OpenSearch veterans launch Infino. Here’s why it matters for agent builders. OpenSearch 老兵创立了 Infino。这对智能体（Agent）构建者来说为何意义重大？
Software vendors like to claim unified systems that act as a so-called single pane of glass for any given coding task, data function, or broader workflow. Agentic data retrieval company Infino hints at that vibe, promising a single retrieval layer for all agents.软件供应商喜欢宣称其系统是统一的，能为任何特定的编码任务、数据功能或更广泛的工作流提供所谓的“单一视图”。智能体数据检索公司 Infino 正是瞄准了这一方向，承诺为所有智能体提供统一的检索层。
In this context, a single retrieval layer means a unified (there’s that word) database system that lets all AI agents query all data from one place.在这种背景下，“单一检索层”指的是一个统一的数据库系统，它允许所有 AI 智能体从同一个地方查询所有数据。
Agents have changed the data query paradigm 智能体改变了数据查询范式
Formally launching its agent retrieval platform on Wednesday, Infino says it built the platform because agents have changed the data query paradigm and don’t query like humans. Infino 于周三正式发布其智能体检索平台。该公司表示，构建该平台的原因在于智能体已经改变了数据查询范式，它们的查询方式与人类不同。
Infino CEO Ekechi Nwokah tells The New Stack that agents are the “largest new consumer of data since the web browser”, but he laments that they are still “being served with a fragmented stack built for a different era,” i.e., data warehouses for SQL, search engines for keywords, vector databases for semantic search, and ETL pipelines to keep them all in sync. Infino 首席执行官 Ekechi Nwokah 在接受 The New Stack 采访时表示，智能体是“自网页浏览器以来最大的新数据消费者”，但他遗憾地指出，它们目前仍在使用“为另一个时代构建的碎片化技术栈”，即：用于 SQL 的数据仓库、用于关键词的搜索引擎、用于语义搜索的向量数据库，以及用于保持这些系统同步的 ETL 流水线。
“We’re collapsing that stack for agents,” Nwokah says. “Our approach is to put the data in Parquet in object storage and let agents search, rank, filter, join, aggregate, and reason over it directly. The result is one copy of the data that all agents read.”“我们正在为智能体精简这一技术栈，”Nwokah 说道。“我们的方法是将数据以 Parquet 格式存放在对象存储中，并让智能体直接对其进行搜索、排序、过滤、连接、聚合和推理。最终结果是所有智能体都读取同一份数据副本。”
As we know from database basics lesson #1, data stores have been traditionally optimized for a human or application that writes one query, waits, and reads the result. 正如我们从数据库基础课程第一课中所学到的，传统的数据存储主要针对人类或应用程序进行优化，即执行一个查询、等待并读取结果。
“An agent behaves quite differently,” Nwokah says. “It issues many small questions to answer one large one, often at the same time, and every result it receives consumes context and can add cost and latency to subsequent model calls. That is a new kind of reader that existing data stores were not optimized for.”“智能体的行为截然不同，”Nwokah 说。“它为了回答一个大问题会同时发出许多小问题，这往往是并发进行的。它收到的每一个结果都会消耗上下文，并可能增加后续模型调用的成本和延迟。这是一种现有的数据存储并未针对其进行优化的新型阅读者。”
“An agent behaves quite differently. It issues many small questions to answer one large one, often at the same time, and every result it receives consumes context, and can add cost and latency to subsequent model calls.”“智能体的行为截然不同。它为了回答一个大问题会同时发出许多小问题，这往往是并发进行的。它收到的每一个结果都会消耗上下文，并可能增加后续模型调用的成本和延迟。”
Nwokah explains that Grep (the command-line tool that searches plain text for line matches on a specific keyword or pattern) returns matches, and the agent goes on to read whole files. Further, he notes that a vector database will return a handful of chunks, and then the agent asks again with additional queries. Additionally, a data warehouse returns rows, and the agent combines them itself. Nwokah 解释说，Grep（一种在纯文本中搜索特定关键词或模式匹配行的命令行工具）返回的是匹配项，随后智能体需要读取整个文件。此外，他指出向量数据库会返回少量数据块，然后智能体又会带着额外的问题再次查询。不仅如此，数据仓库返回的是行数据，智能体还需要自行将其组合。
“Infino is a retrieval layer designed for this data reader dynamic. It gives the agent one interface directly over object storage that combines structured and unstructured queries: ranked searches, exact counts, joins, filters and aggregates at low cost and large scale,” Nwokah clarifies.“Infino 是专为这种数据阅读动态而设计的检索层。它为智能体提供了一个直接覆盖对象存储的接口，结合了结构化和非结构化查询：以低成本和大规模实现排名搜索、精确计数、连接、过滤和聚合，”Nwokah 阐述道。
During the good old years of data search, query patterns were comparatively predictable. When humans or applications queried, search ran on a search database. Analytics ran in a warehouse. Vector databases enabled approximate queries and unlocked a new class of use cases, but they didn’t fundamentally change those query patterns. Each query ran on its own system, with its own copy of the data, its own ingestion pipeline, and its own team to run it.在数据搜索的黄金年代，查询模式相对可预测。当人类或应用程序进行查询时，搜索在搜索数据库上运行，分析在数据仓库中运行。向量数据库实现了近似查询并开启了一类新的用例，但它们并没有从根本上改变这些查询模式。每个查询都在其各自的系统上运行，拥有自己的数据副本、自己的摄取流水线以及专门的维护团队。
Today, those operations can still span a database plus a search engine, a vector database, and a data warehouse. The agent has to call each one and then spend additional model turns combining the results. Infino’s answer is to embed retrieval functions directly in SQL, so the whole question can be expressed in one query.如今，这些操作可能仍然跨越数据库、搜索引擎、向量数据库和数据仓库。智能体必须调用每一个系统，然后花费额外的模型轮次来合并结果。Infino 的解决方案是将检索功能直接嵌入 SQL 中，这样整个问题就可以通过一个查询来表达。
Fewer data tasks, less glue code, centralized policy management更少的数据任务、更少的胶水代码、集中的策略管理
“For developers and data teams, it removes entire pipelines and the time it takes to manage them,” Nwokah enthuses. “That means fewer systems, fewer ingest paths, fewer ETL jobs, fewer schemas to keep aligned, fewer clients in the agent’s tool list, and less glue code to combine their results. It also means one place to govern agents.”“对于开发人员和数据团队来说，这消除了整个流水线以及管理它们所需的时间，”Nwokah 热情地说道。“这意味着更少的系统、更少的摄取路径、更少的 ETL 作业、更少需要保持一致的模式、智能体工具列表中更少的客户端，以及更少的合并结果所需的胶水代码。这也意味着有一个统一的地方来治理智能体。”
He underlines his organization’s method here and says that when an agent’s data lives in a single copy, developers “can enforce hard policies centrally”, so that means governance over who can read what, which rows and columns an agent can see, and what gets logged. 他强调了其组织的方法，并表示当智能体的数据仅保留一份副本时，开发人员“可以集中实施严格的策略”，这意味着可以对谁能读取什么、智能体能看到哪些行和列以及记录什么内容进行治理。
“Compare that with managing agent permissions across MCP gateways and SaaS tools,” urges Nwokah.“将此与跨 MCP 网关和 SaaS 工具管理智能体权限的方式进行比较吧，”Nwokah 敦促道。
“For developers and data teams, it removes entire pipelines and the time it takes to manage them.”“对于开发人员和数据团队来说，这消除了整个流水线以及管理它们所需的时间。”
A developer can enter a single natural-language English question, and it becomes one query that completes in milliseconds, according to Infino. Infino has clarified that the important part isn’t SQL itself; it’s that keyword search, semantic search, filtering, aggregation, and grouping can happen together instead of forcing the agent to orchestrate several systems and reason over a pile of intermediate results.据 Infino 称，开发人员可以输入一个简单的自然语言英语问题，它会变成一个在毫秒内完成的查询。Infino 澄清说，重点不在于 SQL 本身，而在于关键词搜索、语义搜索、过滤、聚合和分组可以同时进行，而不必强迫智能体协调多个系统并对一堆中间结果进行推理。
As an additional challenge uncovered during the platform’s development, Infino noted that even when retrieval is fast, it found that agents spend a lot of time in retrieval loops, i.e., working to formulate a query, search, decide whether the results are good enough, try something else, validate the answer, and repeat. Infino says it integrates targeted inference models with the retrieval engine so these simpler tasks run faster and more cheaply than with frontier models.作为在平台开发过程中发现的额外挑战，Infino 指出，即使检索速度很快，智能体仍会花费大量时间在检索循环中，即：努力构建查询、搜索、判断结果是否足够好、尝试其他方法、验证答案，然后重复上述步骤。Infino 表示，它将目标推理模型与检索引擎集成在一起，因此这些较简单的任务运行速度更快，成本也比使用前沿模型更低。
The data storage challenge: Apache Parquet数据存储挑战：Apache Parquet
The other challenge Infino tackled is how it stores the data. Instead of maintaining separate copies for search, vectors, and analytics, Infino keeps a single copy in Apache Parquet (a column-oriented storage format) on object storage.Infino 解决的另一个挑战是如何存储数据。Infino 没有为搜索、向量和分析维护单独的副本，而是将单一副本以 Apache Parquet（一种列式存储格式）格式保存在对象存储中。
“The core engine is open source, Apache-2.0, on GitHub, and we store all data as valid Parquet files with the search indexes embedded beside the Parquet footer, so any tool that reads Parquet can read the data with or without Infino. Infino Cloud, our hosted service, has some additional proprietary features, like being able to point at your existing Parquet data and make it searchable without building any infrastructure yourself,” adds Nwokah. “核心引擎是开源的，采用 Apache-2.0 协议，托管在 GitHub 上。我们将所有数据存储为有效的 Parquet 文件，并将搜索索引嵌入在 Parquet 页脚旁边，因此任何读取 Parquet 的工具都可以读取这些数据，无论是否使用 Infino。我们的托管服务 Infino Cloud 具有一些额外的专有功能，例如能够指向您现有的 Parquet 数据，无需自行构建任何基础设施即可使其可搜索，”Nwokah 补充道。
Infino stores indexes inside the Parquet files to speed retrieval, while the underlying data remains Parquet. Iceberg and Hudi can manage the same table, and Spark, DuckDB, warehouses, and other Parquet readers can still read it. One copy of the data can therefore serve the different ways an agent needs to query it.Infino 将索引存储在 Parquet 文件内部以加速检索，而底层数据保持为 Parquet 格式。Iceberg 和 Hudi 可以管理同一个表，Spark、DuckDB、数据仓库和其他 Parquet 读取器仍然可以读取它。因此，一份数据副本就可以满足智能体查询数据的不同需求。
Infino also wanted the path from data to query to be simpler than with traditional data stores. Setting up clusters, nodes, shards, and headcount is a lot of complexity just to query data, the company argues. With Infino, developers can point at Parquet or JSON, ingest the data, then point their agent at Infino and start querying.Infino 还希望从数据到查询的路径比传统数据存储更简单。该公司认为，仅仅为了查询数据而设置集群、节点、分片和人员配置是非常复杂的。使用 Infino，开发人员只需指向 Parquet 或 JSON，摄取数据，然后将智能体指向 Infino 即可开始查询。
Who benefits from this technology?谁能从这项技术中受益？
Nwokah lists four major software engineering practitioner groups who he promises can benefit from the approach outlined here:Nwokah 列举了四个主要的软件工程从业群体，他承诺这些群体可以从这里概述的方法中受益：
- Platform or data engineers looking to get their arms around the proliferation of agents, with a far simpler and far less expensive stack to manage. There is no data lock-in (claims Nwokah), as their data stays in their bucket and is readable by any Parquet reader. Agents can combine both structured and unstructured data through one interface.平台或数据工程师：他们希望通过更简单、维护成本更低的技术栈来应对智能体的激增。据 Nwokah 声称，不存在数据锁定问题，因为数据保留在他们自己的存储桶中，并且可被任何 Parquet 读取器读取。智能体可以通过一个接口结合结构化和非结构化数据。
- Security engineers looking to put a proper governance layer around their agents. With a single copy of data, the auth policies an agent runs under are enforced centrally and deterministically rather than across a set of MCP gateways and SaaS connectors.安全工程师：他们希望在智能体周围建立适当的治理层。通过单一的数据副本，智能体运行所依据的身份验证策略可以被集中且确定性地执行，而不是跨越一系列 MCP 网关和 SaaS 连接器。
- ML and data science teams doing feature exploration. Infino searches, filters, joins and aggregates over both structured and unstructured data, cutting the time between messy data and clean features.机器学习和数据科学团队：他们正在进行特征探索。Infino 可以对结构化和非结构化数据进行搜索、过滤、连接和聚合，缩短了从混乱数据到干净特征之间的时间。
- Software developers building agent workflows across code, logs, CI output, issues and other engineering data. Instead of giving the agent a different retrieval system for each corpus, they can give it one query interface across all of them.软件开发人员：他们正在跨代码、日志、CI 输出、问题和其他工程数据构建智能体工作流。他们不必为每个语料库提供不同的检索系统，而是可以为它们提供一个统一的查询接口。
With its current architecture, Nwokah claims that Infino is designed to be roughly 10× cheaper than traditional search or analytics infrastructures. Infino’s own published comparison puts it at about 10.5× cheaper than Elasticsearch and 23× cheaper than OpenSearch for the workload it tested.Nwokah 声称，凭借目前的架构，Infino 的设计成本大约比传统的搜索或分析基础设施便宜 10 倍。Infino 发布的对比数据显示，在其测试的工作负载中，其成本比 Elasticsearch 便宜约 10.5 倍，比 OpenSearch 便宜 23 倍。
Multi-billion-document use cases 数十亿文档的用例
Nwokah confirms that the company is already supporting multi-billion-document use cases in people search, document processing, product analytics, and security. Nwokah 证实，该公司已经在人员搜索、文档处理、产品分析和安全领域支持数十亿文档规模的用例。
“We are particularly optimized for teams with agent use cases that are already on Parquet or considering migrating to Parquet, or alternatively who are looking to replace their search tool with something easier to manage and more scalable,” he concludes.“我们特别为那些已经在 Parquet 上运行、正在考虑迁移到 Parquet，或者正在寻找更易于管理且更具可扩展性的搜索工具替代方案的智能体用例团队进行了优化，”他总结道。
One of the organization’s customers, which Infino did not name, is putting petabytes of data from across their entire company on Infino, according to the company. The Parquet sits in their bucket and is still read by existing tools across different teams. Meanwhile, the product and engineering teams get the scalable agent infrastructure they need to launch AI-native solutions for their customers. Developers can then control their agents from a single interface.据该公司称，其客户之一（未透露名称）正在将整个公司范围内 PB 级的数据放入 Infino 中。Parquet 数据保留在他们的存储桶中，并且仍然可以被不同团队的现有工具读取。同时，产品和工程团队获得了为客户推出 AI 原生解决方案所需的可扩展智能体基础设施。开发人员随后可以从单一界面控制他们的智能体。
Infino CEO Nwokah formed the company alongside co-founder and head of engineering Vinay Kakade, co-founder Asif Makhani, and chief architect Murali Krishna.Infino 首席执行官 Nwokah 与联合创始人兼工程主管 Vinay Kakade、联合创始人 Asif Makhani 以及首席架构师 Murali Krishna 共同创立了该公司。
The group spent years building and operating ML and search systems across Amazon, Google, and LinkedIn, including AWS OpenSearch. The changing nature of agentic data querying is their rationale for building Infino: to enable any agent to query data at scale, in one query, over one copy of the data.该团队在亚马逊、谷歌和 LinkedIn（包括 AWS OpenSearch）构建和运营机器学习与搜索系统方面拥有多年经验。智能体数据查询性质的转变是他们构建 Infino 的理由：使任何智能体都能在单一数据副本上，通过一次查询，大规模地查询数据。
