Generative AI has the potential to dramatically enhance our defense capabilities, if we can harness it effectively and safely. Today, teams are already using generative AI in a variety of ways across the Department of Defense. Recent examples include generating operational orders; understanding huge collections of open-source data for intelligence analysis; and generating and testing software. From text, code, and image generation to research and discovery, the uses of generative AI in the defense sector are broad and deep – and growing rapidly.

We’re seeing an explosion in the number of generative AI use cases, the number of services and teams exploring it, and the novelty and creativity of those use cases. The DoD Chief Digital and Artificial Intelligence Office (CDAO)-sponsored Task Force Lima reported over 180 generative AI use cases in 2023 – and that number has grown dramatically since.

Generative AI’s Challenges

While generative AI is becoming more common in defense, a few challenges have emerged. Among them is computing costs. Generative AI projects at-scale require a large amount of GPU computing power, and despite recent advances like DeepSeek, the costs are much higher than conventional computing power. Further, with the worldwide popularity of generative AI, there have been GPU supply shortages, leading to limitations of scalability.

As an example, NIPRGPT – a Generative AI chat service for defense users – illustrates this dilemma well. Since its release in June 2024, 100,000 users have flocked to NIPRGPT to experiment with creative use cases that include software code generation, summarizing lengthy documents, and drafting reports. This extreme popularity and unexpected volume, indicates a pent-up demand that will only increase. It also underscores the challenge of managing unplanned and unbudgeted costs, as the surge in usage has driven up GPU computing costs. All services face this demand for generative AI and must budget now beyond the traditional operational expenditures.

Another issue for many teams is ensuring human oversight in AI decision making to improve observability and transparency. Teams need to be able to trust the generative AI results and manage its hallucinations, limiting the wildly, demonstrably incorrect responses you can sometimes get from AI tools.

Innovations like retrieval augmented generation (RAG) enhance generative AI results by including domain specific information - which has substantially reduced hallucinations. Whether you realize it or not, you’ve probably experimented with RAG methods when using AI chat services like ChatGPT. Whenever you provide additional information sources in your chat exchange, you’re using a RAG method. For defense use cases, RAG is the most reliable deployment methodology for generative AI services.

Enter Adaptive RAG

Adaptive RAG modifies the user-supplied prompt when the initial generated response fails to meet a standard of quality and relevance threshold. When an initial response is evaluated to be incomplete, inaccurate, or too generic, the adaptive RAG adapts the prompt and searches again for better, more relevant information. It is a much more accurate method that’s also self-reflecting essentially asking whether the response make sense given the prompt.

A traditional RAG solution will retrieve a wide array of information about a topic that meets the prompt, regardless of relevance or quality, and the answer will only be as good as the prompt. Adaptive RAG solutions will instead change both the data prompt and its answer generation to achieve an optimized result.

For example, a defense analyst asks an AI tool for a summary of a classified report. Traditional RAG pulls related documents, but adaptive RAG goes a step further – it rewords the prompt to include more specific terms found in the classified report, reruns the search, and delivers a more complete and useful answer to help the analyst make better decisions.

What’s Ahead for Adaptive RAG in Defense?

Today, as illustrated by the NIPRGPT experiment, the demand for generative AI capabilities is large and creative. The next step to meet that need is to enhance RAG deployment with adaptive approaches so systems can both access domain-specific data as well as create more accurate results – all while limiting hallucinations. This will allow DoD mission partners to drive efficiency and improve result quality.

There is a tremendous amount of promise and potential for adaptive RAG – especially given the need for trust in generative AI responses. That said, the ability to experiment with many RAG and related AI approaches – and to quickly see what works, what doesn’t and what can be built upon for future projects – is setting teams up with the knowledge and skills to make effective, efficient use of AI in all its forms. Doing that is what will bring about continued innovation and discovery, and truly facilitate the art of the possible.