A framework for redesigning RAG systems around historical methodology — preserving source sovereignty, interpretive transparency, and temporal sensitivity where standard architectures undermine them.
Workshop Hands-on with our UN-corpus RAG system · Berlin · 17 July 2026 — details & registration →RAG systems are designed for factual question-answering — find the relevant passages, generate the answer. Historical scholarship demands something different: source criticism before interpretation, temporal sensitivity across decades of discourse, and transparent collaborative reasoning rather than seamless answers.
Query → retrieval → generation in one step. Source selection is a technical optimisation hidden from the researcher. Similarity-based ranking favours recent vocabulary. No built-in space for source criticism. Output presented as answers.
Two separated phases restore the historian's workflow: a Heuristik phase for source discovery and evaluation, followed by an Analyse phase for interpretation. The researcher curates what enters computational reading. Outputs are Zwischentexte (interpretive proposals, not conclusions).
Drawing on Agre's Critical Technical Practice, we embed disciplinary values into system architecture rather than accepting computational defaults as neutral.
Formally decouples corpus construction from interpretation. Researchers examine, critique, and curate retrieved sources before any computational "reading" begins, thus restoring the heuristic phase that standard RAG eliminates.
Enforces proportional retrieval across time periods. Left unchecked, similarity-based search embeds presentist bias, privileging sources whose vocabulary matches modern query terms while suppressing formative periods where concepts emerged.
Post-retrieval evaluation against researcher-defined criteria. Turns algorithmic selection from a black box into a transparent, argumentative process with scored justifications that can be reviewed and contested.
HistoRAG separates the RAG pipeline into distinct phases, each with explicit researcher control points. The architecture is transferable with specific implementations configuring chunking, embedding, and evaluation criteria as well as further features for unique use-cases.
HistoRAG generates what we term Zwischentexte (intermediate texts). These are not answers but interpretive proposals: they lie between retrieved sources and historical argument, offering first proposals for interpretation that the historian can verify, contest, and develop.
"The central question for LLMs in digital humanities is not whether machines can 'read' but how we design systems that make their interpretive interventions visible and contestable, thereby preserving the scholar's epistemic agency throughout."
Our first implementation of HistoRAG, applied to computerisation discourse in Der Spiegel (1950–1979). Tracking how West German society's understanding of automation evolved — from "Elektronenhirn" to "Computer" to "EDV," and from euphoria to anxiety.
Not settled conclusions but Zwischentexte — interpretive proposals the system surfaced for the historian to verify, contest, or discard:
The system proposed that public anxiety about automation surfaced around 1964 — earlier than the canonical 1978 "Computer-Revolution" account — through reader letters that keyword search misses. A hypothesis to test, not a verdict.
The same term appears to carry opposed meanings by speaker position — efficiency for management, existential threat for workers — a pattern legible only at corpus scale.
A proposal that technological anxiety moved upward through the class structure over time, becoming socially charged once it reached the discourse-producing classes.
HistoRAG is a transferable framework. Each instance configures the architecture for a specific corpus and research context.
Computerisation discourse across the full archive — the paper's primary case study uses a clean 1950–1979 window (102,189 articles). Its corpus-trained FastText layer surfaces period-specific vocabulary (e.g. Elektronenhirn) a modern researcher wouldn't know to search.
Every member state's annual address to the General Assembly — 10,969 speeches from ~200 countries — fused at query time with UN Security Council resolutions. The corpus behind our 17 July workshop.
~19,000 parliamentary papers (Drucksachen) of West Berlin's Abgeordnetenhaus, Wahlperioden 1–10 — motions, written questions, bills and reports from the walled city, distinct from East Berlin's Stadtverordnetenversammlung.
The West German federal parliament — plenary debates (Plenarprotokolle) and printed papers (Drucksachen: bills, motions, written questions) fused in one search. The Bonn-era counterpart to the GDR's Volkskammer, and to Berlin's city parliament above.
~98,000 emails from the foundational archive of digital-privacy and crypto-anarchism — May, Hughes, Finney and others. Distinctive feature: thread-aware retrieval that reconstructs whole conversations, not just isolated messages.
All five instances are invite-only — each sits behind a sign-in and is not open to the public. To request access for your own work, write to [email protected]; we add your address to the allowlist and send you the instance link. Researchers connect their own model API key (OpenAI, Anthropic, Google, or DeepSeek), or use a locally hosted, open-weight Qwen model.
Work directly with a purpose-built RAG system on the UN General Debate Corpus — the speeches of every UN member state since 1946 — and explore how scholarly sovereignty can be developed and preserved when researching with large language models. For researchers across the humanities, cultural, and social sciences. No programming required; just bring a laptop.
Can't make it to Berlin? The same format is offered at DH2026 (Daejeon, 27–31 July 2026) by Torsten Hiltmann and Noah Kim-Baumann.
HistoRAG: Embedding Historical Methodology in Retrieval-Augmented Generation Through Critical Technical Practice
Noah J. Kim-Baumann & Torsten Hiltmann · 2026
The framework paper — open-access preprint on arXiv, June 2026.
Kim-Baumann, N. J. & Hiltmann, T. (2026). HistoRAG: Embedding Historical Methodology in Retrieval-Augmented Generation Through Critical Technical Practice. arXiv:2606.18103.
A companion executable-notebook article on the SPIEGELragged case study is under review at the Journal for Digital History.