BattleFish: Knowledge Enhanced LLM Translator for Tactical Environments

The U.S. Army is seeking innovative methods to improve the operational capabilities and interoperability of its tactical forces. Current methodologies for creating, managing and sharing complex sets of data across systems in the tactical domain are largely manual and very time consuming.  Current approaches are often fragmented and do not follow standardized data models, consistent APIs, or reliable semantic meaning of underlying data content terms.  This results in suboptimized tactical systems interoperability, exacerbated by frequent software updates, security vulnerability changes, and new system releases. Ultimately, the Army needs any system to talk to any system, and a way to quickly (think minutes to hours) adapt or refresh when an update, change, or introduction of a new system occurs. Applying Large Language Models (LLMs) and other AI approaches to automate tactical systems integrations and enable data unification and interoperability has great potential to meet this need. CSI proposes a capability called BattleFish, a play on Douglas Adams’ ‘Hitchhiker’s Guide to the Galaxy’ Babel fish, which allowed all creatures in the universe to understand each other.  Our vision for BattleFish involves leveraging an augmented LLM via a knowledge graph-based Retrieval-Augmented Generation (RAG) in an Agentic AI framework to provide a fast, effective and scalable translation capability.  Due to the limitations of computing at the tactical edge, the capability will come in two forms: a dynamic form for live use where computing is available and a static (non-adaptive) translation matrix which is generated from the dynamic form but capable of running on edge computing light-weight hardware. The translation matrix is expected to operate much like a derived small language model, but can be hosted on a standard low-power CPU.  The dynamic form of the BattleFish capability will be capable of generating an updated translation matrix on demand. This will support periodic API changes, content changes, and new message/system end points as a result of system updates, the introduction of new systems into the battlespace ecosystem, and other interface and message changes. Phase I is to design and demonstrate the feasibility of key aspects of the BattleFish concept and approach, particularly the integration of LLMs with knowledge-level RAG to generate an edge-capable translation matrix capable of operating on tactical edge devices.  The approach will be demonstrated as a feasibility prototype, but the core objective is to develop a workable design with an understanding of the features and limitations as well as insights into the interoperability metrics to be evaluated in Phase II.  CSI proposes Dr. Todd Carrico as the PI. Dr. Carrico is a leader in Human-Collaborative AI (H-CAI) decision support.  LtCol John (Danny) Looney recently completed 27 years in the US Army as a Senior Logistician, last servings as V Corps Deputy G4 and will serve as an SME.