Announcements（公告）¶
REMINDER: You can now sign up for the Foundry Newsletter to receive a summary of new products, features, and improvements across the platform directly to your inbox. For more information on how to subscribe, see the Foundry Newsletter and Product Feedback channels announcement.
Introducing AIP Logic [GA]¶
Date published: 2023-12-14
AIP Logic is a no-code development environment for building, testing, and releasing functions powered by large language models. With AIP Logic, you can build feature-rich AI-powered functions that leverage the Ontology without the complexity typically introduced by development environments and API calls. Using Logic’s intuitive interface, application builders can engineer prompts, test, evaluate and monitor, set up automation, and more.
You can use AIP Logic to automate and support your critical tasks, whether that’s connecting key information from unstructured inputs to your Ontology, resolving scheduling conflicts, optimizing asset performance by finding the best allocation, or reacting to disruptions in your supply chain.
AIP Logic's landing page.
Access AIP Logic¶
AIP Logic can be accessed from the platform’s workspace navigation bar or by using the quick search shortcuts CMD + J (macOS) or CTRL + J (Windows). Alternatively, you can create a new logic function from your Files by selecting +New and then choosing AIP Logic, as shown below.
+ New dropdown menu.
What's on the development roadmap?¶
The following features for AIP Logic are currently in development:
- Logic Assistant: Obtain AI-assisted help in writing your prompts and check for missing tools and data. Build more reliable Logic functions and benefit from faster iteration on prompts and reduced error rates.
- Versioning and Branching: Create, manage, and merge different versions and branches of your AIP Logic functions.
- Evaluations: Set up an evaluation and testing framework with your Ontology definition to measure the efficacy of your Logic functions.
Get Started with AIP Logic¶
To get started, visit Getting started or learn more about Core concepts.
Introducing Derived Series [Beta]¶
Date published: 2023-12-14
Derived series, now available as a beta feature on request, provides a new way for time series users to save and replicate calculations performed on time series data within their Quiver analysis. Storing derived series as Foundry resources allows the logic to be shareable and linked to the Ontology, enabling derived series to act like raw time series, calculated on demand without needing additional storage or repeated calculations.
Contact your Palantir representative for enablement.
A derived series is the combination of transformations and/or calculations on raw time series data, saved as a Foundry resource to be reused in a variety of workflows.
Derived series creation¶
Users can now save transformations on time series data from their Quiver analysis as derived series. Time series cards such as time series formula, rolling aggregate, et cetera now have a Save derived series option that converts the entire logic tree in a Quiver analysis to a codex template that can be executed at runtime.
Learn more about creating derived series.
Saving derived series after details and object type select configuration.
Derived series management¶
Furthermore, you can access the derived series management page to manage both the derived series resource as well as codex template. View relevant information about the derived series as well as modify the logic, metadata and republish a new version of the derived series template.
Learn more about managing derived series.
Derived series details in one view.
What's on the development roadmap?¶
The step to save derived series to the Ontology currently necessitates users to manually build and maintain pipelines that link the series to root or sensor object types in order to facilitate the broader use of derived series in analytical or operational applications, comparable to raw time series. We are actively developing full automation in building derived series workflows to ensure a seamless user experience by eliminating the need to manually manage Ontology pipelines.
Get started with derived series¶
For more information, review the following related documentation:
Introducing HyperAuto V2 automatic sync creation [GA]¶
Date published: 2023-12-12
The automatic sync creation feature for HyperAuto V2 pipelines is now generally available, allowing you to select any visible table within your source when configuring a HyperAuto pipeline. In the case where a sync does not already exist, HyperAuto will intelligently create one for you to configure as you wish.
Use tables without data connections syncs as inputs¶
You can now choose tables that do not have data connection syncs as inputs in the Input configuration step. If syncs exist for a chosen input, HyperAuto will default to using the most recently run sync. You can re-configure a selected input via the Configure input table option via the pencil icon and can choose either a different existing sync to use or for a new sync to be created.
Input configuration window now allows tables without data connection syncs as inputs.
From the Input table settings panel, select Create a new sync from your SAP Source for this table, then Save.
Create a new sync option available from the Input table settings panel.
Once the HyperAuto pipeline has been created, you can see how many syncs are being set up (also known as “initialized” in the interface) from the Overview page.
Overview page showing syncs being automatically generated.
HyperAuto will create and run the syncs shown on the Overview page before deploying the pipeline logic.
Learn more from Getting started with HyperAuto.
Introducing the Free-form Analysis widget¶
Date published: 2023-12-12
The Free-form Analysis Workshop widget enables users to independently investigate object data with flexibility within the framework of a Workshop application. Now generally available, the widget allows users to benefit from simple path-based analysis interface driven by Quiver's robust suite of features.
Investigate data directly within Workshop applications with the Free-form Analysis widget.
Seamless data exploration within Workshop applications¶
With the Free-form Analysis widget, users can explore the object data within the Workshop application, and share their investigation with others to reduce duplicative work and enhance their workflow. Previously, when users wanted to dig into data in a Workshop application, they were required to use Contour or Quiver to support their investigation.
Now, with the Free-form Analysis widget, the following use cases will benefit:
- Data exploration: Explore data and create bespoke investigations within an existing Workshop application.
- Root cause investigation: Given an alert, users can build upon a pre-defined set of visualizations and drill into the data in whichever direction is most relevant for their investigation — including exploring linked object types.
- Application prototyping: Builders can review saved analyses to understand common off-roading patterns, then incorporate these into the production workflow.
- Cohort creation: Users can drill down to create custom cohorts, which can be saved as a group for use elsewhere in the application.
Get started with Free-form Analysis widgets¶
To begin using the Free-form Analysis widget, simply search for the widget in the Workshop widget homepage. Then, to configure:
- Provide an Input object set to serve as the base input to the analysis.
- Define how the widget should be configured when there are no cards in a path by setting the Empty state header and the Empty state description.
- Determine the Output object set, which saves the output object set for reference elsewhere within Workshop.
- Optionally, toggle on Enable path saving which will copy individual cards in a path to be added to a Notepad document. Users should note that cards can only be copied if they are in a saved analysis path.
Analysis paths can be saved as either public or private for future reference and can also be opened in Quiver, or copied to Notepad documents.
Free-form Analysis widget configuration
To learn more, see the Free-form Analysis widget documentation.
Introducing expanded LLM integration for AIP developer capabilities¶
Date published: 2023-12-07
We are excited to announce expanded support for large language model (LLM) usage in AIP developer capabilities, including:
- AIP developer capabilities permission management in Control Panel
- Language models in Python transforms
- Language models in Functions on Objects
- Text to Embeddings board in Pipeline Builder
AIP developer capabilities permission management in Control Panel¶
Administrators of AIP-enabled stacks can now take advantage of a new AIP Settings page in Control Panel to manage access to LLMs within AIP developer workflows. From the page, administrators can enable, disable, and choose which users groups are able to build workflows on top of custom and Palantir-provided LLMs.
AIP developer capabilities permission management in Control Panel
Use language models in Python transforms¶
Pipeline authors using Python transforms can now seamlessly build data pipelines that make use of Palantir-provided LLMs and embedding models. Simply take advantage of the Python SDK included in the palantir_models Python package.
Incorporate LLM in your data pipelines with the palantir_models Python package
Example setup¶
The code snippet below demonstrates how a pipeline developer can implement transforms with OpenAI's GPT-4 ↗ in their logic, effortlessly tapping into the potential of LLMs for any data pipeline existing on the platform.
from transforms.api import transform, Input, Output
from palantir_models.transforms import OpenAiGptChatLanguageModelInput
from palantir_models.models import OpenAiGptChatLanguageModel
@transform(
source_df=Input("/path/to/input/dataset")
gpt_4=OpenAiGptChatLanguageModelInput("ri.language-model-service..language-model.gpt-4_azure"),
output=Output("/path/to/output/dataset"),
)
def compute(ctx, source_df, gpt_4: OpenAiGptChatLanguageModel, output):
...
For more information on LLMs available in-platform and usage, consult Palantir-provided models within transforms in the documentation.
Interact with language models in Functions on Objects¶
Users can now create custom logic using Palantir-provided language models with Typescript functions, facilitating workflows like summarization, Q\&A, semantic search, and more. The updated model import panel supports both Palantir-provided and custom-authored models. Typescript classes will be generated for all imported LLMs, providing an intuitive interface for using LLMs within user-authored functions.
A sample of Palantir-provided LLMs - availability may differ between stacks
Example usage¶
For example, the following code demonstrates writing a custom Typescript function using the GPT_4 model to run a simple sentiment analysis on the provided text.
import { GPT_4 } from "@foundry/models-api/language-models";
@Function()
public async sentimentAnalysis(userPrompt: string): Promise<string> {
const systemPrompt = "Provide an estimation of the sentiment the text the user has provided. \
You may respond with either Good, Bad, or Uncertain. Only choose Good or Bad if you are overwhelmingly \
sure that the text is either good or bad. If the text is neutral, or you are unable to determine, choose Uncertain."
const systemMessage = { role: "SYSTEM", content: systemPrompt };
const userMessage = { role: "USER", content: userPrompt };
const gptResponse = await GPT_4.createChatCompletion({messages: [systemMessage, userMessage], params: { temperature: 0.7 } });
return gptResponse.choices[0].message.content ?? "Uncertain";
}
For more information on usage and examples, review the documentation on Language models within Functions.
Text to Embeddings board in Pipeline Builder¶
Pipeline Builder now includes a powerful new Text to Embeddings board. Embeddings are dense vector representations of text designed to capture the semantic meaning of words or phrases from the text to be processed by LLMs. Embeddings convert text into a numerical form that can be processed by LLMs.
The numerical representation (embedding) allows for the comparison of textual data based on contextual similarity rather than just syntactic similarity. For instance, when comparing the words "cat," "dog," and "balloon," embeddings can help determine that "cat" and "dog" are more closely related in meaning than "cat" and "balloon", and use that understanding to drive advanced text analysis and operations.
Using embeddings in Pipeline Builder is particularly beneficial for workflows that involve semantic search as it allows LLMs to perform more effective, nuanced, and accurate searches by evaluating the vectors for similarity.
Example usage¶
This board takes a string column as input and will use the Palantir-provided text-embedding-ada-002 model to create an embedding vector.
Text to Embeddings board configuration
These embeddings can then be added to the Ontology as an Embedded vector property and used downstream in LLM-powered workflows.
Using an embedding as part of your workflow
Learn more¶
To learn more about the features described above, review the following documentation:
- Manage AIP settings with Control Panel
- Learn how to use Language models within Python Transforms
- Learn how to use Language models within Functions
- Embeddings in Pipeline Builder
Additional highlights¶
Foundry Developer Console
Date published: 2023-12-12
Export OpenAPI Specifications in Developer Console | Users can now generate and export OpenAPI specifications for their Developer Console applications as YAML files. This feature can be accessed through the renamed SDK Generation tab, previously known as Version history.
Marketplace¶
Date published: 2023-12-12
Improved Product Image Alignment and Layout in Marketplace | The Marketplace product page now features improved alignment of product images and an enhanced layout.
App Building | Workshop¶
Date published: 2023-12-12
Faster Embedded Module Loading for Large Modules | This update significantly improves the load time of very large embedded modules in Workshop. Users working with large modules will experience load times up to 5 times faster.
Data Integration | Code Repositories¶
Date published: 2023-12-12
Advanced pull request approval policies | Introducing a granular, rule-based system for configuring pull request approvals. With advanced PR approval policies, it is now possible to specify which users or groups should approve a PR based on the files modified. The policy can be edited both in YAML and using the interactive approval policy editor.
Policies are then applied to newly created pull requests. Required approvers are displayed in a new user interface, showing which rules have been satisfied and how many approvals are still required.
To get started using this feature, create a protected branch and then edit the policy from the Branches tab of the settings pane in your Code Repository.
Marketplace¶
Date published: 2023-12-12
Marketplace Modeling Integration Now Generally Available | The Marketplace modeling integration is now generally available, enabling users to package containerized executables that encapsulate various functional logic, such as machine learning, forecasting, optimization, physical models, and business rules.
Administration | Control Panel¶
Date published: 2023-12-12
Enhanced Egress Policies for AWS Hosted Stacks | Users on AWS hosted stacks can now create egress policies to S3 buckets they own in the same region, providing more control over data access. Note that a future update will improve the user experience by providing clear feedback on policy eligibility based on the bucket's region.
Object Monitoring¶
Date published: 2023-12-12
Enhanced Automation View and Edit Modes | The Object Monitoring application now features an improved view mode selector in the automation header, providing users with more options to switch between automation edit and view modes. The execution mode selection has been made more visible, and automation condition icons have been consolidated for a better user experience.
Data Integration | Code Repositories¶
Date published: 2023-12-12
Spark Module Runtimes Upgrade to Spark 3.4 | Foundry's Spark module runtimes have been upgraded to Spark 3.4, providing users with the latest performance and security enhancements. This upgrade applies to all modules greater than the following versions: Python 1.975.0, Java 1.997.0, and SQL 1.861.0. No action is required from users, as modules will be upgraded automatically.
Foundry Developer Console¶
Date published: 2023-12-12
API Token Creation in Developer Console | Users can now generate long-lived API tokens with restricted capabilities in Developer Console for personal use and CI/CD workflows. The initial capability allows installing SDKs with a scope of artifacts:read-artifacts on both the SDK artifacts repository and the Foundry SDK asset bundle. Access the Create API Token dialog in the Application SDK settings pages or through the getting started instructions.
Object Monitoring¶
Date published: 2023-12-12
Enhanced Object Monitoring Features | This update brings several enhancements to the Object Monitoring application. Users can now view failure events from the last 28 days on the landing page, providing better visibility into recent issues. Additionally, more options for deleting automations have been added, making it easier to manage and maintain your workflows. The Actions menu has been reorganized, moving automation options to a separate menu for easier access. Lastly, evaluation latency information has been improved, with clearer wording and added hints for time condition combinations on objects added and removed.
Security | Projects¶
Date published: 2023-12-05
Enhanced Project Contact Selection | Users and groups without contact details can now be added as project contacts, improving flexibility in project management. The updated interface displays the appropriate UI for users without an email address and groups without contact details.
App Building | Workshop¶
Date published: 2023-12-05
New PDF Viewer Widget with Keyword Search | Workshop's PDF Viewer widget is now generally available, featuring keyword search capabilities with text highlighting and auto-scroll on match. This enhancement improves PDF workflows through manual or variable-based inputs.
Administration | Control Panel¶
Date published: 2023-12-05
Expose Cloud Runtime Egress IPs in Control Panel | Users can now access cloud runtime egress IP addresses directly in the Control Panel, making it easier to set up ingress filtering for APIs and cloud sources. This update provides a more contextual and convenient way for users to obtain the necessary IP addresses for their cloud runtimes.
Security | Projects¶
Date published: 2023-12-05
Support for Vertex and Vortex Ontological Dependencies | Projects now support Vertex and Vortex Ontological dependencies, enhancing the Ontology access checker capabilities for more efficient and accurate Ontology management.
Data Integration | Code Repositories¶
Date published: 2023-12-05
Attachments as Function Inputs in Code Repositories | Users can now upload attachments for use as function inputs in Code Repositories, providing a more seamless workflow for using attachments in both Live Preview and published functions. To get started, add a function with an Attachment or Attachment[] input type and open the Live Preview tab to upload your attachments. You can review the documentation on input and output types available when using functions.
Ontology | Ontology Management¶
Date published: 2023-12-05
Action Log Requiredness Toggle for Ontology Management | Introducing a new Requires action log toggle in Ontology Manager. When enabled, any action that edits the specified object type will require an action log rule, ensuring that all edits to objects of that type are logged for greater visibility and traceability.
中文翻译¶
公告¶
提醒： 您现在可以订阅 Foundry 新闻通讯（Newsletter），直接在收件箱中接收平台新产品、功能及改进的摘要。有关订阅方式的更多信息，请参阅 Foundry 新闻通讯与产品反馈渠道公告。
推出 AIP Logic [正式发布]¶
发布日期：2023-12-14
AIP Logic 是一个无代码开发环境，用于构建、测试和发布由大语言模型（large language model, LLM）驱动的函数。借助 AIP Logic，您可以构建功能丰富的 AI 驱动函数，这些函数能够利用本体（Ontology），而无需处理开发环境和 API 调用通常带来的复杂性。通过 Logic 直观的界面，应用构建者可以设计提示词（prompt）、测试、评估和监控、设置自动化等。
您可以使用 AIP Logic 来自动化和支持关键任务，无论是将非结构化输入中的关键信息连接到本体、解决日程冲突、通过寻找最佳分配来优化资产性能，还是应对供应链中的中断。
AIP Logic 的登录页面。
访问 AIP Logic¶
可以从平台的工作区导航栏或使用快速搜索快捷键 CMD + J (macOS) 或 CTRL + J (Windows) 访问 AIP Logic。或者，您也可以从 文件 中选择 +新建，然后选择 AIP Logic 来创建新的逻辑函数，如下所示。
+ 新建 下拉菜单。
开发路线图上有哪些内容？¶
以下 AIP Logic 功能目前正在开发中：
- 逻辑助手（Logic Assistant）： 在编写提示词时获得 AI 辅助帮助，并检查缺失的工具和数据。构建更可靠的 Logic 函数，受益于更快的提示词迭代和更低的错误率。
- 版本控制与分支（Versioning and Branching）： 创建、管理和合并 AIP Logic 函数的不同版本和分支。
- 评估（Evaluations）： 使用本体定义设置评估和测试框架，以衡量 Logic 函数的有效性。
开始使用 AIP Logic¶
要开始使用，请访问 入门指南 或了解更多关于 核心概念 的信息。
推出派生时间序列（Derived Series）[Beta]¶
发布日期：2023-12-14
派生时间序列（Derived Series）现已作为 Beta 功能提供（需申请），为时间序列用户提供了一种在 Quiver 分析中保存和复现对时间序列数据执行的计算的新方法。将派生时间序列存储为 Foundry 资源，使得逻辑可以共享并链接到本体，从而使派生时间序列能够像原始时间序列一样按需计算，无需额外存储或重复计算。
请联系您的 Palantir 代表以启用此功能。
派生时间序列是对原始时间序列数据进行转换和/或计算的组合，保存为 Foundry 资源，可在各种工作流中重复使用。
创建派生时间序列¶
用户现在可以从 Quiver 分析中将对时间序列数据的转换保存为派生时间序列。时间序列卡片（如时间序列公式、滚动聚合等）现在具有 保存派生时间序列 选项，该选项将 Quiver 分析中的整个逻辑树转换为可在运行时执行的 codex 模板。
在配置详细信息和对象类型选择后保存派生时间序列。
管理派生时间序列¶
此外，您可以访问派生时间序列管理页面来管理派生时间序列资源和 codex 模板。查看派生时间序列的相关信息，以及修改逻辑、元数据并重新发布派生时间序列模板的新版本。
在一个视图中查看派生时间序列详情。
开发路线图上有哪些内容？¶
目前，将派生时间序列保存到本体的步骤要求用户手动构建和维护将序列链接到根对象类型或传感器对象类型的管道，以便在分析或操作应用程序中更广泛地使用派生时间序列（类似于原始时间序列）。我们正在积极开发派生时间序列工作流的完全自动化，通过消除手动管理本体管道的需要，确保无缝的用户体验。
开始使用派生时间序列¶
有关更多信息，请查阅以下相关文档：
推出 HyperAuto V2 自动同步创建功能 [正式发布]¶
发布日期：2023-12-12
HyperAuto V2 管道的自动同步创建功能现已正式发布，允许您在配置 HyperAuto 管道时选择源中的任何可见表。如果同步尚不存在，HyperAuto 将智能地为您创建一个，供您按需配置。
使用没有数据连接同步的表作为输入¶
您现在可以在输入配置步骤中选择没有数据连接同步的表作为输入。如果所选输入存在同步，HyperAuto 将默认使用最近运行的同步。您可以通过铅笔图标选择 配置输入表 选项来重新配置所选输入，并可以选择使用不同的现有同步或创建新的同步。
输入配置窗口现在允许将没有数据连接同步的表作为输入。
从 输入表设置 面板中，选择 为此表从您的 SAP 源创建新同步，然后点击 保存。
输入表设置面板中可用的创建新同步选项。
一旦 HyperAuto 管道创建完成，您可以从概览页面查看正在设置（在界面中也称为“初始化”）的同步数量。
概览页面显示正在自动生成的同步。
在部署管道逻辑之前，HyperAuto 将创建并运行概览页面上显示的同步。
推出自由形式分析（Free-form Analysis）小组件¶
发布日期：2023-12-12
自由形式分析 Workshop 小组件 使用户能够在 Workshop 应用程序的框架内灵活地独立调查对象数据。该小组件现已正式发布，允许用户受益于由 Quiver 强大功能集驱动的简单基于路径的分析界面。
直接在 Workshop 应用程序中使用自由形式分析小组件调查数据。
Workshop 应用程序中的无缝数据探索¶
借助自由形式分析小组件，用户可以在 Workshop 应用程序内探索对象数据，并与他人分享调查结果，以减少重复工作并增强工作流。以前，当用户想要深入研究 Workshop 应用程序中的数据时，他们需要使用 Contour 或 Quiver 来支持调查。
现在，使用自由形式分析小组件，以下用例将受益：
- 数据探索： 在现有 Workshop 应用程序中探索数据并创建定制调查。
- 根本原因调查： 给定一个警报，用户可以基于预定义的可视化集进行构建，并沿着与其调查最相关的任何方向深入数据——包括探索链接的对象类型。
- 应用原型设计： 构建者可以查看已保存的分析，以了解常见的偏离模式，然后将这些模式纳入生产工作流。
- 队列创建： 用户可以向下钻取以创建自定义队列，这些队列可以保存为组，以便在应用程序的其他地方使用。
开始使用自由形式分析小组件¶
要开始使用自由形式分析小组件，只需在 Workshop 小组件主页中搜索该小组件。然后，进行配置：
- 提供一个 输入对象集 作为分析的基础输入。
- 通过设置 空状态标题 和 空状态描述，定义当路径中没有卡片时小组件应如何配置。
- 确定 输出对象集，该输出对象集将被保存，以便在 Workshop 中的其他地方引用。
- 可选地，开启 启用路径保存，这将复制路径中的单个卡片以添加到记事本文档中。用户应注意，只有处于已保存分析路径中的卡片才能被复制。
分析路径可以保存为公开或私有，以供将来参考，也可以在 Quiver 中打开，或复制到记事本文档中。
自由形式分析小组件配置
要了解更多信息，请参阅 自由形式分析小组件 文档。
推出针对 AIP 开发者能力的扩展 LLM 集成¶
发布日期：2023-12-07
我们很高兴地宣布扩展了对 AIP 开发者能力 中大语言模型（large language model, LLM）使用的支持，包括：
- 控制面板中的 AIP 开发者能力权限管理
- 在 Python 转换中使用语言模型
- 在对象函数中与语言模型交互
- Pipeline Builder 中的文本到嵌入（Text to Embeddings）面板
控制面板中的 AIP 开发者能力权限管理¶
已启用 AIP 的堆栈的管理员现在可以利用控制面板中新的 AIP 设置页面来管理 AIP 开发者工作流中对 LLM 的访问。从该页面，管理员可以启用、禁用以及选择哪些用户组能够在自定义和 Palantir 提供的 LLM 之上构建工作流。
控制面板中的 AIP 开发者能力权限管理
在 Python 转换中使用语言模型¶
使用 Python 转换的管道作者现在可以无缝构建利用 Palantir 提供的 LLM 和嵌入模型的数据管道。只需利用 palantir_models Python 包中包含的 Python SDK 即可。
使用 palantir_models Python 包将 LLM 集成到您的数据管道中
示例设置¶
下面的代码片段演示了管道开发者如何在其逻辑中使用 OpenAI 的 GPT-4 ↗ 实现转换，轻松利用 LLM 的潜力处理平台上存在的任何数据管道。
from transforms.api import transform, Input, Output
from palantir_models.transforms import OpenAiGptChatLanguageModelInput
from palantir_models.models import OpenAiGptChatLanguageModel
@transform(
source_df=Input("/path/to/input/dataset")
gpt_4=OpenAiGptChatLanguageModelInput("ri.language-model-service..language-model.gpt-4_azure"),
output=Output("/path/to/output/dataset"),
)
def compute(ctx, source_df, gpt_4: OpenAiGptChatLanguageModel, output):
...
有关平台上可用 LLM 及其用法的更多信息，请查阅文档中的 Palantir 在转换中提供的模型。
在对象函数中与语言模型交互¶
用户现在可以使用 TypeScript 函数创建使用 Palantir 提供的语言模型的自定义逻辑，从而促进摘要、问答、语义搜索等工作流。更新后的模型导入面板支持 Palantir 提供的模型和自定义编写的模型。将为所有导入的 LLM 生成 TypeScript 类，为用户编写的函数中使用 LLM 提供直观的接口。
Palantir 提供的 LLM 示例 - 可用性可能因堆栈而异
示例用法¶
例如，以下代码演示了如何使用 GPT_4 模型编写一个自定义 TypeScript 函数，对提供的文本运行简单的情感分析。
import { GPT_4 } from "@foundry/models-api/language-models";
@Function()
public async sentimentAnalysis(userPrompt: string): Promise<string> {
const systemPrompt = "Provide an estimation of the sentiment the text the user has provided. \
You may respond with either Good, Bad, or Uncertain. Only choose Good or Bad if you are overwhelmingly \
sure that the text is either good or bad. If the text is neutral, or you are unable to determine, choose Uncertain."
const systemMessage = { role: "SYSTEM", content: systemPrompt };
const userMessage = { role: "USER", content: userPrompt };
const gptResponse = await GPT_4.createChatCompletion({messages: [systemMessage, userMessage], params: { temperature: 0.7 } });
return gptResponse.choices[0].message.content ?? "Uncertain";
}
有关用法和示例的更多信息，请查阅关于 函数中的语言模型 的文档。
Pipeline Builder 中的文本到嵌入面板¶
Pipeline Builder 现在包含一个强大的新文本到嵌入（Text to Embeddings）面板。嵌入 是文本的密集向量表示，旨在捕获文本中单词或短语的语义含义，以便由 LLM 处理。嵌入将文本转换为可由 LLM 处理的数值形式。
数值表示（嵌入）允许基于上下文相似性（而不仅仅是句法相似性）来比较文本数据。例如，在比较单词 "cat"、"dog" 和 "balloon" 时，嵌入可以帮助确定 "cat" 和 "dog" 在含义上比 "cat" 和 "balloon" 更密切相关，并利用这种理解来驱动高级文本分析和操作。
在 Pipeline Builder 中使用嵌入对于涉及语义搜索的工作流特别有益，因为它允许 LLM 通过评估向量的相似性来执行更有效、更细致和更准确的搜索。
示例用法¶
此面板接受一个字符串列作为输入，并将使用 Palantir 提供的 text-embedding-ada-002 模型来创建嵌入向量。
文本到嵌入面板配置
然后，这些嵌入可以作为 Embedded vector 属性添加到本体中，并在下游的 LLM 驱动工作流中使用。
将嵌入作为工作流的一部分使用
了解更多¶
要了解有关上述功能的更多信息，请查阅以下文档：
- 使用 控制面板管理 AIP 设置
- 了解如何在 Python 转换中使用语言模型
- 了解如何在 函数中使用语言模型
- Pipeline Builder 中的嵌入
其他亮点¶
Foundry 开发者控制台
发布日期：2023-12-12
在开发者控制台中导出 OpenAPI 规范 | 用户现在可以为其开发者控制台应用程序生成并导出 OpenAPI 规范为 YAML 文件。此功能可通过重命名的 SDK 生成 选项卡访问，该选项卡以前称为 版本历史。
Marketplace¶
发布日期：2023-12-12
Marketplace 中改进的产品图片对齐和布局 | Marketplace 产品页面现在具有改进的产品图片对齐和增强的布局。
应用构建 | Workshop¶
发布日期：2023-12-12
大型模块的更快嵌入模块加载 | 此更新显著提高了 Workshop 中非常大的嵌入模块的加载时间。使用大型模块的用户将体验到高达 5 倍的加载速度提升。
数据集成 | 代码仓库¶
发布日期：2023-12-12
高级拉取请求审批策略 | 引入了一个细粒度、基于规则的系统，用于配置拉取请求（pull request, PR）审批。借助高级 PR 审批策略，现在可以根据修改的文件指定哪些用户或组应批准 PR。该策略可以在 YAML 中使用交互式审批策略编辑器进行编辑。
策略随后应用于新创建的拉取请求。所需的审批者会显示在新的用户界面中，显示哪些规则已满足以及还需要多少审批。
要开始使用此功能，请创建一个受保护分支，然后从代码仓库设置窗格的 分支 选项卡编辑策略。
Marketplace¶
发布日期：2023-12-12
Marketplace 建模集成现已正式发布 | Marketplace 建模集成现已正式发布，使用户能够打包封装了各种功能逻辑（如机器学习、预测、优化、物理模型和业务规则）的容器化可执行文件。
管理 | 控制面板¶
发布日期：2023-12-12
针对 AWS 托管堆栈的增强出口策略 | AWS 托管堆栈上的用户现在可以为其在同一区域的 S3 存储桶创建出口策略，从而提供对数据访问的更多控制。请注意，未来的更新将通过根据存储桶区域提供关于策略资格的清晰反馈来改善用户体验。
对象监控¶
发布日期：2023-12-12
增强的自动化视图和编辑模式 | 对象监控应用程序现在在自动化标头中具有改进的视图模式选择器，为用户提供更多选项来在自动化编辑和视图模式之间切换。执行模式选择变得更加可见，并且自动化条件图标已合并以获得更好的用户体验。
数据集成 | 代码仓库¶
发布日期：2023-12-12
Spark 模块运行时升级到 Spark 3.4 | Foundry 的 Spark 模块运行时已升级到 Spark 3.4，为用户提供最新的性能和安全性增强。此升级适用于所有版本高于以下版本的模块：Python 1.975.0、Java 1.997.0 和 SQL 1.861.0。用户无需执行任何操作，模块将自动升级。
Foundry 开发者控制台¶
发布日期：2023-12-12
在开发者控制台中创建 API 令牌 | 用户现在可以在开发者控制台中生成具有受限功能的长期 API 令牌，用于个人使用和 CI/CD 工作流。初始功能允许在 SDK 工件仓库和 Foundry SDK 资产包上安装具有 artifacts:read-artifacts 范围的 SDK。在 应用程序 SDK 设置页面或通过入门说明访问 创建 API 令牌 对话框。
对象监控¶
发布日期：2023-12-12
增强的对象监控功能 | 此更新为对象监控应用程序带来了多项增强。用户现在可以在登录页面上查看过去 28 天的失败事件，从而更好地了解近期问题。此外，还添加了更多删除自动化的选项，使管理和维护工作流更加容易。操作菜单已重新组织，将自动化选项移至单独的菜单以便于访问。最后，评估延迟信息已得到改进，为添加和删除的对象上的时间条件组合提供了更清晰的措辞和添加的提示。
安全 | 项目¶
发布日期：2023-12-05
增强的项目联系人选择 | 没有联系信息的用户和组现在可以添加为项目联系人，提高了项目管理的灵活性。更新后的界面为没有电子邮件地址的用户和没有联系信息的组显示适当的 UI。
应用构建 | Workshop¶
发布日期：2023-12-05
具有关键词搜索功能的新 PDF 查看器小组件 | Workshop 的 PDF 查看器小组件现已正式发布，具有关键词搜索功能，支持文本高亮显示和匹配时自动滚动。此增强功能通过手动或基于变量的输入改进了 PDF 工作流。
管理 | 控制面板¶
发布日期：2023-12-05
在控制面板中公开云运行时出口 IP | 用户现在可以直接在控制面板中访问云运行时出口 IP 地址，从而更容易为 API 和云源设置入口过滤。此更新为用户提供了一种更上下文相关且更方便的方式来获取其云运行时所必需的 IP 地址。
安全 | 项目¶
发布日期：2023-12-05
支持 Vertex 和 Vortex 本体依赖 | 项目现在支持 Vertex 和 Vortex 本体依赖，增强了本体访问检查器的能力，以实现更高效、更准确的本体管理。
数据集成 | 代码仓库¶
发布日期：2023-12-05
附件作为代码仓库中的函数输入 | 用户现在可以上传附件，用作代码仓库中的函数输入，为在实时预览和已发布函数中使用附件提供了更无缝的工作流。要开始使用，请添加一个具有 Attachment 或 Attachment[] 输入类型的函数，然后打开 实时预览 选项卡上传您的附件。您可以查阅关于 使用函数时可用的输入和输出类型 的文档。
本体 | 本体管理¶
发布日期：2023-12-05
本体管理的操作日志必需性切换 | 在本体管理器中引入了一个新的 需要操作日志 切换开关。启用后，任何编辑指定对象类型的操作都需要一个操作日志规则，确保对该类型对象的所有编辑都被记录，以获得更高的可见性和可追溯性。