9月23日上午，在2026云栖大会 AI 搜索分论坛上，阿里巴巴 ATH 事业群 Qoder 技术专家夏晓文发表《Qoder × 阿里云 Elasticsearch：千亿向量检索引擎的生产级落地》主题演讲，系统介绍 Qoder 如何借助阿里云 Elasticsearch Agent 引擎版构建代码上下文检索底座，并分享从多租户隔离、索引构建到高可用、可观测和成本控制的生产实践。
在 Agentic Coding 场景中，模型能否理解一个真实软件项目，不只取决于代码生成能力，还取决于能否在有限上下文中准确找到相关代码、符号、提交记录和 Wiki。上下文是否准确、完整，会直接影响 Agent 后续的代码理解和修改效果。 随着代码库和用户规模持续增长，检索系统也必须同时满足成本、时延、隔离和稳定性要求。
检索不是“塞入更多代码”，而是找到完成任务所需的上下文
Qoder 面向个人和团队提供 Agentic Coding 平台，产品形态覆盖 Desktop、IDE、JetBrains 插件、CLI 和 Mobile，并通过 Agent SDK、Cloud Agents 等方式开放 Agent 能力。不同入口背后有一个共同需求：让 Agent 理解用户正在处理的任务，以及任务背后的项目上下文。
以修改一个功能为例，Agent 不仅要知道相关代码在哪里，还要理解调用关系、项目约定和历史变更。全量代码难以直接放入模型上下文，检索系统需要先定位与当前任务相关的信息，再补充必要依赖和背景。这类检索会贯穿开发任务的多轮交互，单次查询的延迟和成本会直接影响整体体验与服务成本。同时，不同代码库的访问频率差异明显：活跃工作区需要持续、快速地获取上下文，低频代码库则需要低成本保存，并在再次访问时及时响应。Qoder 因此提出多项基础设施目标，包括单个代码库月成本不超过1元、P95 检索耗时不超过100毫秒、支撑千万级租户，并兼顾热点数据访问与冷数据首次查询。（注：这些数字是选型与架构目标，不等同于当前生产实测规模。）
一个 Codebase 对应一个 Slice，让全局规模不成为单次查询负担
面向海量多租户场景，阿里云 Elasticsearch Agent 引擎版通过 Collection 与 Slice 组织业务数据。在 Qoder 的映射关系中，同维度向量进入对应 Collection，一个 Codebase 对应一个 Slice，一段代码 Chunk 对应一个 Document。查询限定在目标 Slice 内，只触达当前代码库的数据，既保持租户隔离，也避免让全局数据规模直接转化为每次检索的扫描负担。
演讲材料展示，在特定参考负载下，热查 P99 约为60毫秒、召回率不低于95%；亿级租户和千亿级向量则代表产品设计容量，而非 Qoder 当前租户数。成本侧采用对象存储承载大容量数据，以 SSD 和内存缓存服务活跃访问，冷数据无需全部长期占用高性能介质。一个 Codebase 对应一个 Slice，是 Qoder 在多租隔离、资源共享和按需访问之间取得平衡的关键。
一次用户查询进入 Qoder 后，会先经过查询处理、改写和意图识别，再通过向量、关键词、符号、提交记录和 Wiki 等通道进行多路召回。结果合并去重后，系统继续构建子图、扩展调用与依赖关系，最后完成排序和上下文融合。夏晓文表示：“ES Agent 引擎版承担其中的向量召回能力，完整检索链路还包括前后的理解、扩展和结果处理。”
从请求成功到业务成功，生产化要覆盖每一个工程环节
索引构建从上传开始：Manager 接收任务并通过消息队列调度 Workers，Workers 调用 Embedding Service 生成向量，再分别写入 KV Cache 和 Elasticsearch。业务侧通过统一接口提供写入、查询和生命周期能力，同时持久化每个 Codebase 的实际后端路由，确保后续读写落到一致的存储目标。
进入生产后，系统需要处理大量容易被忽略的细节。Collection 和 Slice 在首次非空写入时按需创建；Bulk 请求按报文大小分批提交，并逐项检查执行结果，因为 HTTP 成功并不代表每条写入都成功；查询必须明确指定 Slice 和过滤条件，限定召回范围；底层分数还要经过归一化处理，与上层阈值语义保持一致。
安全与可靠性同样贯穿全链路。演讲介绍，Qoder 在 Elasticsearch 中只保存向量及必要的位置等元数据，不保存原始代码或代码 Chunk 正文。查询侧采用多节点和副本配置，客户端区分读写地址，并对适合重试的暂态错误进行处理；运维侧通过 Trace 和 Codebase 串联入口、实际后端、HTTP 请求、Bulk 单条结果和耗时。生产可靠性的验收标准不是“接口返回成功”，而是业务读写成功、数据完整，并能够在故障后恢复。
千亿级文档进入生产运行，性能、吞吐与成本按明确口径评估
根据演讲披露，Qoder 当前相关存储文档规模已达到千亿级，写入资源为30多台 ECS，查询资源为5台以上 ECS；最高支持查询 QPS 1000，Agent 并发检索 QPS 超过20，写入峰值超过1GB/s。以每个代码库1万条向量、共享集群及指定折后价格估算，单 Codebase 月均基础设施成本约为0.016元。该数字是特定假设下的基础设施分摊估算，不是固定价格，也不代表完整业务成本。
在演讲展示的 Qoder 测试口径中，Elasticsearch 纯向量召回阶段的 P50 和 P95 分别为11.73毫秒和14毫秒，同类产品分别为30.47毫秒和36.25毫秒，P95 降幅约61.4%；写入带宽展示值为2GB/s，对比400MB/s，约为5倍；以同类产品总体成本为100%，Elasticsearch 对应值约为15%。上述时延、写入和成本数据采用不同测试与测算条件，不应组合外推为所有负载下的统一结果；实际表现会随数据规模、资源配置、并发和缓存状态变化。
从代码检索走向代码、记忆与知识的融合检索
代码检索只是 Agent 上下文工程的一部分。Qoder 还计划将同一套工程能力扩展到用户与项目记忆、企业知识库和跨仓库知识：记忆需要关注隔离、时效和删除机制；企业知识库需要在文档与 Wiki 混合检索时执行权限过滤，并保留来源和版本；跨仓库知识则需要关联代码、Issue 与评审记录，为跨项目任务补充可追溯背景。
未来，Agent 可先识别任务目标与授权范围，再选择代码、记忆或知识库检索器，对不同来源的结果统一去重和重排。在有限 Token 预算内保留最有帮助的内容，同时保留证据来源，是融合检索继续演进的核心方向。 演讲提出，Qoder 将继续探索代码、记忆与企业知识的融合检索，让 Agent 获得更相关、来源更明确、也更可解释的上下文。
本文根据2026云栖大会 AI 搜索分论坛现场演讲材料整理。文中规模、性能、吞吐与成本数据均以演讲披露的版本、配置、测试和测算口径为准；未来场景属于规划方向，不代表相关能力已全部上线。
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 23 September, at the AI Search Sub-forum of the 2026 Clouds Congress, Alibaba ATH consortium Qoder technical expert Sha Xiaowen delivered a keynote address entitled " Qoder x Alibaba Cloud Elasticsearch: Ten hundred Billion Vector Retrieval Engine Production Level Down ", a systematic presentation on how Qoder built the code context Retrieval Base using the Alibaba Cloud Elasticsearch Agent engine version and shared production practices from multi-tenant isolation, indexing to high availability, observable and cost control.
In the Agecoding scene, the model's ability to understand a real software project depends not only on the ability to generate the code, but also on the accuracy of the relevant codes, symbols, submission records and Wiki in the limited context. The accuracy and completeness of context directly affect the understanding and modification of the subsequent Agent code. As the number of coding repositories and users continues to grow, the retrieval system must simultaneously meet the cost, latency, isolation and stability requirements.
Retrieving is not "plugging into more codes," but finding the context required to complete the mission.
Qoder provides a platform for individuals and teams, with product forms covering Desktop, IDE, Jetbrains plugins, CLI and Mobile, and opens up Agent capabilities through Agent SDK, Cloud Ages, etc. There is a common need behind the different entrances: to get Agent to understand the tasks being handled by the users and the project context behind the tasks.
For example, to modify a function, Agent not only knows where the code is, but also understands call relationships, project engagements and historical changes. Full code is difficult to put directly into model context, and the retrieval system needs to locate information relevant to the current mission and to supplement the necessary dependencies and background. This type of search interacts with multiple development missions, and delays and costs of single searches directly affect overall experience and service costs. At the same time, the frequency of access varies significantly between different code libraries: active workspaces require continuous and rapid access to context and low-frequency code libraries require low-cost storage and timely response when re-visiting. Qoder therefore proposes a number of infrastructure objectives, including a single code repository with a monthly cost of less than $1; a P95 search time of less than 100 milliseconds; support for tens of thousands of renters; and a balance between hot spot data access and cold data for the first time. (Note: These figures are selected and structured targets that are not equal to the current measured scale of production.
A Codebase corresponds to a Slice so that global scale does not become a single query burden
Alibaba Cloud Elasticsearch Agen, an engine version, uses Colleaction and Slice business data for large-scale multi-tensors. In the Qoder map relationship, the same dimension vector enters the Collaction, one Codebase corresponds to one Slace, one code to one Document. The query is limited to the target Slice and only touches the data of the current code library, keeping the tenant isolated and avoiding the direct transformation of global data scales into a scanning burden for every search.
The presentation showed that under a specific reference load, thermal inspection P99 was about 60 ms, with a recall rate of no less than 95%; the hundreds and hundreds of billions of vectors represented product design capacity rather than the current number of Qoder tenants. The cost side uses object storage for high capacity data to be accessed with active access to SSD and memory cache services, and cold data do not require full long-term occupancy of high performance media. One Codebase corresponds to one Slice, which is the key to Qoder balancing multi-lease segregation, resource sharing and on-demand access.
After a user query enters Qoder, the search is processed, rewrited and intended to be identified, followed by multiple circuits through the vector, keywords, symbols, submission records and Wiki. As a result of the merger, the system continues to build the sub-graph, expand the call and dependencies, finalize the sorting and context integration. Ha Xiaoven states: “ES Agent engine version undertakes the vector recall capability in it, and the complete search chain also includes ex ante and ex ante understanding, extension and result processing.”
From successful requests to successful operations, production must cover every piece of work.
Index construction starts with uploading: Manager receives jobs and dispatches them through message queues. The operational side provides writing, query and life-cycle capabilities through a unified interface, while at the same time sustaining the actual back-end route for each Codebase to ensure that subsequent reading and writing reach a consistent storage target.
Once in production, the system needs to deal with a large number of easily neglected details. Colletion and Slice are created as needed when writing for the first non-empty; Bulk requests are submitted in batch by size, and the results are checked on a case-by-case basis, as the success of HTTP does not mean that every writing is successful; queries must clearly specify the Slice and filter conditions, and limit the recall range; the bottom fractions are also to be normalized, consistent with the upper threshold syntax.
Safety and reliability also cut across the chain. Presentations, Qoder saves only metadata such as vectors and necessary locations in Elasticsearch and does not save original or code Chunk text. The search side is configured with multiple nodes and copies, the client differentiates between reading and writing addresses and handles temporary errors suitable for retry; the side passes through Trace and Codebase serial access, physical backend, HTTP request, Bulk single result and time-consuming. acceptance criteria, which produces reliability, is not “a successful interface return”, but a successful business reading and writing process, complete data and able to recover after failure.
Hundreds of billions of documents entering production operations with performance, throughput and cost assessed by clear calibre
According to the presentation, Qoder currently has hundreds of billions of levels of relevant storage documents, with more than 30 ECS written and more than five ECS asked for; maximum support is given to query QPS 1000, Agent and retrieve QPS over 20 and write at peaks above 1GB/s. The average monthly infrastructure cost per code repository, estimated at 10,000 vectors, shared clusters and specified discount prices, is approximately 0.016. This figure represents an estimate of the infrastructure share under a given scenario and is not a fixed price and does not represent the full cost of operations.
In the Qoder test caliber presented in the presentation, the P50 and P95 of 11.73 ms and 14 ms for the pure vector recall phase of Elasticsearch were respectively 30.47 ms and 36.25 ms for the same product, with a reduction of about 61.4 per cent for P95; the written bandwidth display of 2GB/s, compared to 400MB/s, was about 5 times; the total cost of the same product was 100 per cent and the Elasticsearch equivalent was about 15 per cent. The above-mentioned latency, writing and cost data are subject to different test and measurement conditions and should not be extrapolated to the harmonized results under all loads; actual performance will vary according to the size of the data, the allocation of resources, and the status of cache.
Retrieval from code to code, memory and knowledge.
The code search is only part of the Agent context project. Qoder also plans to extend the same set of engineering capabilities to user-project memory, enterprise knowledge base and cross-warehouse knowledge: memory requires attention to seclusion, timeliness and removal mechanisms; enterprise knowledge base requires permission filtering and retention of source and version when documents are mixed with Wiki; cross-warehouse knowledge requires associated codes, Issue and evaluation records to supplement the retrospective background for cross-project tasks.
In the future, Agent can first identify mission objectives and mandates, then select a code, memory or knowledge base retriever, and reorder and reorder results from different sources. Maintaining the most helpful elements within the limited Token budget while preserving sources of evidence is the central direction in which integrated retrieval continues to evolve. It was suggested that Qoder would continue to explore the integration of codes, memories and business knowledge, allowing Agent to obtain a more relevant, well-sourced and interpretable context.
This paper is based on live presentations from the AI Search Sub-Forum at the 2026 Clouds Congress. The size, performance, throughput and cost data in the text are based on the version, configuration, testing and calibration of the presentation; the future scene is in the planning direction and does not represent that all relevant capabilities are on line.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.