Traité de documentation : le livre sur le livre, théorie et pratique

TL;DR: In 1934, a Belgian lawyer-turned-bibliographer named Paul Otlet published a 400-page treatise arguing that the book is a bad container for knowledge: documents should be decomposed into atomic facts, each fact stored once with its source, linked to related and contradicting facts, and served remotely to a screen on your desk — he called it a "télescope électrique." He then tried to build it with index cards and accumulated roughly fifteen million of them before the project collapsed. The web vindicated the easy half of this vision (remote access to documents). The hard half — knowledge as a provenance-tracked graph of claims rather than a shelf of PDFs — was infeasible until LLM-scale information extraction, and it is exactly the substrate that scientific synthesis still lacks.

The librarian who wanted to delete the book

Otlet is usually filed under "visionary who anticipated the web," which undersells him. With Henri La Fontaine he had already built, starting in the 1890s, the Universal Bibliographic Repertory — an attempt to catalog every published work on standardized index cards — and the Universal Decimal Classification, a faceted classification scheme still in use in European libraries. The Traité de documentation (1934) is the theoretical summa of that forty-year project, and its central move is more radical than hypertext.

The book, Otlet argued, is an accident of printing economics. It bundles many statements together, in a fixed order, with redundancy across books, and it forces the reader to re-extract the facts every time. His "monographic principle" says: cut the bundle apart. The unit of knowledge should be the individual statement — a dated, sourced, classified assertion — recorded once on its own card. The universe of these cards, cross-linked by classification codes, becomes what he called the Universal Book: a single, continuously updated, deduplicated representation of what humanity knows, in which any given fact appears once and every document is just a view into the network.

On top of this substrate he described the access layer, and this is the part that gets quoted: work desks without books, equipped with a screen and a telephone line, querying a distant repository; documents "téléphotés" — transmitted as images — on demand; eventually machines that answer questions rather than return documents. He imagined the whole thing as a "mechanical, collective brain." This is 1934: eleven years before Bush's Memex, three decades before Nelson coined "hypertext," fifty-five before Berners-Lee's proposal.

Why it could not work then

Every stage of that pipeline required a technology that did not exist, so Otlet substituted the only universal machine available: clerical labor.

Decomposition had no machine. Extracting facts from a document is reading plus judgment. A trained indexer at the Mundaneum could produce perhaps a few dozen good cards per day. Global scientific output in the 1930s was already on the order of hundreds of thousands of papers per year (a rough figure; growth curves from Price's later work suggest this order of magnitude). The extraction task alone needed tens of thousands of full-time expert readers, forever. Otlet's institute never had more than a small staff plus volunteers.

Storage had no compression and no random access worth the name. The Repertory reached roughly 15 million cards. As text this is tiny — a card holds maybe 100–300 characters, so the whole collection is on the order of a few gigabytes, less than one modern phone photo library. But physically it filled halls of cabinets in the Palais du Cinquantenaire in Brussels, and "queries" were answered by staff walking the drawers, responding to postal requests for a fee. Latency: days to weeks. Concurrency: one clerk per drawer.

Linking was write-once and one-directional. A cross-reference on a card is a pointer that can never be updated in bulk. There is no backlink index, no recomputation when classifications change. The UDC gave Otlet a global address space — genuinely clever — but maintaining consistency across millions of cards by hand does not scale; it anti-scales, since every new card can invalidate old links.

Transmission barely existed. Facsimile over telephone lines was demonstrated in the 1920s, and television was a laboratory curiosity; Otlet extrapolated from both. But bandwidth to the home was a voice call, and there was no switching infrastructure for document retrieval on demand.

The project also died politically: the Belgian government withdrew support in the 1930s, and much of the collection was damaged when German forces cleared part of the premises in 1940. Otlet died in 1944, largely forgotten until W. Boyd Rayward's historical work from the 1970s onward revived him.

What changed

The web solved transmission and storage in the 1990s — this is the well-worn part of the Otlet story, and Rayward, Buckland, and later the documentary The Man Who Wanted to Classify the World made the case that the Mundaneum was a "paper Google." But note what the web actually built: remote access to documents. The unit stayed the page, the PDF, the bundle. Google ranks bundles; citation graphs link bundles; RAG retrieves chunks of bundles and re-extracts facts at query time, wastefully, every single time — precisely the redundancy Otlet wanted to eliminate.

The missing machine was one that reads. Claim extraction from scientific text is a task that resisted classical NLP — SemRep and biomedical relation extraction produced brittle triples, open information extraction produced noise — and became genuinely workable only with large language models. An LLM today can take a paper and emit its core claims with hedging level, evidence type, and grounding spans, at a marginal cost of well under a cent per abstract. At roughly 200 million papers in Semantic Scholar's corpus, a claim-level pass over all abstracts is a project costing on the order of $10⁵–10⁶ — a lab-scale budget for what required a civilization-scale clerical workforce in 1934. Entity resolution, claim canonicalization ("is claim A the same assertion as claim B?"), and contradiction detection — the merge step in Otlet's pipeline, which is harder than extraction — are also now approximately solvable with embeddings plus NLI-style judgment, though far from solved.

A 2026 Mundaneum

A serious revival is a claim graph over the scientific literature, built as infrastructure:

• Nodes: canonicalized claims, each with an explicit statement, entity links, and a distribution over hedges ("X inhibits Y in mouse models" ≠ "X inhibits Y").
• Edges: asserts (paper → claim, with grounding span — Otlet's provenance requirement, non-negotiable), supports, contradicts, refines, assumes. The contradiction edges are the payoff: the graph should surface, automatically, that field F contains 40 papers asserting P and 15 asserting ¬P, partitioned by experimental condition.
• Dynamics: claims carry belief state updated as new papers arrive — a living literature review, or in Otlet's terms, the Universal Book that is never finished.
• Query layer: the electric telescope, which now means an LLM interface that answers questions by traversing claims and citing spans, not by re-reading PDFs.

Reuse from Otlet: the monographic principle itself; the insistence on a universal address space for claims (Wikidata QIDs and ontologies play the role of UDC codes); the separation of the knowledge substrate from the documents that feed it. Replace: manual classification with learned extraction; the single central institution with a federated, versioned, disagreement-tolerant graph — Otlet's monism (one true fact per slot) is the one part of his vision we should reject, since science is exactly the domain where claims conflict and the conflicts are the signal.

The honest technical risks: extraction hallucination (mitigated by mandatory grounding spans and verification passes), claim identity being genuinely context-dependent (the same sentence means different things under different experimental setups), and the graph rotting as ontologies drift. These are research problems, not blockers.

Descendants, partial vindications, open ground

The lineage is real but every branch stopped short. Bush's Memex (1945) and Nelson's Xanadu inherited the linking, not the decomposition. The Semantic Web (2001) inherited the atomic-statement idea as RDF triples but assumed humans would author them; it got Wikidata — a genuine, thriving Otletian artifact for entities and simple facts — but never a claim layer over the literature. Nanopublications (Mons, Groth, and colleagues, ~2009 onward) are Otlet's cards reinvented for biomedicine: one assertion plus provenance plus publication metadata; adoption stayed niche because authoring was manual. Semantic Scholar, SciFact-style claim verification, Elicit, and Microsoft's GraphRAG each cover one stage of the pipeline. To my knowledge, no one has shipped the full loop — corpus-scale extraction, canonicalization, contradiction detection, cumulative synthesis — as maintained public infrastructure. That is the open ground, and it is more an institutional problem than a modeling one now, which is fitting: Otlet's failure was also, in the end, institutional.

Where to read it

The Traité is available in a digital edition at books.openedition.org/emsha/482; the original was published in Brussels in 1934 (the edition metadata at the link differs from the original imprint — I have not verified the details of this particular re-edition). The book is long and repetitive; the sections on the monographic principle and on the future of the book (the "electric telescope" passages, in the latter part of the treatise) are the essential reading. Read alongside: W. Boyd Rayward's The Universe of Information (1975) and his translated Otlet anthology International Organisation and Dissemination of Knowledge (1990), which are the scholarly foundation for everything above; Vannevar Bush's "As We May Think" (1945) for the contrast — Bush's Memex is personal and associative where Otlet's Mundaneum is collective and structured; and the nanopublication literature for the closest modern attempt at cards done right.