Creating Clarity
By Andras Hejj, Co-Founder and CTO, motif
We spent two years trying to make AI work for financial advisory, and the first year we kept hitting the same wall everyone else hits.
You connect an LLM to your data. You add RAG. It answers questions, sounds confident, and then cites a regulation that was updated four months ago like nothing changed. Or it treats a Q3 earnings miss and a Q1 beat as interchangeable because both live in the same vector space. Or a client asks “why did European bank stocks sell off” and the model retrieves three vaguely related paragraphs instead of tracing the obvious chain from ECB policy change to sovereign yield movement to bank margin compression to equities following.
Any junior analyst could do that in their head. AI couldn’t. So we stopped trying to make retrieval smarter and started building something structurally different.
Everyone does RAG. Almost nobody does it well.
RAG is the default architecture for AI products right now. Chunk your documents into pieces, turn them into vectors, find the most similar chunks when someone asks a question, feed those chunks to an LLM. It works for simple Q&A and it’s why every AI product can answer basic questions about your data.
But it has no idea what caused what or that relationships change, because it pattern-matches on semantic similarity, not on actual structure.
GraphRAG is supposed to fix this. Build a knowledge graph, retrieve from the graph. Microsoft published good research on it and the community has been running with it. The concept is right.
But building a knowledge graph that’s actually useful means you have to understand what you’re putting into it. Not chunk it. Not embed it. Understand it. Extract the entities, map the relationships, identify the claims, tag when things happened, score how confident you are, track what contradicts what.
That’s expensive, slow, and takes serious compute.
And it does not scale to indexing the entire internet, which is what most AI companies are trying to do, so they skip it. They build thin graphs with flat edges that tell you “these two things are connected” and nothing else, which is barely better than the vectors they were trying to replace. We went the other way.
What Clarity actually does
We don’t try to ingest everything. We understand everything we ingest, whether that’s analyst research, earnings reports, regulatory filings, macro data or proprietary material.
When a document enters Clarity, we break it apart into meaning, not chunks. Who are the actors, the people, the companies, the institutions? What events happened, policy decisions, earnings releases, geopolitical shifts? What claims were made, by whom, based on what evidence? And how does all of it connect?
Every entity gets a temporal layer. Not a timestamp on a database row but a full history of when it first appeared, how it changed, what caused the change, and whether newer information contradicts older information. Every relationship between entities carries its own metadata, including which sources established it, how confident those sources are, when it was last validated, and how it’s evolved.
We call this Contextual Edge Enrichment. In most knowledge graphs, an edge between “Federal Reserve” and “US Interest Rates” is a flat line that tells you they’re connected and nothing else. In Clarity, that edge is a first-class object. It knows when the connection was established, who said so, how reliable they are, and what changed since.
The difference is between “these things are related” and “here’s how, since when, according to whom, and with what confidence.” One is a search result and the other is reasoning.
Think about how an actual analyst works. They don’t memorise documents and fuzzy-match against them later. They build a mental model. They know the players, the relationships, the history, the conflicts and they update that model as new information comes in. They can trace a chain of cause and effect in their head because they’ve structured the information, not just stored it.
Clarity is that mental model, externalised, continuously updated and shared across every agent in the system.
The agents
Once the graph exists, agents operate on it. Not one agent but swarms of them. There’s solid research behind this approach. MiroFish, an open-source multi-agent prediction engine, showed that specialised agents with structured knowledge access outperform single-model approaches on complex reasoning. We took that pattern and adapted it for wealth advisory.
Research agents scan and ingest continuously. They don’t just add data, they challenge the existing graph. New earnings report contradicts last quarter’s forecast? Both states get stored, with the causal transition between them. Two sources disagree? The agents weigh authority against recency before anything gets committed, and that’s how confidence scores get built.
When someone asks a question, analysis agents traverse Clarity. They walk causal chains with confidence that decays across depth, pulling time-stamped, sourced evidence at every step. “What caused European bank stocks to sell off” becomes ECB policy to sovereign yields to margin compression to specific equity movements, with every link grounded and every step traceable.
Scenario agents go further. “What if the Fed cuts rates by 100 basis points?” The agent finds the Fed entity in Clarity, traces the forward causal chain through every connected node, and for each one linked to a real asset, estimates direction and magnitude of impact. Counterfactual reasoning over structured, temporal data. This is work rooted in causal knowledge graph research, encoding not just what’s connected but how and why, applied to live financial data.
Quality agents check everything. Every score cites specific evidence, following approaches like Brain-in-the-Fish. Not “factual accuracy: 85%” but “85%, supported by slide 3 which correctly reports the ECB decision, weakened by slide 5 which overstates the yield impact.” Traceable, auditable, and the kind of thing you can hand to a regulator.
A real example
March. ECB announces its rate decision. Inside Clarity, the policy entity updates with the new decision, tagged with source and timestamp. Causal edges propagate: sovereign bond entities get updated yield signals, European bank equities get margin compression indicators, sector-level scores recalculate. By the time the agents finish traversing the chain, the full path from policy decision to portfolio recommendation is traced, sourced and ready. No analyst wrote a memo or updated a spreadsheet because the system reasoned through it.
RAG would have retrieved the press release, but Clarity understood what it meant.
Why this matters if you’re a financial institution
Compliance. Every output in Clarity traces to specific nodes and edges, each with provenance and temporal context. When a regulator asks how a recommendation was formed, you show them which data points, when they were validated, and what the confidence was. We built for MiFID II, GDPR, and the EU AI Act from day one, not as an afterthought. Clarity is live under Swiss financial supervision, backed by institutional partners including Liminal, Temasek’s Web3 venture studio.
Personalisation that moves. Clarity tracks how preferences, risk appetite, and decision patterns evolve. Behaviour shifts, the graph reflects it, and advisory adapts.
Integration in weeks. Clarity ships as an API and SDK. It connects to your existing onboarding, client portals and advisory apps with configurable tone, language and prompts. You don’t hire a graph database team because the system is already populated, already reasoning and already under regulatory supervision.
We’ve built something novel, and it’s actually different to what exists in the market today. If you want to see what that looks like in practice, start the conversation at chatwithmotif.com.
