内容来源：https://aibusiness.com/agentic-ai/glean-s-model-aims-redefine-enterprise-search-with-ai

内容总结：

谷歌云特约报道：企业部署生成式AI应从信息检索入手

随着生成式AI技术加速渗透企业市场，业界专家指出，企业应首先聚焦那些能提升人类信息处理体验的应用场景。企业搜索服务商Glean近日发布的新型AI搜索模型Waldo，揭示了行业专业化模型的重要性，同时也凸显了搜索厂商面临的市场变局。

Waldo于4月28日发布，是一款强化学习驱动的智能代理搜索模型。其工作流程是：在通用大模型介入之前，Waldo先执行搜索任务，完成后将结果交给具备推理能力的前沿模型进行后续处理。据Glean介绍，Waldo负责搜索，前沿模型则负责生成最终回应。该模型基于英伟达的Nemotron 3 Nano开源模型构建。

Glean借此向企业传递一个核心理念：由于大多数智能代理在执行任务前都需要先搜索信息，因此搜索是所有代理应用的基础。相较于直接使用通用大模型，采用专门的搜索模型进行信息检索，再用推理模型进行逻辑分析和内容提取，效率更高、成本更低。Glean表示，这一做法旨在帮助企业节省使用大型前沿模型的高昂费用。

Forrester Research分析师罗恩·柯伦指出：“随着企业代理应用迭代至第二、第三代，管理者对实际成本、效果和性能的关注度日益上升。在整体代理工作流中，针对特定任务的定向模型正变得越来越有吸引力。”

作为一家以企业搜索为核心业务的厂商，Glean在智能代理AI时代转向帮助企业实现最精准、最直接的信息检索，是顺理成章的路径。其他搜索厂商如Moveworks和Genspark也纷纷向代理搜索方向转型。

Futurum Group分析师布拉德利·希明表示：“在特定领域拥有专长的公司——无论是通过软件、专业服务还是其他服务——都可以利用定向模型将这些知识转化为商业价值。”他补充说，这些厂商还可以像Glean那样对模型进行微调或精简提炼。

不过，柯伦认为，Glean的专业化模型策略使其在众多厂商中脱颖而出，可能成为这家总部位于帕洛阿尔托的公司的差异化优势。“由于Glean以SaaS平台运行，它在理解企业用户搜索习惯方面，可能比其他服务商更有深度。”

但柯伦也提醒，搜索与检索的实际情况可能不像Glean描述的那样简单。例如，数据检索可以通过API或MCP连接器等多种方式实现，而前沿模型在直接文件访问方面也已相当成熟，能够无需过多用户输入即可读取、导航并分析文件。“就在六七个月前，前沿模型还很难通过浏览企业目录来找到匹配查询的答案，但现在，大多数前沿模型及其配套工具都已具备这一能力。”

Glean等搜索厂商面临的另一挑战在于，企业的需求正在变得越来越多元化。柯伦观察到：“过去几年，企业对高端企业搜索的期望从‘能给我一个企业版谷歌吗’，悄然变成了‘能给我一个企业版ChatGPT吗？’”。需求的转变，加上搜索已成为智能代理AI的基础环节，正迫使Glean这样的厂商重新审视自身的价值定位。

尽管如此，希明认为Waldo的发布对整个AI行业而言是一个积极信号。“像Glean这样带着自有解决方案进入市场的厂商越多，整个行业的发展就越健康。”

中文翻译：

由谷歌云赞助
选择您的首个生成式AI应用场景
要开始使用生成式AI，首先应聚焦于能够改善人类与信息交互体验的领域。

沃尔多强调了企业内部对特定任务模型的需求，同时揭示了搜索厂商正面临的格局变化。
企业搜索厂商Glean推出的新型智能体AI搜索模型，凸显了企业使用领域专用及特定任务模型的重要性，但也折射出搜索厂商面临的挑战。

4月28日发布的沃尔多，是一种强化学习智能体搜索模型，在基础模型启用前运行。其职责是搜索，完成后将任务交给基础模型进行检索与推理。据Glean介绍，沃尔多负责搜索，基础模型生成回答。该模型基于英伟达Nemotron 3 Nano开源模型构建。

通过沃尔多，Glean向企业传达一个理念：由于许多智能体在执行任务前必须先搜索信息，从这层意义上看，搜索是所有智能体应用的基础。
因此，更优的方案是使用专用搜索模型执行搜索，再采用推理模型进行推理与检索。Glean的方法基于一个普遍认知：领域专用模型往往比通用基础模型更高效。Glean表示，其目标是为企业节省使用大型基础模型的成本。

福雷斯特研究公司分析师罗文·柯伦指出：“当企业开始迭代至第二、第三代智能体时，它们会更关注这些技术的实际成本、效果与性能。在整体智能体工作流与执行中，这类针对性更强的模型吸引力正日益增强。”

由于Glean的核心业务是企业搜索，在智能体AI时代转向帮助企业实现最精准、直接的信息检索，是顺理成章的选择。其他搜索厂商如Moveworks和Genspark，也已被迫向智能体搜索转型。

Futurum集团分析师布拉德利·希明表示：“在特定领域具备专长的企业……无论是通过软件、专业服务还是其他服务，都能利用定向模型将知识转化为收益。”他补充道，这些厂商还能对模型进行微调或蒸馏——Glean正是通过使用Nemotron 3 Nano实现了这一点。

柯伦认为，Glean的专用模型独树一帜，可能成为这家帕洛阿尔托厂商的差异化优势。
“由于Glean以SaaS平台形式运行，在理解企业用户搜索行为方面，可能比其他服务商更具深度优势。”他表示。

他补充说，搜索与检索可能并不像Glean描绘得那么简单——用一个专用模型搜索，另一个模型检索与推理。首先，数据检索方式多样，如使用API或MCP连接器。其次，基础模型已能熟练实现直接文件访问——这种方法允许模型在无需用户过多输入的情况下读取、导航和分析文件。柯伦指出，六个月或八个月前，基础模型还不太可能通过浏览企业目录来匹配查询请求。
“如今，大多数基础模型以及围绕它们构建的工具框架都已具备这些能力。”柯伦继续说道。

Glean及其他搜索厂商面临的另一挑战是，企业需求日益多元化。
“过去几年，企业需求已从‘能否给我一个企业版谷歌’（谈论高端企业搜索时）演变为‘能否给我一个企业版ChatGPT？’”柯伦说。他分析道，企业诉求的变化，加之搜索成为智能体AI的基础环节，意味着Glean这类厂商需重新审视自身价值主张。

不过，希明认为，沃尔多的发布对整个AI行业而言是一个积极信号。
“像Glean这样带着自有解决方案进军市场的厂商越多，整个行业的发展就越有利。”他表示。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Waldo highlights the need for task-specific models within enterprises while revealing the changing landscape search vendors face.
Enterprise search vendor Glean’s new agentic AI search model underscores the importance of enterprises using domain- and specialized-task-specific models, but it also illuminates the challenge facing search vendors.
Waldo, released on April 28, is a reinforcement learning agentic search model that runs before a frontier model is used. Its job is to search, and once it has finished, Waldo hands off to a frontier model that retrieves and reasons. According to Glean, Waldo searches and the frontier model generate the response. The model was built on Nvidia’s Nemotron 3 Nano open model.
With Waldo, Glean is making the case to enterprises that, since many agents first have to search for information before performing the tasks that they’re being asked to do, in that sense, search underlies all agentic applications.
So, it is better to have a search-specific model that performs the search, then use a reasoning model for reasoning and retrieval. Glean’s approach is based on the popular idea that domain-specific models are often more effective than generic frontier models and Glean says it aims to help enterprises save on the cost of using large frontier models.
“As enterprises start to get to the second and third generation of their agents, they’re starting to get more concerned about what the actual cost, effectiveness and performance is of these things,” said Rowan Curran, an analyst at Forrester Research. “There is an increasing appeal for these more directed models for tasks within overall agentic workflows and executions.”
With Glean being at its core an enterprise search vendor, the pivot to helping enterprises achieve the most accurate, direct retrieval of information is natural in the age of agentic AI. Other search vendors, such as Moveworks and Genspark, have had to move toward agentic search as well.
“Companies that have some expertise in a given area … be that through software or through professional services or through any service are in a position to translate that knowledge into cash using a targeted model," said Bradley Shimmin, an analyst at Futurum Group. He added that those vendors could also fine-tune or distill a model, which is what Glean did by using Nemotron 3 Nano.
However, Glean’s specialized model stands out and could be a differentiating factor for the Palo Alto-based vendor, Curran said.
“They may have some advantage given that they run as a SaaS platform in terms of understanding the way that their enterprise users search more deeply than some of the other providers,” he said.
He added that search and retrieval will likely not be quite as simple as Glean has presented it, with a specialized model for search and another for retrieval and reasoning. For one, there are different ways to retrieve data, such as using APIs or MCP connectors. Also, frontier models have become adept at direct file access, a method in which the model can read, navigate and analyze files without much user input. Curran added that, six or eight months ago, it would have been unlikely for frontier models to look through an enterprise directory to find an answer that matched the query presented.
“Now you do have these capabilities in most of the frontier models as well as the harnesses that are being built around them,” Curran continued.
Another challenge that Glean and other search vendors face is that enterprises are making increasingly diverse demands.
“Over the past couple of years, we have gone from … ‘Can you give me Google for my enterprise’ when they were talking about high-end enterprise search, which is now transformed into ‘Can you give me ChatGPT for my enterprise?’” Curran said. The change in what enterprises are looking for, along with search being a fundamental part of agentic AI, means that vendors like Glean are reconsidering their value proposition, he added.
However, Waldo’s release is a promising sign for the entire AI industry, Shimmin said.
“The more vendors like Glean that push into these marketplaces with their own solutions, the better off the whole industry is,” he said.