Sovereign Library Intelligence
An agentic AI framework for trustworthy library collections stewardship and access: locally hosted, evidence-linked, and under institutional control.
AI that libraries can trust, and keep control of
Libraries and archives hold millions of digitized items that are poorly described. AI can help generate that missing metadata, but only if libraries can trust it and keep control of their own collections.
Standard generative models are prone to hallucination, producing plausible but fabricated facts and citations. At the same time, commercial cloud AI often treats cultural heritage data as training material, ingesting institutional metadata into proprietary models without consent. Libraries risk trading long-term control of their heritage for short-term automated convenience.
SLI is a three-year applied research project investigating how agentic AI architectures, combining structured reasoning with tool-based verification, can be designed to meet library standards for transparency and local stewardship. The system drafts archival metadata, ties every field to the specific portion of the source that supports it, and checks that evidence against professional authority files, all within the institution's own infrastructure.
The project produces a validated design framework for privacy-preserving AI, a benchmarking methodology for evidence quality, and an open-source toolkit that institutions of any size can adopt.
Four questions guide the investigation
Reducing factual errors
Which combinations of structured reasoning steps and verification methods most effectively reduce factual errors under real library constraints?
Evidence-linked description
How can AI-generated descriptions of text, images, and video be directly linked to the specific portions of archival materials that support them?
Measuring quality
To what extent can an expert-validated toolkit accurately quantify the groundedness and fidelity of AI-generated archival metadata?
Sovereign vs. cloud
Under what conditions does a locally hosted, domain-focused model demonstrate higher factual precision and institutional alignment than proprietary baselines?
Three technical innovations
Together they let an AI system describe archival materials the way a professional would: step by step, with evidence, and inside the library's own walls.
Agentic verification
A pipeline of coordinated agents: a Planner that decomposes archival tasks, an Extraction agent that identifies entities, dates, and locations, and a Verification agent that validates them against authority files, mimicking professional archival reasoning.
Evidence-linked metadata
Each generated field is tied to a specific source segment: character spans and line IDs for text, regions for page images, and timecodes for audio and video, producing a verifiable audit trail.
Institutional sovereignty
A gateway governance pattern keeps every model interaction inside the institution's perimeter. Open-weight models run on local hardware, and a sovereign knowledge base of LCSH, FAST, Getty AAT, and local finding aids grounds every decision.
Evaluation moves beyond final-answer scoring to trace-based auditing of intermediate reasoning steps, with two domain-specific metrics: