公司指南

作家企業人工智慧

WRITER is a San Francisco company, founded in 2020, that sells a full-stack generative AI platform to large businesses: its own Palmyra language models, a graph-based retrieval system called Knowledge Graph, and tools for building and monitoring AI agents.

聆聽此指南AI 語音敘述 · ~8 min

本頁閱讀時間 9 分鐘
  1. 概述
  2. 重點摘要
  3. 深入探討
  4. Grounding a policy question-answering agent in company documents
  5. 現實世界的實施
  6. 作家企業人工智慧的未來
  7. 付諸實踐
  8. 資料來源與延伸閱讀
  9. 不斷探索
  10. 常見問題

概述

It matters as one of the few enterprise AI vendors that trains its own frontier-scale models rather than reselling another lab's.

重點摘要

  1. WRITER is an enterprise AI company founded in 2020 by May Habib and Waseem AlShikh that trains its own Palmyra language models.
  2. Palmyra X5, released in April 2025, has a 1 million token context window and launched at $0.60 per million input tokens.
  3. Knowledge Graph is WRITER's graph-based retrieval-augmented generation system for grounding answers in a company's own files and web pages.
  4. AI HQ, launched April 10, 2025, combines a low-code Agent Builder, agent monitoring and a library of more than 100 prebuilt agents.
  5. WRITER raised a $200 million Series C at a $1.9 billion valuation in November 2024, bringing total funding to $326 million.
  6. WRITER states that it does not train its models on customer data, inputs or outputs.

深入探討

WRITER, styled in capitals by the company, is an enterprise software company that builds generative AI for large organizations. It was founded in 2020 by May Habib, its chief executive, and Waseem AlShikh, who had earlier co-founded Qordoba, a company that helped businesses localize products for new markets, according to TechCrunch. Its platform now centers on building AI agents, programs that use a language model to plan and carry out multi-step tasks with tools. Its pitch is a full stack: WRITER trains its own models, runs its own retrieval system and provides the application layer, so a company buys one integrated system rather than assembling parts from several vendors.

The model layer is the Palmyra family of large language models, launched in 2023. A large language model (LLM) is a neural network trained on vast amounts of text to predict the next token, a word or word fragment, and then tuned to follow instructions. In October 2024 WRITER released Palmyra X 004, which it said was trained almost entirely on synthetic data, text generated by other models rather than collected from people, for about $700,000 in compute. In April 2025 it released Palmyra X5, an adaptive reasoning model with a 1 million token context window, meaning it can take in hundreds of thousands of words in one prompt. WRITER said X5 was trained with synthetic data for about $1 million in GPUs, supports tool calling, image inputs and more than 30 languages, and scored 19.1% on OpenAI's MRCR 8-needle long-context test. WRITER's developer documentation now lists Palmyra X6, also with a 1 million token window, for agentic work, and schedules X4 and X5 for deprecation on December 14, 2026.

The retrieval layer is Knowledge Graph. Retrieval-augmented generation (RAG) is the technique of fetching relevant passages from a company's own documents and placing them in the model's prompt so answers are grounded in that material. Most RAG systems store text chunks as vectors, lists of numbers that represent meaning, and fetch the chunks whose vectors are most similar to the question. WRITER describes Knowledge Graph as graph-based RAG, which it claims is more accurate than vector-based retrieval. In a graph approach, entities such as products, people and policies and the relationships between them are stored as nodes and edges, so retrieval can follow connections rather than relying only on similarity. The documentation shows the workflow: create a graph with a name and a description of its purpose, upload files or add website URLs, scope it to particular teams, and attach it to an agent or query it through the API.

The application layer is AI HQ, launched on April 10, 2025. It combines a low-code, drag-and-drop Agent Builder for IT and business teams, observability tools that monitor agent activity in real time and flag anomalies, and a library of more than 100 prebuilt agents for finance, healthcare, retail and other functions. WRITER says the platform is powered by Palmyra models and graph-based RAG, and that agents can be grounded through Knowledge Graph and document uploads. Customers named in WRITER and AWS materials include Accenture, Intuit, Mars, Marriott, Prudential, Qualcomm, Salesforce, Uber and Vanguard.

The company's funding reflects investor interest in this approach. In November 2024 it raised a $200 million Series C at a $1.9 billion valuation, co-led by Premji Invest, Radical Ventures and ICONIQ Growth, bringing total funding to $326 million, according to TechCrunch. Participants included the venture arms of Salesforce, Adobe, IBM, Workday and Citi, as well as Accenture. TechCrunch reported that WRITER said it serves hundreds of clients but did not disclose revenue.

Compared with the alternatives, WRITER sits between two common enterprise choices. One is buying a general assistant from a frontier lab, such as an enterprise chatbot plan, and connecting it to company data. The other is building in-house on a cloud platform with open or proprietary models plus a separate vector database and orchestration framework. WRITER's argument is that owning the model, retrieval and agent layers together gives better accuracy and control, as in its claim that graph-based retrieval beats vector retrieval. Those are vendor claims. Independent comparisons of graph-based and vector-based RAG depend heavily on the data and questions, so buyers should test on their own documents.

Limitations follow from the category. Like every LLM, Palmyra models can produce fluent but wrong answers, and RAG reduces but does not eliminate this when retrieval misses the right passage or the source documents are themselves out of date. A 19.1% score on an 8-needle long-context test shows that a 1 million token window does not mean reliable recall of every detail in it. Agents that take actions in business systems add risks of wrong or unintended actions, which is why observability and permission scoping matter. Model churn is real too: several specialized Palmyra models, including Palmyra Med and Palmyra Fin, were deprecated on July 13, 2026, so integrations must plan for migrations.

In practice, adopting WRITER means an enterprise contract for the platform, or API access billed per token. The developer documentation lists Palmyra X6 at $2 per million input tokens and $8 per million output tokens, and X5 at $0.60 and $6. WRITER's security page states that it does not train its models on customer data, inputs or outputs, describes a zero data retention approach with configurable deletion schedules, offers dedicated private cloud deployment, and lists SOC 2 Type II and ISO/IEC 27001, 27701 and 42001 certifications. Data governance teams should still review current terms, since available controls depend on contract and configuration.

04工作範例

Grounding a policy question-answering agent in company documents

  1. Collect a small set of current internal policy PDFs and note three questions whose correct answers you already know.

  2. Create a Knowledge Graph through the WRITER API with a name and a description that says it holds HR policies.

  3. Upload each file through the files API and attach it to the graph, then check that every file shows as processed.

  4. Attach the graph to a no-code agent, or query it through chat completions with the graph as a retrieval tool.

  5. Ask your three test questions and compare each answer and its cited passages with the known correct answer.

  6. Add one outdated policy file, ask again, and observe whether the agent cites the old or the new version.

它顯示了什麼

A working policy agent and a small accuracy check. If answers cite the right passages for all three questions, retrieval is working; if the outdated file wins, remove stale documents, since retrieval cannot tell which version is current.

現實世界的實施

Building an internal research agent in WRITER's AI HQ Agent Builder, a low-code tool launched in April 2025 that can call systems such as Salesforce, Workday, Microsoft, Adobe and Atlassian.

Grounding an agent's answers in company documents by creating a Knowledge Graph through the API, uploading files and website URLs, and attaching the graph to a no-code agent.

Calling Palmyra X5 through Amazon Bedrock, where AWS listed it at launch in April 2025 at $0.60 per million input tokens and $6 per million output tokens with a 1 million token context window.

Starting from one of more than 100 ready-made agents in WRITER's library, such as an FDA guidance agent or a company financials research agent, whose early beta users included Uber, Salesforce and Franklin Templeton.

Running a long-document workflow, such as reviewing a large contract set in one prompt, using Palmyra X5, which AWS said processes a full 1 million token prompt in about 22 seconds.

作家企業人工智慧的未來

WRITER's recent releases show where it is heading. AI HQ in April 2025 shifted the product from writing assistance toward agents, and the developer documentation now positions Palmyra X6 for agentic work and long-running enterprise workflows. The April 2025 X5 launch emphasized a 1 million token context window and fast tool calling for multi-step agents. Open questions are common to enterprise AI: whether agents can be trusted to take actions in core business systems without close supervision, how well very long context windows support accurate recall given modest long-context benchmark scores, and how vendors prove the accuracy advantages they claim for their retrieval methods. The consolidation of specialized models into general ones, seen in the July 2026 deprecation of Palmyra Med, Fin, Creative and Vision, suggests that domain needs are increasingly served by one general model plus company data rather than separate specialist models. How WRITER competes as frontier labs sell their own enterprise agent platforms remains to be seen.

付諸實踐

為什麼這很重要

  • 供應商策略. 供應商路線圖會影響您的團隊接下來可以建立的功能。
  • 成本與預算. 商業條款和部署選項會影響長期成本和風險。
  • 風險與安全. 公司激勵措施塑造了產品預設、安全態勢和開放性。

留意

  • 發佈公告可能會超過實際生產工作流程的穩定性。
  • API 定價或政策轉變可能會在一夜之間打破假設。
  • 單一供應商依賴性增加了鎖定和遷移成本。

後續步驟

  1. 使用您自己的任務和資料集評估提供者。
  2. 在整合之前查看隱私、安全和法律條款。
  3. 維護跨模型或供應商的後備計劃。
  4. 監控發行說明,以便路線圖的變更不會讓團隊感到意外。

資料來源與延伸閱讀

  1. TechCrunchGenerative AI startup Writer raises $200M at a $1.9B valuation
  2. Amazon (AWS press release)Writer Releases New Adaptive Reasoning LLM Palmyra X5 With 1M Context Window to Scale Next Era of Enterprise AI Agents
  3. WRITERWRITER launches AI HQ press release
  4. WRITER developer documentationChoose a model
  5. WRITER developer documentationCreate and manage a Knowledge Graph
  6. WRITERSecurity and trust

不斷探索

免費電子報

獲取每日 AI 簡報

每個工作日上午提供三則經過驗證的 AI 新聞,以淺顯易懂的英文撰寫。永久免費,無廣告。

每個工作日一封電子郵件。一鍵取消訂閱。我們絕不出售或分享您的地址。

測試一下自己

Take the Writer Enterprise AI quiz

每個答案皆有即時回饋,通過課程後可獲得具備可驗證 ID 的可分享證書。

開始測驗

支持免費的 AI 教育。AI Understanding 是一個 501(c)(3) 非營利組織——絕無廣告,絕無付費牆。 捐款

常見問題

What is Writer Enterprise AI?

WRITER is a San Francisco company, founded in 2020, that sells a full-stack generative AI platform to large businesses: its own Palmyra language models, a graph-based retrieval system called Knowledge Graph, and tools for building and monitoring AI agents. It matters as one of the few enterprise AI vendors that trains its own frontier-scale models rather than reselling another lab's.

Writer 與許多其他生成型人工智慧公司有何不同?

Writer 建立並訓練自己的 Palmyra LLM,而不是簡單地轉售第三方模型。

誰是 Writer 的主要客戶群?

Writer 專注於企業客戶,為業務工作流程提供安全、受管控的 AI。

作家強調什麼檢索方法來減少幻覺?

Writer 使用基於圖形的檢索增強生成來根據結構化公司資料中的回應。

What should an enterprise check before relying on Writer data-use or training commitments?

Writer’s published materials describe its own models and default retention settings, while administrators can enable third-party models; customers should check current terms and configuration.

作家的共同創辦人是誰?

Writer 由 May Habib 和 Waseem Alshikh 於 2020 年共同創立。