 100%，在同一参考负载下，Agent 引擎版整体约为 30%，即降低约 70%；其中存储约 20%，内存约 25%，CPU 约 50%。收益主要来自存算分离和按需缓存，低频租户不再占用常驻资源。
再看查询体验。在同一负载下，Agent 引擎版热查询 P99 约 60 毫秒，向量召回率 95% 以上；自建 Elasticsearch 的延迟和召回率分别约为 2 秒和 70%。冷查询需要从 OSS 读取数据，约 600 毫秒，可以通过预热减少首次查询的等待。
从企业知识库到 Agent 记忆，搜索成为可复用的基础能力
目前，阿里云 Elasticsearch 售卖页已上线 Agent 引擎版的开通入口，用户可以启用 Slice Collection 海量多租户检索，并分别配置索引节点和查询节点。围绕租户的数据空间，产品还提供三项能力：预热（warm slice），在用户打开工作区或进入会话时提前加载数据；复制（copy slice），用于建立独立环境；分支（branch slice），基于写时复制，可以秒级创建轻量分支，目前已发布 Beta 版。
这些能力适用于企业知识库与 RAG、AI Coding、法律和金融等行业 AI，以及 Agent 记忆和手机、车载助手等终端记忆场景。以 AI Coding 为例，每个代码仓库或工作区都是一个独立的检索空间，Agent 执行任务时，可以跨文件检索源码、测试和注释，联合使用向量、关键词与过滤条件。单个数据集的需求是：10 万以上文件，全量索引 P95 不超过 1 分钟，更新间隔不超过 10 秒。
杨孔仕在总结中表示，面向 Agent 的搜索基础设施，核心是让“查询范围由租户自身的数据决定”。当 Agent 数量、知识空间和长期记忆持续增长时，只有让持久数据与活跃计算解耦，并按租户边界组织数据和查询，搜索基础设施才能兼顾规模、体验、成本与运维效率。
注：亿级租户、千亿级向量为架构设计容量。文中性能与成本数据是百亿文档、万级知识库参考负载下的结果，实际表现取决于数据集、资源、缓存与检索参数；成本收益随活跃数据比例、资源配置与访问负载变化。DiskBBQ 的构建加速比随数据集、参数与资源配置变化，分支的实际创建耗时受元数据规模与系统负载影响。
阿里云 Elasticsearch Agent 引擎版：help.aliyun.com/zh/es/user-…
阿里云 AgenticSearch：help.aliyun.com/zh/open-sea…
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On the morning of September 23, at the AI Search sub-forum of the 2026 Cloud Habitat Congress, Yang Kongse, Senior Technical Expert of the Alibaba Cloud Smart Group, presented a keynote address entitled " Searching Infrastructure for the Age of Agent: Billion-class Tenants, 100-billion-way Alibaba Cloud Elasticsearch Age Engine Edition ", in which the system described how Alibaba Cloud was stored, leased-house-level search and non-state calculations to respond to the magnitude of data generated by the Agent applications for independent access to space, cold heat load disparity and sustained growth.
The presentation revealed that the Alibaba Cloud Elasticsearch Agent engine version was designed to design the architecture capacity at the hundreds of billions of vectors; under the reference load of the tens of millions of documents and the 10,000-degree knowledge base, thermal query P99 was about 60 milliseconds, with a vector recall rate of over 95 per cent, and the overall cost was about 70 per cent lower than self-built Elasticsearch.
In the Age of Agent, search infrastructure has to deal first with the search for space “mass but low frequency”.
Enterprise knowledge banks, code workspaces, industry databases and personal memory are becoming the external context repeatedly called by Agent. They present a new contradiction to infrastructure: the increasing amount of data space that needs to be independently retrieved and the long-term preservation of data, but only part of it is really active at the same time.
“Data need to be preserved for a long time, but resources cannot be kept on standby for each data.” Yang Kongsei indicates. In the traditional Elasticsearch deployment, a tenant index can easily form a mass fraction, and sharing the index requires additional tenant filtering; vector index and multiple copies of the SSD take over resources for long periods; and storage binding also means that both amplification and failure recovery move data. As the volume of total data continues to grow, the number of invalid readings and calculations increases when a search continues to touch a large number of other tenants.
Therefore, the search infrastructure for Agent should not only expand its capacity but also balance search experience, resource elasticity and cost. The user target listed in the presentation was: to retrieve P95 below 100 ms, with a no less than 95% recall rate; in the case of existing programmes, the search took more than one second (no fraction indicated).
OSS loads persistent data, calculates that cache is active only Set
The Agent engine version provides permanent integration of index data, pre-written logs and metadata into OSS; the Writing Layer (index nodes) builds the index, the Query Layer (Query nodes) searches, and the two tiers can be configured separately. Local memory and SSD at nodes are used primarily for cache and temporary spaces, and there is no need to move complete data sets before scaling up or replacing nodes.
Yang Kongsei compared this mechanism to “desks and databases”: the desktop contains only information currently in use and long-term information is kept in the library. Data on active tenants load on demand entered cache and cold data remained in OSS; operations can also pre-heat target areas in advance, reducing waiting for first visits. Making resource inputs dependent on active loads, rather than full data size, is key to reducing long-tailed tenants ' resident costs.
The control system will observe CPUs, memory, disks and I/O water levels at the writing and query layers, respectively. Computation of an automatic resource build-up scheme is on line in October 2026, when the capacity can be recovered at high water levels by adding nodes and low water levels; the number of nodes in the speech is indicative of the mechanism of growth curve.
Draw the tenant boundary with Slace to allow queries to reach their own data.
On the data organizer level, one Slice corresponds to a tenant's search space, and the same tenant's data is stored on a continuous basis; on the other hand, Slice Collegiation brings a large amount of Slice under a single logical portal and the back-end physical index is expanded as needed. The operations need not be re-engineered over and over again because of changes in the bottom index, and queries, cache and preheating can also be carried out around clear tenant boundaries.
With regard to vector search, DiskBBQ first aggregates the close vector into a cluster with a layer K-Means, then encodes the vector quantization as 1 bit, and writes in a cluster order to the disk, creating data that can be read from the whole block. 1 bit encoding is 1/32 of the 32-bit float, but this is not equivalent to 32 times the overall storage compression. DiskBBQ is clearly constructed faster because it does not need to build the map: on the basis of 1 million vectors, DiskBBQ takes 94 seconds, HNSW takes 1054 seconds, almost 10 times faster.
Under the multi-tenant scenario, Sliced DiskBBQ further brings the tenant boundary into the vector index: directly locates the current tenant ' s data range when searching, and other tenants ' data is zero I/O, zero points. “A search of a knowledge base should not be accompanied by an increasing number of irrelevant readings and calculations because of the growth of other knowledge bases.” Yang Kongsi said.
Under the same load, costs are reduced by about 70%, queries are faster and recall more fully.
Cost first. The cost of self-built Elasticsearch is 100 per cent, and under the same reference load, the Agent engine version as a whole is about 30 per cent, or about 70 per cent lower; about 20 per cent storage, about 25 per cent memory and about 50 per cent CPU storage. The proceeds were mainly from savings separation and on-demand cache, with low-frequency tenants no longer occupying resident resources.
And then look at the query experience. Under the same load, the Agent engine heat query P99 is about 60 ms, with a vector recall rate of more than 95%; self-built Elasticsearch has a delay and recall rate of about 2 seconds and 70% respectively. Cold queries require data to be read from OSS, about 600 milliseconds, and can reduce the waiting for the first query by preheating.
From the Enterprise Knowledge Bank to Agent Memory, search becomes the basic ability for reuse
At present, the Alibaba Cloud Elasticsearch sales page is now available on the Agent engine version of the portal, which allows users to access the Slice Colletion Mass Multiple Tenant Retrieval and to configure index and query nodes, respectively. The product also provides three capabilities around the tenant ' s dataspace: preheating (warm slice), which loads data well in advance when the user opens the workspace or enters a session; copying (copy slice), which is used to create an independent environment; and branch (branch slice), which can create a light branch in seconds based on reproduction at writing, and is currently available in publish Beta.
These capabilities apply to the business knowledge base and industries such as AG, AI Coding, Law and Finance AI, as well as to Agent's memory and terminal memory scenes such as mobile phones and car-mounted assistants. In the case of AI Coding, each code repository or workspace is a separate search space, and Agent is able to access source code, test and comment across files, using vectors, keywords and filter conditions jointly. The requirements for a single data set are: more than 100,000 files, full index P95 not exceeding 1 minute and update no more than 10 seconds.
In his summary, Yang Kongsi stated that the core of the search infrastructure for Agent was to allow “the scope of the inquiry to be determined by the tenant's own data”. When the number of Agents, the knowledge space and long-term memory continue to grow, search infrastructure can be balanced against scale, experience, cost and operational efficiency only if lasting data are decoupled with active calculations and data and queries are organized by tenant boundaries.
Note: Billion renters, hundreds of billions of vectors are structure design capacity. Performance and cost data in text are the result of 10 billion documents, 10,000-degree knowledge base reference loads, the actual performance of which depends on data sets, resources, cache and search parameters; cost-benefits vary with active data ratios, resource allocation and access loads. The construction of DiskBBQ accelerates with changes in data sets, parameters and resource allocation, and the actual creation of branches is influenced by metadata size and system loads.
Alibaba Cloud Elasticsearch Engine Edition: help.aliyun.com/zh/es/user-...
Alibaba Cloud AgencySearch: help.aliyun.com/zh/open-sea...
Auto-fetched and translated by AI Atlas. All rights belong to the original author.