Ten Thousand Years of Tagging

We recently built a new classification system for Source Library — six independent facets that tag every book by tradition, domain, form, cultural sphere, era, and epistemic mode. While designing it, we realized something: most of the key documents in the history of knowledge classification are books we already have. So we read them.

What follows is the deep dive. Every text marked with a link is available in Source Library.

1. The Catalog: Callimachus at Alexandria (~245 BCE)

The first known library catalog was the Pinakes, compiled by the poet Callimachus at the Library of Alexandria. It organized roughly 500,000 scrolls into six classes: rhetoric, law, epic, tragedy, comedy, and lyric poetry. Within each class, authors were listed alphabetically, with biographical notes and a list of works.

This was a single-axis system: one scroll, one category, filed by genre. The categories were literary forms, not subjects. If you wanted to find everything about astronomy, you had to already know that Eudoxus wrote about astronomy and look him up by name. The Pinakes is lost, but its logic — sort by type, then by author — persisted for two thousand years.

2. The Ten Categories: Aristotle (~350 BCE)

Before anyone classified books, Aristotle classified reality. His Categories (part of the Organon, which we have in Greek manuscript and the Oxford translation) proposed that everything that can be said about anything falls into one of ten categories: Substance, Quantity, Quality, Relation, Place, Time, Position, State, Action, and Passion.

These aren’t library categories — they’re the grammar of existence. But they set the template for all Western classification: there exists a finite set of fundamental types, and everything in the world can be assigned to one. Every classification system since is either extending Aristotle or rebelling against him.

We hold multiple Aristotle editions including a Vatican Greek manuscript (Vat.gr.244), a Bodleian manuscript (MS Barocci 87), and the Rhetoric, Poetic, and Nicomachean Ethics (399 of 407 pages translated).

3. The Tree: Porphyry’s Isagoge (270 CE)

Six centuries later, the Neoplatonist Porphyry wrote a short introduction to Aristotle’sCategories that became more influential than the original. The Isagoge (which we have in a Greek manuscript with the commentaries of Proclus and Ammonius) demonstrated classification through binary branching: start with the most general category (Substance), then split with a differentia. Corporeal or incorporeal? Living or non-living? Rational or irrational?

The result is the Tree of Porphyry — the first hierarchical classification diagram. It became THE model for over a millennium. When we designed our faceted vocabulary, we borrowed Porphyry’s key insight: every tag value carries a one-sentence differentia explaining what distinguishes it from its neighbors.

The problem with trees: they force false choices. A book on Paracelsian medicine is both alchemy and medicine, but a tree makes you pick one branch.

4. The Emanation: Pseudo-Dionysius (~500 CE)

While Porphyry classified by type, an anonymous Syrian monk proposed a different model. In the Celestial Hierarchy (Ficino’s translation, 140 of 142 pages translated), knowledge is not a tree of types but a cascade of intensity. The divine Good emanates outward, each level receiving a diminished portion.

From page 60: “Just as every specific order of things is traced back to its own single Head, so too must the universal order of all things be ultimately referred to one universal Head of all.”

Dionysius suggests not all labels serve the same purpose. Some tags are “source” tags close to the author’s own terminology. Some are “bridge” tags that connect traditions. Some are “user-facing” tags that help browsers. These are different ranks in a Dionysian hierarchy of tags.

5. Beyond the West: Chinese Sibu, Islamic Fihrist, Indian Padarthas

The Western story from Aristotle to Porphyry was not the only game. Three other civilizations independently invented comprehensive classification systems — and two of them predate most European innovations.

The Chinese Four Divisions (Sibu), formalized during the Jin dynasty (3rd c. CE), organizes all written knowledge into four classes: Classics (jing), History (shi), Philosophy (zi), and Literature (ji). The Siku Quanshu (1782), the largest collection of books in Chinese history at 36,000 volumes, was organized using this system. It’s still the basis of Chinese library classification today. Our collection includes the Sancai Tuhui (Illustrated Encyclopedia of the Three Realms) (96 pages, fully translated) — a Ming dynasty encyclopedia that organizes heaven, earth, and humanity into systematic visual catalogs.

Ibn al-Nadim’s Kitab al-Fihrist (987 CE) is the Islamic world’s Gessner — 550 years earlier. It catalogs every Arabic book known to a Baghdad bookseller, organized into ten sections: Holy Scriptures, Grammar, History, Poetry, Theology, Jurisprudence, Philosophy, Legends, Doctrines of non-Muslims, and Alchemy. Each section has subsections and author biographies. It covers roughly 10,000 works — an astonishing scope for the 10th century.

The Indian padarthas (categories of reality) from Vaisheshika philosophy proposed six fundamental categories: Substance, Quality, Action, Generality, Particularity, and Inherence. S.R. Ranganathan, who invented faceted classification in 1933, explicitly credited the padarthas as his inspiration. The categories of Indian philosophy became the structural logic of modern library science.

6. The Combinatorial Turn: Ramon Llull (1305)

The most radical model came from a 13th-century Majorcan mystic. Llull’s Ars Brevis proposed that knowledge is not a tree at all. Instead, it emerges from the combination of a small set of independent principles.

Nine Absolute Principles (Goodness, Magnitude, Duration, Power, Wisdom, Will, Virtue, Truth, Glory) combined pairwise in a 36-cell table. From page 11:“Each principle taken by itself is entirely general... when one principle is contracted to another, then it is subalternated, as when one says ‘great goodness.’”

A book isn’t in a category — it combines attributes. Nine principles give 36 pairs, 84 triples, 126 quadruples — far more distinctions than 126 flat categories. Llull needed only nine because he understood: the number of categories should be small enough to hold in mind simultaneously. The power comes from combination, not enumeration.

Athanasius Kircher expanded this in Ars Magna Sciendi (1669, 532 pages, fully translated), adding a question dimension: each letter maps to Whether? What? Of what? Why? How much? When? Where? How?

7. The Reformers: Ramus (1543) and Gessner (1545)

Peter Ramus’s Dialecticae Institutiones (1543) proposed replacing Aristotelian logic with dichotomous division: take any subject, split it into two, split each part into two, repeat. Where Porphyry’s tree was metaphysical, Ramus’s was pedagogical — not classifying reality but organizing how to teach it. These “Ramist tables” conquered Protestant education across Europe and, crucially, crossed the Atlantic.

Two years later, Conrad Gessner published the Bibliotheca Universalis — the first attempt to catalog every book ever printed (~12,000 works). Alphabetical by author, then reclassified by 21 subject divisions in the companion Pandectae. Gessner’s innovation: multiple access points to the same content. Author, subject, date — all first-class entry points. Our faceted system has six.

8. The Cognitive Tree and the Universal Language: Bacon (1623) and Leibniz (1666)

Francis Bacon’s De Augmentis Scientiarum (fully translated, 696 pages) grounded classification in the cognitive operations of the knower. From page 126: “The truest division of human doctrine is that which is taken from the threefold Faculty of the Rational Soul. History is referred to Memory, Poesy to Imagination, Philosophy to Reason.”

This suggests a tagging dimension library science mostly ignored: not just what a book is about, but what kind of thinking it requires. We built this as the “epistemic mode” facet. Bacon also invented desiderata — mapping what knowledge ismissing — a tagging system for gaps.

Gottfried Wilhelm Leibniz took Llull’s combinatorial vision further. In his Dissertatio de Arte Combinatoria (1666) and the lifelong project of a characteristica universalis, Leibniz imagined a universal symbolic language where all knowledge could be represented as combinations of primitive concepts — and reasoned about mechanically. He explicitly built on Llull. We hold multiple volumes of Leibniz’s philosophical writings and his mathematical writings. His dream of a calculus ratiocinator — a machine that could compute with categories — anticipated both Ranganathan’s faceted classification and modern knowledge graphs.

9. The Empirical Superstructure: Robert Hooke (1705)

Bacon drew the tree. Robert Hooke tried to fill it in. His Posthumous Works (594 pages, 10 translated) included “A General Scheme, or Idea of the Present State of Natural Philosophy” — the “superstructure” on Bacon’s foundation.

Where Bacon organized knowledge by cognitive faculty, Hooke organized nature by observational method: what instruments you need, what senses are involved, what scale of phenomena. Light, sound, motion, gravity, magnetism — each domain broken down by the type of experiment needed to investigate it. This is a different axis: not what knowledge is about, not what the mind does, but what tools you need. The “General Scheme” is mostly untranslated — a priority for our OCR pipeline.

10. Classification as Curriculum: Comenius (1651) and Samuel Johnson of Yale (1752)

Comenius argued in Naturall Philosophie Reformed by Divine Light (308 pages, fully translated) that classification is curriculum. You learn simple before complex, and the classification system itself is a learning path. From page 15: “They must be used in this order: that we begin with sense, and end in revelation.”

A century later, this idea crossed the Atlantic. Samuel Johnson (1696–1772) — not the English lexicographer, but the first president of King’s College (now Columbia) — created an Encyclopaedia of Philosophy that organized all knowledge into Ramist-style dichotomous trees blended with Lockean empiricism. Published in his collected Career and Writings, it became a textbook at King’s College and shaped how the colonial generation organized knowledge.

The connection to the founding of the republic is not metaphorical. The men who wrote the Constitution were products of this educational system. Separate powers, enumerated rights, hierarchical jurisdiction — these reflect the Ramist habit of dividing any complex subject into a branching structure of named parts. The chain: Ramus (1543) → Protestant universities → Samuel Johnson at Yale/King’s (1752) → colonial curriculum → the founding generation’s mental models.

11. The Great Classifiers: Linnaeus (1735) and Diderot (1751)

Carl Linnaeus published the Systema Naturae in 1735 — just 21 pages in the first edition, but it contained the most successful classification system ever created. Binomial nomenclature (Kingdom → Phylum → Class → Order → Family → Genus → Species) is Porphyry’s tree made operational for biology. It’s still in use 290 years later. We hold the 1735 first edition and several companion works, including Linnaeus’s Hortus Cliffortianus (1737).

In 1751, Diderot and d’Alembert published the first volume of the Encyclopédie. Its famous “Système Figuré des Connaissances Humaines” diagram — viewable in the ARTFL Encyclopédie Project — is Bacon’s cognitive tree (Memory/Imagination/Reason) realized at industrial scale. 72,000 articles, 17 volumes of text, 11 volumes of plates. The Enlightenment’s operating system for knowledge.

12. The Library Scientists: Dewey (1876), Otlet (1905), Ranganathan (1933)

Melvil Dewey published his Decimal Classification in 1876: 10 top-level classes, 100 divisions, 1,000 sections. Every book gets one number. It’s Porphyry’s tree with decimal notation. Still used by most public libraries.

The Belgian bibliographer Paul Otlet took Dewey further. Starting in 1905, he and Henri La Fontaine created the Mundaneum — a vast paper-based knowledge system in Brussels containing over 12 million index cards cross-referenced by subject. Otlet expanded Dewey into the Universal Decimal Classification, adding a notation for combining subjects (a book on “chemistry of food in France” could be expressed as a compound number). In 1934, he described a “réseau mondial” (world network) that would connect all knowledge through electric signals — essentially imagining the internet.

The real breakthrough came from India. S.R. Ranganathan invented Colon Classification(1933) — the first truly faceted system. Instead of one category per book, every book gets one tag from each of five fundamental facets: Personality, Matter, Energy, Space, and Time. “History of Indian medicine in the 18th century” becomes L:2:f:44:N.

Ranganathan explicitly credited the padarthas of Vaisheshika philosophy as his inspiration. He was Llull’s intellectual heir: both understood that a small number of independent axes, freely combined, generates more distinctions than any flat list. Every time you filter by price + brand + rating + color on Amazon, you’re using Ranganathan’s idea.

13. The Digital Revolutions: Bush (1945), Folksonomy (2004), Knowledge Graphs (2012)

In 1945, Vannevar Bush published “As We May Think” in The Atlantic, describing the memex — a desk-sized device that stores all of a person’s books and records, accessed through associative trailsrather than hierarchical filing. “The human mind operates by association,” Bush wrote. “It should be possible to beat the speed and permanency of the brain.” The memex never got built, but it directly inspired Ted Nelson’s hypertext and, through him, the World Wide Web.

The Web brought a classification revolution nobody expected: folksonomy. Starting with Delicious (2003) and Flickr (2004), users tagged content freely with no controlled vocabulary. The “tag cloud” era. Twitter hashtags (2007) extended this to real-time discourse. The power: anyone can tag anything. The weakness: synonyms, typos, no structure. “Alchemy,” “alchemy,” “alchemia,” and “transmutation” become four unrelated tags.

Knowledge graphs emerged as a structured alternative. Google’s Knowledge Graph (2012) and Wikidata (2012) model knowledge not as categories but as entities and relationships. Aristotle isn’t “filed under Philosophy” — he’s an entity with properties (born: Stagira, teacher of: Alexander, student of: Plato) connected to other entities. This is fundamentally different from all tree and facet models: there are no categories at all, only a web of typed links. It’s closer to Bush’s associative trails than to Porphyry’s branching tree.

14. The Embedding: Vector Clustering (2020s)

Neural language models introduced a completely different approach. Instead of humans choosing categories, models embed texts as points in high-dimensional vector space. Texts with similar meaning end up near each other. Clustering algorithms find the groupings.

We did this with Source Library — embedding 5,993 book summaries and discovering 48 clusters. The algorithm found groupings no human would have proposed: a “Thirty Years’ War Pamphlets” cluster connecting Frankfurt book fair catalogs with Rosicrucian texts. But embeddings have no labels, no explanations, no stability. Blavatsky’s Isis Unveiled ended up in “Christian Kabbalah” — wrong as a label, but revealing as a neighborhood.

15. LLMs as Classifiers (2024–2026)

Large language models changed what’s possible. They can read a book’s title, author, year, language, and summary, understand what kind of text it is, and assign tags from a controlled vocabulary — with the judgment of a specialist librarian and the speed of a database.

This is new. Previous automated approaches were either statistical (keyword matching, embeddings) or rule-based (Dewey’s ten classes). LLMs combine the flexibility of human judgment with the scale of automation. The Library of Congress has been experimenting with AI-assisted subject heading assignment. OCLC (the organization behind WorldCat) is exploring LLM-powered cataloging. Academic libraries are testing GPT-based classification against human catalogers and finding competitive accuracy.

What none of these projects have done, as far as we know, is combine LLM classification with a philosophically grounded faceted vocabulary. Most AI cataloging projects bolt an LLM onto existing systems (Dewey, LCSH). We designed the vocabulary itself from first principles, using the classification theories of Llull, Bacon, Porphyry, Gessner, and Ranganathan. Which brings us to what we built.

16. The Synthesis: Source Library’s Faceted Tags (2026)

Our new system combines all of these ancestors:

• Llull’s combinatorial logic: six independent facets, freely combined. 68 tag values produce roughly a million unique intersections.
• Porphyry’s differentia: every tag value has a one-sentence boundary definition.
• Bacon’s cognitive grounding: the “epistemic mode” facet records how a text generates knowledge.
• Gessner’s multiple access points: no single facet is primary.
• Ranganathan’s PMEST: orthogonal facets whose intersection uniquely locates any text.
• Folksonomy’s lesson: controlled vocabulary matters. We didn’t let the model invent tags.
• Knowledge graph thinking: books connect to other books through shared facet values, not just shared categories.
• Embedding discovery: the 48 clusters remain as a “surprise me” feature.

A Gemini model reads each book’s metadata and selects tags. The cost to classify 13,000 books: about 70 cents.

The Loop

Aristotle → Porphyry → Dionysius → Chinese Sibu → Ibn al-Nadim → Llull → Ramus → Gessner → Bacon → Leibniz → Hooke → Comenius → Samuel Johnson → Linnaeus → Diderot → Dewey → Otlet → Ranganathan → Bush → folksonomy → knowledge graphs → embeddings → LLM facets.

What strikes us is how slow the real breakthroughs were. The tree model lasted 1,200 years (Porphyry to Bacon). The enumerative model lasted 140 years (Dewey to embeddings). Llull’s combinatorial insight was 700 years ahead of Ranganathan. Ramus’s pedagogical trees shaped how a nation organized its government, and nobody in library science seems to have noticed. The Chinese Sibu system and the Islamic Fihristboth predate European innovations by centuries. The ideas were there all along — they just took time to be heard across traditions.

There’s something fitting about a library that contains Aristotle, Porphyry, Llull, Ramus, Gessner, Bacon, Leibniz, Hooke, Comenius, Kircher, Samuel Johnson, Linnaeus, and Diderot using their ideas to organize itself. The books taught us how to classify the books.