From Digital Library Records to Trust-Grounded Virtual Experts for Knowledge Retention

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Long-lived technical organizations depend on senior subject matter experts whose knowledge is only partially preserved in reports, publications, and archival metadata. Digital libraries preserve those records, but they rarely provide centralized, expert-scoped, question-answering access to the knowledge of people who leave or change roles. We present an implemented knowledge-retention prototype, built on ALICE, an architecture for large language model (LLM)–knowledge-graph interaction and grounded evidence provenance, that creates virtual experts from digital-library records. Its core contribution is a data-organization layer for expert-scoped knowledge: author-scoped ingestion from the NASA Technical Reports Server (NTRS), Docling-derived chunking, evidence-aligned triple extraction into per-expert Kuzu graphs, and a lifecycle model for creating, refreshing, and retiring expert knowledge bases; a trust-aware, persona-conditioned conversational interface is one application built on that layer. We also introduce Trust-Propagated Path Retrieval (TPPR), a graph retrieval mechanism that scores multi-hop evidence chains by the geometric mean of their edge certainty scores. On standard multihop question answering (QA) benchmarks, the ALICE retrieval architecture with GPT-4o achieves 56.5% exact match (EM) with TPPR on 2WikiMultihopQA and 32.5% EM with high-recall retrieval on MuSiQue. Oracle analysis indicates that MuSiQue is limited primarily by retrieval coverage, with a 37.5% retrieval failure rate versus 8.5% for 2WikiMultihopQA. The analysis snapshot includes two expert collections spanning 74 publications, 5,866 chunks, 25,534 entities, and 22,734 extracted relations. A formative masked pilot with 10 NASA domain experts suggests that virtual-expert responses are rated substantially more human-like than baseline LLM responses and can be identified above chance, while remaining too small for inferential claims.