Palantir Foundry AIP

What is Foundry and AIP?

Palantir Foundry is known as one of the most popular end-to-end data platforms. As we previously introduced in our Foundry 101 article, it is an operational platform that holistically encapsulates your data processing, analytics, BU, and ML needs and allows for informed decision-making for both technical and non-technical members of your organization.

Following the recent developments of Large Language Models and GenAI business applications, Palantir augmented their offering by integrating multiple GenAI capabilities into their platform.

In April 2023, the company launched its Artificial Intelligence Platform, aka AIP. AIP integrates large language models into privately operated networks and has been continuously and seamlessly integrated directly into the platform toolbox and functionalities since its inception.

In this article, we will present the different ways you can leverage the integration of LLMs into the platform, either from built-in tools or for custom implementations.

AIP Assist

“AIP Assist is an LLM-powered support tool designed to help users navigate, understand, and generate value with the Palantir platform“

AIP assist is available across the entire platform as a chatbot. Using a combination of NLP and third-party LLM, it allows you to query Palantir’s documentation in natural language.

Another feature and one of the latest additions to the AIP ecosystem is the ability to query custom documentation directly via AIP assist. This feature, released in May 2024, allows you to add a custom documentation repository to the API assist knowledge base. This feature makes RAG implementation very straightforward inside the platform. It does require some minimal additional work as you will need to create a documentation of your documentation:

• For each documentation markdowns that you might have you can in an overview.md file define @name and @description of your custom documentation. This description will be given to the LLM so that it knows when to query the documentation.

• Give read permission to AIP by renaming the pre-generated file _aip-assist.json into aip-assist.json.

• Add your documentation to AIP in the control panel.

AIP Integration to Repository

Anytime you are inside the repository you will see a purple button “Ask AIP Assist”. As shown below, different tools are available for developers to assist them in their debugging and documenting tasks.

The “Explain code” and “Find bugs and fix compiler errors” are preconfigured actions/prompts. You can either highlight a specific code snippet or use the entire file. Upon selecting one of the above actions, the AIP assist dialogue will open, giving you either the code explanation or suggestions to fix your code. The AIP assist also has the ability to suggest a few actions, such as “running the preview“ or ”find and replace string. ”

The translate tool allows you to convert your code from/to Python, Java, Javascript, SQL, or TypeScript.

Another way AIP helps us to develop inside the code repository is by suggesting autocomplete of your code, acting as a co-pilot.

AIP Integration to Pipeline Builder

Foundry has increasingly invested in improving its low-code solution pipeline builder. This tool allows you to create complex and versioned processing pipelines. With AIP, they’ve extended the capabilities of the pipeline builder with features such as the generate feature, LLM functions, or the explain feature. As the low-code alternative of the code repository targets less technical users, this feature enables business users with good technical requirements understanding to manipulate data efficiently.
As with any dialogue interface, it takes some practice and prompt engineering to get it right. Ultimately, even if it is not perfect in a more complex generate request, it can still significantly improve development time as it will create transform blocks that you can then manually modify or tweak.

One task I found very suitable for generating content is string cleaning or regex generation. This generate feature is besides directly integrated in the transform block.

Palantir Foundry also provides access to custom and predefined LLM functions within the pipeline builder. Allowing us to create transforms powered by LLMs.

Model Catalog

The model catalog is an AIP application that allows you to access and use most of the popular LLMs including but not limited to GPT, Claude and Llama.

Upon selecting a model, the model catalog application allows you to easily test different prompts for each model in a playground or sandbox environment.

Each of those models can be used inside either a TypeScript or Python function. Note that media input is not yet supported in TypeScript but should be released soon.

AIP Logic

“AIP Logic is a no-code development environment for building, testing, and releasing functions powered by LLMs.“

AIP logic can be seen as the no-code alternative to the model catalogue for LLM-powered functions. AIP logic reduces the complexity coming from API calls and the development environment and makes it easy for your LLM to interact with your ontology. For example, you can define any of your ontology objects or object sets as the function input. AIP logic functions are also capable of either returning a value or, more interestingly, editing ontology objects.

Conclusion

The release of AIP is a great addition to the Foundry environment. Palantir’s approach of leveraging generative artificial intelligence for the platform itself via AIP assist, as well as making multiple LLMs easily usable for custom implementation, shows by example the potential of GenAI and empowers users to try it out.

Foundry also strongly focuses on making powerful low/no-code alternatives to plain code to allow more users to contribute. We’ve seen those efforts on the platform already with Pipeline Builder, which allows you to create complex and versioned data pipelines, and we are seeing it again now with the AIP Logic application.

Before AIP, Foundry faced some limitations for tasks that required deep learning solutions. With the release of the latest LLMs, parts of those tasks can now be performed more easily on the platform. Overall, AIP contributes to the broader adoption of Generative AI in the industry by making it accessible and scalable. Future releases that will, for example, allow media input for the typescript function, as mentioned above, will also have a big impact on the potential capabilities of existing applications such as workshops.