Lượm lặt

Glean is an enterprise software company that makes the information scattered across a company's apps searchable and answerable in one place. It was founded in 2019 in Palo Alto, California, by Arvind Jain, who is its CEO, with Piyush Prahladka, Tony Gentilcore and T.R. Vishwanath. Its products are Glean Search, a conversational tool called Glean Assistant, and Glean Agents, a platform for building and running AI agents, meaning software that carries out multi-step tasks rather than only answering a question. All three sit on the same foundation: an index of the customer's own documents, messages, tickets and people.

The first step is connection. Glean's documentation describes connectors, small integrations for each app, that sync content and metadata into the customer's search index: titles, bodies, comments, attachments, owners, timestamps and app-specific fields. Connectors also copy each app's access control lists, the records of who may open which item, and fetch people and group data so the same person can be matched across systems. Where an app supports it, they collect activity signals such as views, edits and shares. Sources are kept current with incremental crawls and, in some cases, webhooks or push APIs, while fast-changing sources such as email or calendar can be fetched live at query time instead of being indexed.

The second step is ranking. In a 2023 engineering post, Glean's authors argued that vector search alone, which matches the meaning of a query to the meaning of a document using numeric embeddings, is not enough for workplace content. They described a hybrid approach that combines classic keyword retrieval with neural embeddings, followed by a reranking step that blends dozens of signals, including keyword match, semantic similarity, freshness and personalization, into one relevance score. Personalization comes from what Glean now calls the Enterprise Graph, a knowledge graph whose nodes are content, people and activity, and whose edges include links between documents, people who work together, and who created, edited or shared what.

The third step is generation. Glean Assistant uses retrieval-augmented generation (RAG): it retrieves the most relevant items the user is allowed to see, passes them to a large language model, and asks the model to answer from that text. The same 2023 post explains why Glean prefers this to fine-tuning, which means further training a model on company data. Fine-tuned knowledge goes stale, cannot respect per-user permissions, cannot point to its sources, and can raise the rate of hallucination, the model stating things that are not true. With RAG, the index can be updated continuously, permissions are enforced before the model sees anything, and the cited documents can be inspected. The post also describes using a language model for query planning, turning one request into several searches, such as GitHub pull requests plus Slack messages.

The company's milestones are well documented. Glean raised a $15 million Series A in March 2019, co-led by Kleiner Perkins and Lightspeed, and launched its search product in September 2021. A $100 million Series C led by Sequoia Capital in May 2022 valued it at $1 billion. Valuations then rose quickly: $2.2 billion in February 2024, $4.6 billion after a Series E of more than $260 million in September 2024, and $7.2 billion after a $150 million Series F led by Wellington Management on June 10, 2025, according to TechCrunch. Glean Agents first appeared in February 2025 and became generally available on May 20, 2025, alongside a model hub giving agents access to 15 large language models through Amazon Bedrock, Google Vertex AI and Azure OpenAI, and a server for the Model Context Protocol (MCP), an open standard for connecting AI agents to tools. On December 8, 2025, Fortune reported that Glean had reached $200 million in annual recurring revenue (ARR), double the $100 million it reported nine months earlier, counting only software subscriptions.

Glean competes in a crowded field. Large productivity suites build assistants that work best on their own content, such as documents and mail stored in one vendor's cloud. Glean's pitch is the opposite: a neutral layer across many vendors' apps. Companies can also build their own RAG systems with open-source vector databases and a model API, which gives full control but leaves them to solve connectors, permission syncing, identity matching and ranking themselves, the parts Glean's documentation spends most of its effort on. The trade-off is cost and dependence on one vendor against the engineering time of building and maintaining the same pipeline in-house.

The limitations follow from the design. Answers can only be as good as what is connected and indexed, so a missing connector or a stale crawl means a missing or outdated answer. Permission mirroring is only as accurate as the source app's own sharing settings: if a sensitive file is shared too widely in the source, it will be searchable for everyone who has access, which is why Glean's documentation tells administrators that opening a result may trigger a second check in the source system and points them to governance tools. Activity signals and available permissions differ by connector. RAG reduces hallucination but does not remove it, so a generated answer can still misread or misstate a correct source, which is why citations matter.

In practice, Glean is sold to organizations rather than individuals. Fortune reported that its contracts run one to three years. Its documentation offers two cloud deployment models. In Glean Hosted, Glean runs a single-tenant instance, isolated from other customers, on Google Cloud. In Customer Hosted, formerly called Cloud-Prem, Glean deploys and operates the instance inside the customer's own Google Cloud or AWS account, and the customer pays the hosting costs. SiliconANGLE reported in May 2025 that a partnership with Dell Technologies also lets enterprises run the platform on their own premises. A rollout typically involves an administrator authorizing each connector, checking that identities and groups map correctly, and testing permission behavior with test accounts before opening the tool to all employees.