CHARLIE Brings Multi-Agent RAG to Digital Forensics, Keeping Evidence Off the Cloud

Digital forensics teams routinely drown in evidence: seized devices, chat logs, financial records, surveillance footage transcripts, all of it needing to be cross-referenced without breaking the chain of custody. A newly published system called CHARLIE proposes a way to let AI agents do that cross-referencing without ever sending sensitive material outside institutional walls.

Detailed in a paper submitted to arXiv on July 1, 2026 (arXiv:2607.05428), CHARLIE is described as a multi-agent Retrieval-Augmented Generation system purpose-built for structured evidential processing in forensic settings. The work was first presented at RELAF 2026, the inaugural Workshop on Reasoning with Evidence in Law Enforcement and Forensics, held alongside ICAIL 2026 in Singapore.

Why On-Premise Matters Here

The defining design choice behind CHARLIE is where it runs. Rather than routing queries through external cloud APIs, the system operates entirely on-premise within an institution's own infrastructure. For forensic and law-enforcement contexts, that is not a minor architectural preference: evidentiary integrity and data sovereignty are legal requirements, not conveniences. A cloud-dependent RAG pipeline that phones home to a third-party model provider introduces exactly the kind of chain-of-custody ambiguity that can get evidence thrown out or a case challenged in court. By keeping retrieval, reasoning, and storage local, CHARLIE's authors argue the system avoids that exposure entirely.

From Classical RAG to Agent Orchestration

CHARLIE's technical core marks a deliberate departure from standard retrieval-augmented generation. The paper describes a shift from a classical RAG setup, where a single model retrieves context and generates an answer, toward an agent-based orchestration model. In practice, this means the workload is split across specialized components: local retrieval modules pull relevant material from case files, task-decomposition logic breaks investigative queries into smaller sub-tasks, a structured memory layer keeps track of what has been established across documents, and verification mechanisms check outputs before they're surfaced to an investigator.

That combination is aimed squarely at a problem generic chatbots handle poorly: reasoning across large, messy document sets while keeping every claim traceable back to its source. CHARLIE is built to scale across multi-document extraction tasks and to support what the authors call longitudinal forensic intelligence, meaning insight that accumulates and stays coherent as a case grows over weeks or months, rather than being reconstructed from scratch with each query. Traceability and auditability are treated as first-class requirements throughout, not an afterthought bolted onto the output.

What It Signals for Legal AI

CHARLIE lands at a moment when courts and regulators are increasingly skeptical of black-box AI tools touching evidence. A system explicitly engineered around auditability and local data control, rather than raw generative fluency, suggests where forensic AI tooling may be headed: less about chat-style convenience, more about defensible, inspectable reasoning chains that can survive cross-examination. The RELAF workshop's own existence, focused specifically on reasoning with evidence in law enforcement, points to a growing research community forming around exactly these constraints.