Why Knowledge Graphs Generate Alpha
Who this is for
- Crypto VCs: spotting second-order risks, hidden correlations, and narrative shifts.
- Investors: reducing blind spots as domains scale faster than intuition.
- Business Developers & operators: understanding ecosystem structure, not just surface metrics.
- Researchers: combining symbolic structure with statistical models.
- Engineers: building systems that reason, not just retrieve.
New Knowledge as Information Theory problem
Knowledge
What is knowledge? For the purpose of this article, knowledge is a compressed model of the world that allows an agent — human or artificial — to make better decisions with fewer observations.
Knowledge must be sufficient and adequate. There’s a debate about truth vs. usefulness, about justification (is a lucky guess knowledge?), and apparently useless knowledge. That debate aside, we are looking for knowledge that helps an agent operate more efficiently.
In Web3, everyone I ran into is trying to maintain some kind of an internal model of the industry — by going to conferences, watching podcasts, and of course, spending an inordinate amount of time on crypto Twitter.
The problem occurs when no matter what you do, your internal world model is not complete and useful enough for what you are trying to do.
This is becoming a problem for everyone: investors, BD, marketers, entrepreneurs, degens, even engineers.
Needless to say, AI agents are in the same position, even though they can churn significantly more data.
Principle of Most Surprise
In information theory, information is not about truth, but about surprise. An observation carries information proportional to how much it changes our beliefs.
Data that confirms what we already expect has low informational value. Data that forces us to revise our model of the world has high informational value.
To catch Alpha, we need to focus on surprising information that can lead to better decisions. That’s what we call insight.
As we shall see, graphs make rare relationships cheap to query, making insights easier to find.
For example, the fact that a DeFi protocol X is deployed on an L1 network Y is nothing surprising (Baseline). We expect it to be deployed somewhere, and the question is just where exactly. Low signal.
But here is something non-obvious.
StarkNet (L2, 31 investors) and Lido (liquid staking) share 5 investors: Multicoin Capital, Paradigm, Coinbase Ventures, Three Arrows Capital, and Alameda Research. Two of those shared investors (3AC, Alameda) went bankrupt — and their positions in both tokens created cascading sell pressure across sectors that seemed unrelated.
Similarly, Ethena (synthetic dollar, 30 investors) and Morpho (lending) share 6 investors including Pantera Capital (tier 1), GSR, Nascent, Brevan Howard Digital. Cross-sector co-investment that creates hidden correlation: a liquidity crisis at any shared investor propagates to both.
A surprising bit of information that ends up being useful is an insight. Which may lead to alpha — the excess return everyone is looking for.
Graphs Help Maximise Surprise
We are talking about Graphs as an abstraction from discrete math, with vertices and edges. Not plots with axes X, Y.
Graphs are awesome, because an edge between two vertices may exist or not, and multigraphs can have even more than one edge. Graphs don’t magically create insight. They reduce the cost of representing and discovering non-obvious relationships.
A row in a table will always have some values, even if it is NULL for some column. A node in a graph can happily be alone — and when there’s an edge, it might mean something.
Graphs make these relationships explicit. Protocols are deployedOn Blockchains. Persons workOn Protocols, they workAt Companies. Tokens are usedBy Protocol. People can use Contracts and so on.
Not All Edges Are Useful for Insight Generation
The existence of a relationship between a Protocol and a Blockchain is obvious and implied. The fact that Aave is available on Base, and not on HyperEVM, is useful information, but is far from an insight. Every Protocol node will have an edge to a Blockchain.
A graph will happily store an edge from each Protocol to one or more Blockchains, but this is where it makes no difference to just storing it in a table.
On the contrary, a link between a Token and Person, especially combined with the fact that this Person is part of a Company developing a Protocol might be an insight.
Criteria for an insightful edge:
- Rarity
- Cross-domain
- Actionable
- Historically predictive.
Consider Uniswap. A spreadsheet of its investors reads like a who’s-who of crypto venture capital: a16z, Paradigm, BlackRock, Polychain Capital, Coinbase Ventures, and seven others. A separate spreadsheet of on-chain insider transactions for the Uniswap protocol shows 1,680 transactions with $154 million in net selling (data as of Feb 2026 from Token Terminal). Neither spreadsheet references the other.
In a knowledge graph, these two facts live three hops apart: Token:UNI → FUNDED_BY → InvestmentFund on one side, and Token:UNI → GOVERNS_PROTOCOL → Protocol:uniswap ← FOR_PROJECT ← InsiderTransaction on the other. The graph makes the divergence traversable: twelve tier-1 VCs entered through equity rounds and remain on the cap table, while on-chain insiders holding liquid tokens are exiting. Run the same traversal on Euler and you get the opposite signal — 26 investors and $34 million in net insider buying. Same edge types, opposite conclusions. The insight isn’t in either table. It’s in the path between them.
Ontologies, RDF, Triples
In late 2025 and early 2026 my feed is full of discussions about Ontologies, their role in modeling reality, their necessity for LLMs and agent interoperability. What is this all about?
Get Valeriy Zamaraiev’s stories in your inbox
Join Medium for free to get updates from this writer.
Ontologies are just knowledge about how to organize knowledge.
Graphs are useful, but they are just one way of modeling the world, which falls short when you want to model the world precisely and unambiguously, in a scalable, maintainable way.
In the example above, it might be important that Protocols can be deployedOn Blockchains, but not work for them — that one is between Persons and Companies. Companies, on the other hand, work on Products, of which Protocols are a kind, and there are other kinds of products.
Ontologies help answer the questions: “what is this?”, “is A a kind of B?”, how can A, B and C be related in principle. They provide the vocabulary and constraints that reasoning can operate on — the ontology defines that Man is a subclass of Mortal and that Socrates is a Man; logic then derives that Socrates is mortal.
Many examples of how this could be useful: Bitcoin is simultaneously a Blockchain, a network protocol, and with some tolerance, a coin (BTC). There is also a distinction between coins and tokens, if anyone cares about this still. Capturing this with just one node BTC in a graph would leave out a lot of useful information. Modern graph engines have ways of tackling this using multiple labels, but this explicitly has not much to do with graphs per se.
Resource Description Framework (RDF) is a technical name for the concept of organizing knowledge in the most generic but robust way. It has a bit of an academic flavor, and a few very important, scalable (and expensive) implementations.
Ironically, RDF was the initial vision for Web3 — Semantic Web.
From the perspective of RDF, knowledge is a collection of facts expressed as triples, which, it turns out, can model anything and everything in the world.
Why do we even call a bag of triples “Knowledge Graph” then? Where are the graphs?
Graphs in RDF are not explicit. They are emergent. Because triples are everywhere and they define everything and anything, treating each subject and object as a node, and every predicate as a relationship will be excessive.
It is only worth looking at the predicates that have value in the sense we have described above — that define edges that bring insightful information.
AI Agents
In 2026, any discussion is incomplete without discussing AI Agents. It is fascinating that connectionist AI (neural networks) that had won the AI race, is now turning to the long abandoned Symbolic AI, aka Good Old-Fashioned AI or GOFAI, to complement it in the areas where connectivism is still weak — being grounded in specific facts and logical reasoning.
Constructing the Knowledge Graphs
No surprise, if we need to extract Entities/Relationships from text, LLMs are the way to go in 2026.
But agentic flows could push this further — not only extract, but actively find what’s missing in the graph, obtain the data source, deciding along the way among multiple data sources, and fill in the gaps in the KG.
Consuming the Knowledge Graph Data
It turns out, AI agents are helpful not only in constructing Knowledge Graphs, but also in consumption.
As we are working on DYOR.bot, the current alpha version agent is able to explore the Knowledge Graph iteratively, without much prior knowledge about its structure.
Thanks to the ontology and agentic workflows, a sequence of questions like: “ok, let me see which node the user means”, “ok, let’s see which attributes and links are supported”, “ok, let’s take a look at other related types” — all expressed in suitable SPARQL queries — lead to the desired outcome.
An agent with a suitable ontology untangles complex knowledge magnitudes faster than a human would do — no surprise, writing queries is just coding.
Agents excel where the search space is large, structured, and partially unknown.
Given a proper graph (Labeled Property Graph, as opposed to RDF store), agents can find the results easier and faster — because the nodes and relationships are explicit — at an apparent loss of generality, and ability to dig very deeply.
What this means is that we need both: RDF as a canonical, expressive, foundational truth, and LPG as derived, fast, agent-friendly projection.
This is the architectural approach we converge on at DYOR.bot after considering many constraints.
Limits and Costs of Knowledge Graphs
Knowledge Graphs are not a universal solution. For a lot of things, vector search plus LLMs is cheaper, faster, and simply good enough. If all you need is retrieval, summarization, or a one-off answer, graphs are usually overkill.
Graphs also don’t magically produce insight. They just make it cheaper to notice relationships that are easy to miss otherwise. They only start to matter when structure itself is part of the signal: when provenance matters, when relationships are rare or asymmetric, and when context has to survive across time and domains.
The real cost of Knowledge Graphs isn’t storage or compute, it’s attention. Someone has to decide what things are, how they can relate, and which distinctions are worth keeping. Ontologies evolve. Relationship extraction is noisy and needs constant correction. None of this is free. The payoff is also not linear: value tends to look disappointing early on and only compounds once the same structure gets reused across many questions, agents, and workflows.
Knowledge Graphs are the wrong tool when:
- the questions are static
- the world you’re modeling doesn’t really change
- you don’t care where an answer came from
- your insights come from aggregation, not relationships.
Used in the wrong place, they turn into expensive abstractions. Used in the right place, they preserve meaning exactly where intuition and unstructured search start to fail.
Takeaway
In fast-moving domains like crypto, the bottleneck isn’t access to information, it’s maintaining a coherent internal world model.
Knowledge Graphs don’t replace judgment, but they make rare, meaningful relationships visible before they become obvious.
