Why No Schema Can Describe the Whole World — and What It Means for Search

The article builds on the earlier thesis that reality is relational and that a properly constructed relational model is consistent and infinitely extensible, and it reveals the other side of that claim: the model is extensible precisely because it is incomplete — and incompleteness is not a defect but the price of consistency, which the author supports with three independent arguments (Gödelian, cardinality-based, and Kolmogorovian). From this he draws consequences for the architecture of information systems and for the shift from deduction to induction: whereas in the closed world of finite domains the deductive querying of the relational model works, the open world of the internet demands induction — full-text, vectors, and language models are not a primitive substitute for an elegant SQL query, but a necessary tool for understanding a world deeper than any finite schema.

Imagine an online shop that sells socks. A Product table, attributes color=red, size=39. The SQL query works perfectly. The model is consistent, clear, extensible. Now imagine you want to search every online shop on the internet. One has color, another has colour, a third hid it in a free-text description "beautiful red socks with stripes." What was structured at the level of a single system turned, from the perspective of the whole, into an unstructured heap.

This transition is not an organizational accident. It is an inevitable consequence of how the world is arranged — and of how any finite model runs up against its limits. The previous article formulated the thesis that reality is relational and that a correctly constructed relational model is consistent and infinitely extensible. Both hold. But this analysis shows that the story has another side: the model is extensible precisely because it is incomplete. And incompleteness is not a defect — it is the price of consistency.

This conclusion rests on three independent arguments — the Gödelian one (a database schema as a formal system), the cardinality one (a countable schema against an uncountable reality), and the Kolmogorovian one (the incompressibility of general configurations). It has direct consequences for the architecture of information systems: full-text search and vectors are not a primitive alternative to an elegant SQL query. They are an essential tool for coming to know a world that is deeper than any schema.

Where the original article left off

The original article started from the axiom of the consistency of reality and from the observation that every M:N relationship, under deeper analysis, reveals a hidden entity. From this it derived six steps: reality is consistent; a relational model with exclusively 1:N relationships is isomorphic to the part of reality it describes; extending the model means adding new entities and relationships; a new aspect of reality cannot contradict the existing ones; adding to the model cannot create a contradiction; and this holds for any extension.

The article honestly identified step two — the isomorphism between model and reality — as the weakest link in the chain. This analysis shows that the problem runs deeper.

Does reality have a bottom?

The hierarchy of matter provides a direct empirical argument for the infinite extensibility of the relational model:

molecule → atoms → nucleus + electrons → protons/neutrons → quarks → (strings?)

Each level adds new entities and new 1:N relationships without disturbing the higher ones. A water molecule still has two hydrogens and an oxygen — the fact that a proton consists of three quarks breaks nothing at the higher level. This is exactly the pattern of incremental extension without destructive changes that the original article predicts.

The Planck length (~1.6 × 10⁻³⁵ m) is often interpreted as the limit below which the notion of space loses physical meaning. Quantum gravity, string theory, and loop quantum gravity problematize this limit in different ways. For our argument, however, it does not matter whether the depth is truly infinite — it is enough that it is practically unlimited relative to the capacity of any finite model.

At the quantum level a complication arises. Entanglement — two photons sharing a state nonlocally — poses a challenge for a purely relational interpretation. Is it a 1:N relationship through a hidden entity, namely a shared quantum state? Or an entirely different type of relationship?

Ontic structural realism (Ladyman, French) offers an answer: it is precisely here that relations turn out to be more primary than relata. Entangled photons have no independent properties — they are their relation. At the quantum level, relations are not between entities. Relations are the entities.

The Gödelian moment: consistency, or completeness

The original article claimed that the relational model is consistent and infinitely extensible. That is true — but it is a weaker claim than the reader automatically assumes. For infinite extensibility does not imply that the model can fully describe reality.

The stronger formulation reads: reality is infinitely deep. At every level of description there are further entities, further relations, further attributes that are not yet captured in the model. No finite model can be a complete description of reality — it is always merely an approximation.

A correctly constructed relational model has two properties. Consistency — no internal contradiction ever arises, which follows from the isomorphism with a consistent reality. And incompleteness — there are always aspects of reality that the model does not capture. A system is either complete or consistent — but not both.

A database schema as a formal system

The previous version of the article formulated the relationship to Gödel's theorem cautiously, as a parallel or an analogy. But the connection is stronger. A relational database schema is a formal system in the Gödelian sense:

Tables correspond to predicates. Rows are ground facts. Integrity constraints — primary keys, foreign keys, CHECK, UNIQUE, NOT NULL — are axioms. SQL queries, views, and triggers act as inference rules. And query results are theorems — derived statements.

This is not a metaphor. Codd constructed the relational model on the foundations of first-order predicate logic — relational algebra is a subset of it. SQL is a language for formulating statements and queries within this system. And the key point: every practical database schema is rich enough to express the arithmetic of natural numbers. A single table with one INTEGER column and integrity constraints suffices. This satisfies the condition of Gödel's first incompleteness theorem.

What then follows from it: there exist true statements about the data that cannot be derived from the schema, the integrity constraints, and the stored data. In database terms: there exist valid questions that a structured query over the given schema cannot answer.

Socks as proof

Let us return to the opening example. Within a single online shop, the model is consistent and sufficient for its domain. But from the perspective of a search engine that searches across all online shops, the structured data of one system is an unstructured heap in the higher system.

And now the Gödelian leap: the problem is not the indiscipline of schema designers. Even if a single global schema for socks existed — and one in fact does, schema.org defines a Product type with the attributes color, size, material — it would still be impossible to define in advance a schema for everything that will ever be on the internet.

Tomorrow someone invents a product that fits no existing category. The day after, someone writes a poem about socks — and it has different relevant attributes than color and size. A year from now someone creates a medical dataset on the effect of compression stockings on varicose veins — and the relevant attributes are pressure in mmHg, material composition, clinical study.

Every schema extension is legitimate and consistent — but no finite schema covers all future extensions. And that is exactly what Gödel's theorem says: a formal system cannot be simultaneously consistent and complete.

Three independent paths to the same conclusion

The Gödelian argument is not the only one. The incompleteness of a finite model of reality can be supported by three independent paths, each from a different mathematical tradition.

The Gödelian argument (logic). A database schema is a formal system rich enough for arithmetic. Gödel's first theorem applies directly: there exist true statements about the data that the system cannot derive. Strength: formally the strongest. Limitation: it requires accepting the schema–formal-system mapping.

The cardinality argument (set theory). The set of all possible relations in reality is uncountable — every real-valued parameter (temperature, position, time) generates a continuum of possible values and relationships. Any finite schema captures countably many types of relations. A countable subset of an uncountable set is always a proper subset. The schema is necessarily incomplete. Strength: it does not require Gödel, Cantor suffices. Limitation: it assumes that the relevant relations form an uncountable set — but physics empirically supports this, as continuous quantities generate a continuum of states.

The Kolmogorov argument (information theory). The Kolmogorov complexity of an object is the length of the shortest program that generates it. A database schema is a compression algorithm — it captures regularities in the data. But the incompressibility theorem says that almost all strings are incompressible — they have no structure shorter than themselves. Applied to reality: almost all configurations of reality are schematically incompressible — they cannot be captured by a schema shorter than the data itself. Strength: information-theoretically elegant. Limitation: it assumes that reality contains incompressible configurations — but noise, chaos, and quantum randomness empirically support this.

Three arguments from three traditions, and yet they converge: a finite structured model cannot be a complete description of reality.

Why this is not a flaw

A model is useful not because it exhausts reality, but because it approximates it consistently. And whenever we find that the approximation is not enough, we can extend it without destroying what already exists. This is exactly that infinite extensibility — it is not a promise of completeness, it is a promise of safe growth.

A city map is useful even though it does not record every crack in the pavement. Its value lies in the fact that what it does record, it records correctly — and it can be supplemented with a layer of cracks without breaking the layer of streets.

From deduction to induction: why reality cannot be searched by rules

A relational database operates in the closed-world mode (the closed-world assumption): everything that is not in the database is false. An accounting system — I know what an invoice, a customer, a VAT rate are. The schema is finite, complete for the given domain. I query in a structured way, I define a class, a relation, an attribute, and the system deduces the answer. It works precisely because the domain is finite and known in advance.

Reality is an open world. I do not know in advance which entities exist. I do not know which attributes are relevant. I do not know which relations hold among them. And above all — I do not know what I do not know. In an open world, the falsity of a statement cannot be inferred from the absence of a record. One can only say: it is not in my model.

From this follows a fundamental consequence for search. Search in a closed world is deductive — from rules to data. "Find an entity of type X with attribute Y in relation Z." The system applies rules to the stored facts and deduces the answer. The result is deterministic and complete within the model.

Search in an open world must be inductive — from data to patterns. "Find anything that is semantically close to this query." The system does not search according to predefined classes and relations — it estimates relevance from surface signals. The result is probabilistic and always incomplete.

And this is not a choice in system design — it follows from the nature of knowledge itself. You cannot deductively search something whose structure you do not know. It would be like querying with SQL a table whose schema does not exist.

Every search engine is an inductive machine

This insight illuminates the entire history of web search.

PageRank (Google, 1998) — induction from the link graph. It does not say "this page is about X." It says: many other pages link to this page, it is probably important. A heuristic over unstructured reality.

BM25 (Robertson & Walker, 1994) — induction from frequency statistics. It does not say "this document contains an entity of type Y." It says: a term from the query occurs in the document more often than chance would warrant, it is probably relevant.

Vector similarity (BERT and successors, 2018/2019+) — induction from distributional patterns. It does not say "this text is in relation R with the query." It says: the distribution of contexts of the two texts is similar, they probably speak about the same thing.

Large language models (GPT, Claude, 2022+) — induction from the entire training corpus. They do not derive from rules — they generate an answer on the basis of learned patterns. When they "know" the answer, it is because they have seen enough inductive signals — not because they deduce it from an axiom. A subtle nuance: chain-of-thought reasoning resembles deduction — the model chains steps, applies rules, derives conclusions. But the premises it starts from are themselves induced from the training data. It is deduction over induced premises — a hybrid in which the inductive foundation determines the quality of the deductive superstructure.

None of these methods can be complete — and this is not a failure of development. It is a consequence of the Gödelian limit.

The logical chain in seven steps

The whole argument closes into a single chain:

1. Reality is relational (Codd; ontic structural realism — Ladyman, French; Whitehead; Kaipayil).

2. A relational model of reality is consistent and extensible (the original article).

3. But reality is infinitely deep → a finite schema will not describe it (Gödel, Cantor, Kolmogorov).

4. What is structured locally is unstructured globally (the online-shop example).

5. Searching the unstructured = an open world → the deductive approach fails.

6. We must switch to induction: full-text, vectors, language models.

7. Induction is always incomplete — but it is intellectually honest.

Vector embeddings and knowledge graphs: approaches and their limits

A vector embedding (BERT, language models) is an attempt at an implicit relational model. When a model encodes text into a vector, it captures semantic relationships — but not explicitly as entities and links, rather as distances in a latent space. It is an approximation of an infinitely complex relational structure by a finite-dimensional vector.

From this follows: an embedding is always lossy. It reduces an infinitely deep relational structure to a finite number of dimensions. The quality of an embedding is a measure of how well this reduction preserves the relations relevant to a given task — not all relations in general.

Knowledge of the domain determines the appropriate modeling approach — and the type of inference. At one end of the spectrum stand systems with complete knowledge of a finite domain — an accounting system, a civil registry — where a pure relational model with SQL and pure deduction works. A step further are systems with partial knowledge, such as an online shop or a news portal, which combine a structured schema with full-text search — deduction with induction. A general web search engine has minimal a priori knowledge of structure and relies on full-text and vectors — almost pure induction. And at the opposite end of the spectrum stand purely vector-based systems and language models, which operate without any predefined structure and perform pure induction — semantic similarity, exploration of an unknown space.

This scale is not a ladder from worse to better. It is a scale of epistemic honesty — each approach acknowledges a different extent of prior knowledge and chooses the corresponding type of inference.

RAG and hybrid search (Vespa, Elasticsearch with vectors) are an attempt at a compromise: what you can structure, you solve with a relational model, filters, metadata, facets — by deduction. What you cannot structure, you solve with full-text and vectors — by induction. The moment of the query decides which layer dominates.

Knowledge graphs (Wikidata, Google Knowledge Graph) represent an attempt at explicit relational extraction from unstructured data — converting an open world into a closed one. They work — but they are always incomplete. Wikidata has over 100 million items and roughly 1.65 billion statements. The web has hundreds of billions of pages. The ratio of captured relations to existing ones is negligible — and always will be, regardless of the resources invested. A knowledge graph is a useful partial deduction over an otherwise inductive system — not a replacement for induction.

Counterarguments

"The relational model can handle unstructured data too"

Technically, any text can be stored in a relational database — a CLOB or TEXT column. But in doing so we give up what the relational model offers: structured queries. Stored text without extracted relations is unstructured data in a relational wrapper — functionally equivalent to a full-text index.

"Language models can extract relations automatically"

They can — and they do (named entity recognition, relation extraction, structured output). But the extraction is always incomplete (the model captures only the relations it was trained on), error-prone (hallucinations, bad disambiguation), and requires a predefined schema, which brings us back to the problem of a priori knowledge.

Language models push the boundary of what can be structured — but they do not remove the fundamental limit.

"Gödel's theorem does not apply to databases"

The objection has two versions. The weaker one: a database schema is not a formal system. This version is refuted by the mapping described above — the relational model is built on first-order predicate logic, tables correspond to predicates, integrity constraints to axioms.

The stronger version: Gödel's theorem concerns the derivational capabilities of a system, not the descriptive ones. A database does not derive theorems, it stores facts. This is a legitimate objection. The answer: an SQL query is a derivation — it combines stored facts with rules (the schema, JOINs, WHERE clauses) and produces new statements. A VIEW is a derived statement par excellence. And integrity constraints are axioms that the system enforces.

One open question remains: Gödelian incompleteness concerns recursively axiomatizable systems. A database schema is finite (hence trivially recursively axiomatizable), but it is a question whether its expressive power always reaches the level needed for Gödel's theorem. That is why two independent arguments stand above — the cardinality one and the Kolmogorov one — which prove incompleteness even without Gödel.

What follows from this

For the theory of modeling. The relational model is the best known tool for modeling the known structure of reality. Its consistency and extensibility are real and practically valuable. But its domain of application is bounded by the degree of a priori knowledge of the structure — and with it, the type of inference we perform over the model.

For the architecture of information systems. Correct architecture is not a choice between a relational and an unstructured approach. It is a decision about where the boundary of your knowledge of the domain lies — and assigning the right tool and the right inference to each zone. Known structure → relational model, deduction. Partially known → hybrid. Unknown → full-text, embedding, language model, induction.

For coming to know the world in general. Every model of reality is a map, not the territory. A map can be consistent — and it should be. But it will never be complete. And the pursuit of completeness at the expense of consistency — cramming unstructured data into a rigid schema, deductively querying an inductive problem — is worse than admitting incompleteness.

The transition from deduction to induction is not a concession — it is intellectual honesty. Where you know the rules, deduce. Where you do not, induce. And never pretend that you are inducing when you are in fact guessing — nor that you are deducing when you in fact have no axioms.

Open questions

1. Formalizing the isomorphism. Can the relationship between a relational model and the corresponding part of reality be formalized using the notion of isomorphism from category theory? If so, the argument about consistency and incompleteness would become a formal theorem.

2. Quantum relational ontology. Is entanglement modelable as a 1:N relationship through a hidden entity (a shared quantum state), or does it require an entirely new type of relation?

3. The embedding as an approximation. Is there a theoretical limit to the quality of an embedding as a function of the dimensionality and the complexity of the relational structure it approximates? The Johnson–Lindenstrauss lemma suggests there is — it shows that n points in a high-dimensional space can be projected into O(log n / ε²) dimensions while preserving their pairwise distances with error ε.

4. A practical test. Can it be empirically verified that systems designed according to this hierarchy (deduction where you know, induction where you do not) exhibit lower technical debt and higher extensibility than systems with a uniform approach?

5. The limits of induction. Is there a theoretical limit to how well an inductive system (BM25, embedding, language model) can approximate deductive precision? And is this limit constant, or does it shift as the dimensionality of models grows?

Conclusion

The original thesis — what is relational is infinitely extensible, as long as it does not contradict reality — holds up. But its full meaning emerges only with an addition: the model is extensible precisely because it is incomplete. Incompleteness is not a defect, it is the price of consistency.

And this price has direct architectural consequences. Search in reality cannot be performed by rules alone — classes, relations, attributes — because reality is infinitely more complex than any finite schema. The deductive approach works in the closed world of finite domains. For the open world — the internet, news, maps, general human knowledge — the only option is induction: full-text, vectors, language models.

Full-text search of unstructured data is not a primitive alternative to an elegant relational query. It is an admission that the world is deeper than any schema we can give it. And that admission is more accurate than the illusion of structure where no predefined structure exists.

Sources and references

The relational model and database theory

• Codd, E. F. (1970). A Relational Model of Data for Large Shared Data Banks. Communications of the ACM, 13(6), 377–387. — The founding text of the relational model. Codd built relational algebra on the foundations of set theory and first-order predicate logic. Available: doi:10.1145/362384.362685

Mathematical foundations of the incompleteness arguments

• Gödel, K. (1931). Über formal unentscheidbare Sätze der Principia Mathematica und verwandter Systeme I. Monatshefte für Mathematik und Physik, 38, 173–198. — The first incompleteness theorem: in every sufficiently rich consistent formal system there exist true statements that the system cannot prove.

• Cantor, G. (1891). Ueber eine elementare Frage der Mannigfaltigkeitslehre. Jahresbericht der Deutschen Mathematiker-Vereinigung, 1, 75–78. — The diagonal argument proving that there are more real numbers than natural numbers (the uncountability of the continuum). The basis of the cardinality argument.

• Kolmogorov, A. N. (1965). Three Approaches to the Quantitative Definition of Information. Problems of Information Transmission, 1(1), 1–7. — The definition of algorithmic (Kolmogorov) complexity. The incompressibility theorem: almost all strings have no structure shorter than themselves.

Ontic structural realism

• Ladyman, J. (1998). What is Structural Realism? Studies in History and Philosophy of Science, 29(3), 409–424. — The distinction between epistemic and ontic structural realism. The thesis that relations are ontologically more primary than relata.

• French, S. & Ladyman, J. (2003). Remodelling Structural Realism: Quantum Physics and the Metaphysics of Structure. Synthese, 136, 31–56. doi:10.1023/A:1024156116636 — A defense of ontic structural realism in the context of quantum mechanics.

• French, S. & Ladyman, J. (2011). In Defence of Ontic Structural Realism. In: Bokulich, A. & Bokulich, P. (eds.), Scientific Structuralism. Boston Studies in the Philosophy and History of Science, vol. 281. Springer. doi:10.1007/978-90-481-9597-8_2

• Ladyman, J. & Ross, D. (2007). Every Thing Must Go: Metaphysics Naturalized. Oxford University Press. — A comprehensive defense of ontic structural realism with an argument for the "scale relativity of ontology."

• Ladyman, J. (2020). Structural Realism. In: Zalta, E. (ed.), Stanford Encyclopedia of Philosophy. plato.stanford.edu — A survey article summarizing the main positions and arguments of structural realism.

Relational ontology

• Whitehead, A. N. (1929). Process and Reality: An Essay in Cosmology. Macmillan. — Process philosophy. Reality as a network of interconnected processes (events), not independent substances.

• Kaipayil, J. (2009). Relationalism: A Theory of Being. Bangalore: JIP Publications. — Ontic relationalism: being is relational, entities are constituted by their relations. PhilArchive

• Kaipayil, J. (2003). Human as Relational: A Study in Critical Ontology. Bangalore: Jeevalaya Institute of Philosophy. — An application of relational ontology to the human person.

Search and information retrieval

• Robertson, S. E. & Walker, S. (1994). Some Simple Effective Approximations to the 2-Poisson Model for Probabilistic Weighted Retrieval. In: Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 232–241. ACM/Springer. — The key paper introducing BM25, to this day the most widely used ranking function in full-text search.

• Robertson, S. E. & Zaragoza, H. (2009). The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval, 3(4), 333–389. doi:10.1561/1500000019 — A survey of the probabilistic framework for IR, including BM25 and BM25F.

• Brin, S. & Page, L. (1998). The Anatomy of a Large-Scale Hypertextual Web Search Engine. Computer Networks and ISDN Systems, 30(1–7), 107–117. — The original paper describing PageRank and the architecture of Google.

• Devlin, J. et al. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Proceedings of NAACL-HLT 2019. (Preprint: arXiv:1810.04805, October 2018.) — A model for contextual text representation that laid the foundation for semantic search using vector embeddings.

Theory of dimensionality reduction

• Johnson, W. B. & Lindenstrauss, J. (1984). Extensions of Lipschitz Mappings into a Hilbert Space. Contemporary Mathematics, 26, 189–206. — The Johnson–Lindenstrauss lemma: n points in a high-dimensional space can be projected into O(log n / ε²) dimensions while preserving their pairwise distances with error ε — where n is the number of points, not the original dimension of the space.

The concept, structure, and editorial direction of this article are the work of the author, who prepared the content outline, established the key theses, and directed the entire creative process. Generative AI (Claude, Anthropic) was used as a tool for research, fact-checking, and elaborating the author's outline.

The author edited the outputs on an ongoing basis, verified the key findings, and approved the final version. No part of the text was published without human review. All factual data were verified against the publicly available sources listed in the text.

This procedure complies with the requirements of Art. 50 of EU Regulation 2024/1689 (AI Act) on the transparency of AI-generated content. #poweredByAI