Private RAG for a Historical Research Archive

Context

A historical research project needed to work with 10,000+ pages of WWII-era archival material — German military records and RAF operational logs — available only as scanned microfilm reels. No OCR, no index, no way to ask "where is X mentioned?" without weeks of manual reading. The material is research-sensitive: sending it to a cloud AI service was not an option.

Solution

An end-to-end private pipeline. OCR tuned for degraded historical typescript in two languages, with structured Markdown output. Semantic + keyword hybrid search over the full corpus, with optional LLM re-ranking for difficult queries. A multi-tenant visibility model — shared and private document collections with server-side access filtering as the single security chokepoint — so multiple researchers work on one system without seeing each other's private material. A purpose-built frontend: reel browser, page-level navigation, lazy-loaded thumbnails. Everything runs on dedicated infrastructure; no document ever leaves it.

Stack

Outcome

A previously unsearchable archive answers questions in seconds. Researchers query in natural language across German and English sources simultaneously. The system is the reference deployment behind the AI Assessment offer — the same architecture, adapted per client.