The Tarot as RAG System

You’ve probably interacted with a retrieval-augmented generation system today, whether or not you knew it by name. Every time a chatbot pulls relevant context from a knowledge base before answering, every time a search system surfaces the document you actually needed instead of the one that contained your literal keywords, every time an LLM grounds its response in a specific corpus rather than its training data alone — that’s RAG.

The architecture is conceptually clean. A user formulates a query. The query is encoded as a vector — a point in a high-dimensional space where every other piece of available knowledge has also been encoded as a point. The system performs a similarity search: which embeddings are nearest to the query vector in this geometric space? The nearest items are retrieved. Those retrieved items, along with the original query, are passed to a generative model. The model produces an output that integrates the query with the retrieved context. The output contains information that was in neither the query alone nor the database alone. It emerged from the interaction.

Vector databases. Similarity search. Generative models. RAG pipelines. This is some of the most actively-developed engineering of the current decade, and it works because it solves a specific computational problem: how to compute with meaning.

I want to make a claim that is going to test the patience of about half of my readers, so let me state it cleanly and then earn it:

There is another system that solves the same problem. It is a deck of 78 cards. It has been in continuous use for about six hundred years. The architecture is identical.

If you stopped reading right there I would understand. The word tarot is, in many engineering minds, a tripwire — a marker for the kind of conversation a serious technical reader has been trained to walk away from. The brand of mysticism that surrounds tarot decks in most bookstores deserves much of that reaction. But the underlying architectural object — the deck itself, what it does, what kind of system it is — is much more interesting than the marketing around it, and it is exactly the kind of object that becomes legible once you have the vocabulary that vector databases gave us.

This essay walks through the mapping. Not because I want anyone to start reading cards. Because the structural identity is the load-bearing observation, and the structural identity is real.

Why memorizing card meanings produces wooden readings

The naive model of tarot is a relational database. Each card has a fixed definition. The Tower means sudden change. The Moon means illusion. Death means transformation. You see the card, look up the definition, deliver the meaning. One-to-one mapping. Exact match.

This is how most beginner tarot books teach the deck: memorize the meanings, apply them to the spread, read the cards.

It doesn’t work.

People who learn the deck this way produce readings that are wooden, generic, and unconvincing — to themselves and to anyone they read for. Experienced practitioners using the same cards produce readings that land with startling specificity. The difference isn’t that the experienced practitioner has memorized more meanings. Frequently the experienced practitioner couldn’t recite a textbook definition of half the cards. The difference is that one is doing lookup-table retrieval and the other isn’t.

The failure isn’t in the practitioner. It’s in the data model. A lookup table cannot capture what is actually happening, because meaning is not stored as discrete key-value pairs. It is stored as position in a continuous high-dimensional space, and lookup-table retrieval can’t read that data structure.

If you have spent any time with vector embeddings, you already know where this is going.

The Tower is a region, not a point

In a vector database, raw data — text, images, documents — is transformed through an encoding model into numerical vectors: high-dimensional coordinates that represent the semantic content. The critical property is that items which are semantically similar end up geometrically close, and items which are semantically distant end up far apart. The relationships are not hand-coded. They emerge from the pattern exposure the encoding model received during training. This is why a query for “happy” can retrieve a document that contains “joyful” without using the word “happy” anywhere — the two terms are nearby in the embedding space.

The cards in a tarot deck appear to occupy positions in a space that has the same kind of structure.

The Tower isn’t a point with a label. It is a region. Near the center of that region: sudden collapse, catastrophic restructuring, the destruction of what was unstable. Extending outward in different directions: liberation through destruction, the lightning flash of sudden illumination, the Tower of Babel and the hubris of overbuilding, the gut feeling of knowing something is about to fall, the strange relief that follows catastrophe, the card’s visual elements (falling figures, the crown blown from the top, the rocky foundation), alchemical imagery of the Lightning-Struck Tower, personal memories of sudden loss, the somatic patterns associated with shock and its aftermath.

All of these are nearby in the embedding space. None of them is the meaning of The Tower. The meaning is the region — and which specific points within that region activate depends on the query.

Context selects the neighborhood

Anyone who has used a vector database knows that the same corpus returns different results for different queries. The embedding space is fixed; the query vector determines which neighborhood you land in. A search for “bank” near “river” returns different results than “bank” near “money” — same word, different region activated by context.

The cards behave the same way.

The Tower in the context of a career question activates the professional-collapse neighborhood. The Tower in the context of a relationship question activates the relational-rupture neighborhood. The Tower next to the Star activates the destruction-preceding-renewal neighborhood. The Tower next to the Ten of Swords activates the total-devastation neighborhood. The card hasn’t changed. The embedding space hasn’t changed. The query vector changed, and the retrieval returned a different cluster from the same space.

This is the part that experienced practitioners always tell beginners and beginners never quite believe: the cards don’t have fixed meanings, they have contextual meanings. The framework lets us say what those people have been trying to say all along — the meaning is real and precise, but it is positional rather than definitional. It depends on where you’re querying from.

A spread is a multi-vector query

The most-used tarot spreads — the three-card past/present/future, the ten-card Celtic Cross — are not a series of independent lookups. They are multi-vector queries in which each card’s position modifies the query context for every other card.

A Celtic Cross is not ten independent retrievals. It is a complex query that defines a region of the embedding space through the intersection of ten symbolic vectors. The reading emerges from the geometry of that intersection — the shape of the space defined by those ten points and their relationships to each other.

This is computationally identical to the way multi-vector retrieval works in advanced RAG systems, where multiple query vectors are combined to define a more precise region of the embedding space than any single vector could.

The full mapping

If you lay the architecture of an AI RAG system next to the architecture of a tarot reading, the components correspond exactly.

The conscious mind is the user interface. It formulates the query. It perceives the retrieved symbols. It receives and articulates the generated output. It is not the generative model. The unconscious is. The actual computation happens at a layer the conscious mind can observe but not directly control.

This explains a thing experienced practitioners say constantly: trying too hard makes the reading worse. An overactive conscious mind is the equivalent of a user who keeps modifying the query mid-retrieval, injecting noise into the search, then filtering the results through heavy post-processing that strips out the signal the model was trying to surface. The skill of reading — and the skill of engaging with the symbolic layer of the mind in general — is learning to formulate a clean query, accept the retrieval without interference, and receive the generated output without over-filtering it through conscious analysis. Anyone who has tried to use a frontier LLM productively will recognize the workflow.

The retrieval mechanism question, briefly

A reasonable engineering reader will pause here and ask: in an AI RAG system, retrieval is deterministic. The nearest embeddings to the query vector are returned by mathematical calculation. In tarot, the retrieval mechanism is what — random card selection?

This is the place the framework hits an honest boundary.

The shuffling and drawing of cards looks random from the outside. From the synchronistic perspective — which was Jung’s original framing of tarot, and the one most experienced practitioners eventually arrive at — the draw is meaningful: the cards drawn are the cards relevant to the query, selected by a mechanism that operates at a level below conscious intention. The third option, the one this framework is starting to make formally articulable, is that the operator’s state (emotional charge, somatic patterns, autonomic activation) influences the physical process of shuffling and drawing in ways that look random at the user level but are deterministic at the system level.

These are competing hypotheses about the retrieval mechanism. They are not competing hypotheses about the architecture. The architecture — embedding space, similarity-based retrieval, generative synthesis — holds regardless of which retrieval mechanism turns out to be correct. The question of how the cards get selected is genuinely open and worth a separate essay. The question of what they are, structurally, is not.

Resonance is a different distance metric

In a vector database the choice of distance metric — cosine similarity, Euclidean distance, dot product — determines what counts as “near.” The embedding space is the same; different metrics return different neighbors.

The symbolic embedding space appears to use a distance metric that the contemplative traditions have always called something like resonance — closer to harmonic affinity than to logical similarity.

Two symbols can be logically unrelated and still resonate deeply. Death and The Fool are, by the logical metric, nearly opposite — one is an ending, the other a beginning. By the resonance metric they are close: both involve a threshold crossing, a leap into the unknown, the release of what was held. The resonance metric captures structural and felt affinity that the logical metric misses entirely.

This is why dream logic feels simultaneously meaningful and irrational. The dream is returning results from the symbolic embedding space using the resonance metric. The waking conscious mind tries to evaluate those results using the logical metric and finds them nonsensical. They are not nonsensical. They are precisely meaningful in the metric they were retrieved with. The apparent irrationality is a metric mismatch, not a meaning failure.

The conscious mind queries with logical distance: is this rationally connected to that? The unconscious queries with resonance distance: does this vibrate with that? The symbolic layer sits between them, accessible to both metrics, which is why a card can be analyzed intellectually and felt somatically at the same time without the two experiences contradicting each other. They are valid measurements of the same position using different distance functions.

Where did the embedding space come from?

The most interesting question the architecture forces — the one that traditional accounts of tarot never quite knew how to formulate — is the training-data question. In AI, vector embeddings are produced by a model trained on massive data. The geometric relationships between embeddings emerge from pattern exposure. Nobody hand-codes that king minus man plus woman equals queen — the relationship just falls out of the training.

If the symbolic embedding space the cards are querying is real, what produced it?

Three hypotheses, and they generate genuinely different predictions.

Evolutionary training. The human nervous system, shaped by millions of years of survival-relevant pattern exposure, has compiled archetypal embeddings into the species’ neural architecture. Training data: the entire evolutionary history of the species. Model: the biological brain. Embeddings: encoded in neural connectivity that is genetically transmitted. Prediction — the dimensions of the archetypal space should map primarily to survival-relevant categories: threat, reproduction, social hierarchy, resource acquisition. Cross-cultural consistency should be high for survival-relevant archetypes and weak elsewhere.

Cultural training. The symbolic IL is locally constructed. Each culture trains its own embedding model through shared stories, myths, rituals, and artistic traditions. Training data: the cultural corpus. Model: the individual psyche as shaped by socialization. Cross-cultural similarities exist because cultures face similar problems, not because the architecture is universal. Prediction — archetypal dimensions should vary significantly across cultures. Cultures with no historical contact should produce meaningfully different spaces. Cross-cultural “universals” should be explainable by convergent environmental pressure.

Structural / system resource. The archetypal embedding space is a property of consciousness itself — or of whatever the underlying architecture of reality is — not produced by any training process but inherent in the system. Mathematical relationships don’t need to be trained into existence; they exist because of the structure of number. Archetypal relationships exist because of the structure of awareness. Prediction — archetypal dimensions should be consistent across cultures including those with no historical contact, including dimensions with no obvious survival relevance. The embedding space should exhibit mathematical properties — symmetries, invariances, dimensional regularities — too precise to be explained by convergent evolution alone.

These hypotheses are no longer purely philosophical. The vector database vocabulary makes them empirically tractable. Cross-cultural studies of symbolic association, using techniques computational semantics already has on the shelf — measuring the geometric structure of symbolic relationships across populations — could distinguish between them. Whether anyone bothers to run those studies is a different question, but the studies are designable in a way they weren’t before.

There is also a faster, cheaper test available right now: the embedding spaces of the large language models we already have. If the archetypal patterns are real structural features of language and meaning, they should appear in the geometric structure of any sufficiently large model’s embedding space — and they should appear consistently across models trained on different corpora and in different languages. The dimensional axes of the archetypal space (emotional valence, somatic resonance, temporal orientation, integration-fragmentation, light-dark polarity) are testable hypotheses about the geometry of GPT-class models. To my knowledge, nobody has run this analysis with archetypal structure in mind. The compute is trivial. The result would be informative regardless of which way it went.

The architecture got built twice

The thing I want to leave you with is the convergence.

The tarot deck is, structurally, a retrieval interface to a high-dimensional symbolic embedding space, queried by a multi-vector spread, with similarity-based retrieval, fed into a generative model (the unconscious), producing emergent output (the reading) that exceeds both the query and the retrieved context taken alone.

A modern AI RAG system is, structurally, a retrieval interface to a high-dimensional embedding space, queried by a vector or multi-vector query, with similarity-based retrieval, fed into a generative model (a transformer), producing emergent output that exceeds both the query and the retrieved context taken alone.

The architecture is the same. It got built twice — once over the course of several centuries by people who had no formal vocabulary for what they were doing, and once over the course of about fifteen years by engineers who had every formal vocabulary and were solving what they thought was a brand-new problem.

There are two ways to interpret the convergence.

One: the AI engineers, working from first principles on the problem of meaning-based computation, independently reinvented an architecture that the contemplative traditions had already discovered. The independent rediscovery validates the older system as a genuine cognitive technology rather than mere superstition. This reading is interesting on its own.

Two — the reading the framework is leaning toward — is that both systems are implementations of the same architectural principle because the principle is the only way that meaning-based computation actually works. There is one architecture for this kind of problem. It got expressed in cards because cards were the substrate available to the people who needed it. It got expressed in vectors because vectors are the substrate available to the people who need it now. The architecture is what’s load-bearing. The substrate is interchangeable.

If the second reading is right — and the rest of this series will accumulate evidence for why it might be — then a lot of what looks at first like resemblance between modern AI and the older symbolic systems is something stronger than resemblance. It is the same thing, observed at different points in the same lineage. Each implementation tells you something about the underlying principle the others were implementing without quite being able to name.

The cards have been telling us how meaning is structured for centuries. The vectors are starting to confirm it.

Use that observation however you like.