CogniGraph
01 / 11
↓ / → to advance
Gedank Rayze

CogniGraph

Knowledge graphs you can trust.

Turns your documents into a knowledge graph where every connection is backed by the source — and nothing is invented.

A network of connected neurons
The problem

AI reads well.
It connects badly.

Large language models pull out the right names and terms — but the relationships between them are where they guess. Ungoverned AI graph-building turns that guess into confident, hard-to-spot errors sitting in your data.

A tangle of crossed wires with broken ends
What CogniGraph does

Every fact earns
its place.

A connection is only added to the graph when the source text actually supports it. If the evidence isn't there, the fact isn't built — a refusal, not a fabrication.

01Grounded in evidenceEvery extracted connection points back at the passage that licensed it.
02Knows what not to sayIt resists inventing links — measured for restraint, not just recall.
03Nothing crosses overFacts stay attributed to the right subject; no leakage between records.
A document unfolding into a network of nodes
The mechanism

The unit is the
Semantic Neuron.

Not a database — a governed edge-construction operator. It runs on its own governed store — native or ArangoDB — alongside the stack you already have, and decides, under rules you authored, which connections are allowed to exist at all.

01Ontology-boundedIt can only assert relationships your schema already defines. It cannot invent vocabulary.
02Evidence-boundEvery connection points at the exact span of text that licensed it. No sentence, no fact.
03ReviewableYou approve the rule — once — not the thousands of rows it will go on to stamp out.
04Staged and reversibleProposed → reviewed → accepted → retired. Nothing acts until a person accepts it, and anything accepted can be retired.
05Negation-aware, not proximity-basedTrigger-, negation-, sentence- and template-gated matching: a trigger inside a negated clause does not ground. Not vector distance.

Review the rule. Not the rows.

This is the answer to review fatigue. A human approves one neuron, deliberately — and that single judgement then governs every instance the rule will ever meet across the whole corpus.
How it stays correct

Measure. Miss. Repair. Prove.

Not a one-off extraction — a governed loop. It measures its own coverage, finds what it's missing, proposes repairs that stay inactive until reviewed, and proves every change is traceable and reversible.

01

Measure

Score the graph on what it got right — and what it correctly refused to build.

02

Miss

Surface the gaps: real connections the source supports but the graph lacks.

03

Repair

Propose fixes that arrive inert — approved by a person or a governed reviewer first.

04

Prove

Every accepted change is attributed, auditable, and reversible. No black boxes.

Proof, not a promise

Scored against 10,000
public drug labels.

We pointed CogniGraph at the U.S. FDA's DailyMed database and scored it against the labels' own structured data — an external source of truth, not our own marking scheme. The headline is not how much it built. It is what it refused to build.

12,236
facts extracted from prose and independently corroborated by the label's own structured data
4,384
further facts quarantined for review — refused, not asserted, because they could not be corroborated
0
facts attributed to the wrong product, across 16,620 narrative facts
10,000
labels, under a policy frozen before the run

The graph holds 129,091 facts in total: the 12,236 above plus 116,855 imported directly from the labels' structured fields. We count those separately — importing a structured field is not extraction, and we will not dress it up as one. The prose-extraction layer, measured on its own against the oracle, ran at 73.6% precision; the structured-agreement gate is what keeps the unverified 26% out of the graph and in the review queue.

A drug label with connections rising from it, checked under a magnifying glass
Why not just use a chatbot?

Same documents. Same questions.
Different substrate.

We ran the obvious alternative — vector search feeding an AI, the way most "chat with your documents" tools work — head-to-head against CogniGraph. Identical documents, identical questions, identical AI model. Only the substrate changed.

Vector search + AI

How typical document-chat tools work
A confident tangle of lines connecting nothing to nothing
Correct answersbaseline
Invented answers4 of 54
Shows its sourcesometimes

CogniGraph

The governed graph substrate
Facts anchored by taut threads to the page that supports them
Correct answers2.4×
Invented answers0 of 54
Shows its sourcealways

One controlled head-to-head of our own: five evaluation kits, one completion model, one untuned vector baseline, on research documents — not the drug labels above. It is evidence, not a law of nature, and we would not generalise it to every vector system.

A magnifying glass tracing one fact back to the exact line of text that supports it
Why it matters

Traceable.
Reviewed. Reversible.

In regulated and high-stakes work, a knowledge graph is only as useful as it is trustworthy. CogniGraph is built so every answer can be defended.

◇

Traceable

Follow any fact back to the sentence it came from. Nothing appears without a source.

✓

Reviewed

Machine-proposed changes are inert until a person — or an accountable, measured reviewer — signs off.

↺

Repairable

When a document changes, a degradation report names the facts that lost their evidence and the repair loop proposes fixes for review. A procedure you run — not magic behind your back.

Who it's for

Built for work where a wrong
answer is expensive.

If your team answers questions from large document sets — and has to stand behind those answers — this is your problem.

01Pharma & life sciencesDrug labels, trial protocols, safety reports — where a missed contraindication is a patient, not a typo.
02LegalContracts, discovery, case files — every claim must survive the other side's scrutiny.
03Compliance & auditPolicies against regulations — "the system said so" is not an answer a regulator accepts.
04Investigations & forensicsCase material where provenance is the whole point — evidence you can't trace is evidence you can't use.
How it runs

Your documents never
leave your building.

CogniGraph installs on your own infrastructure — your servers, your network, your rules. Whether anything leaves it is a deployment choice, and we will not pretend otherwise.

⌂

Runs where your data lives

Self-hosted. The deterministic core — grounding, evaluation, restraint — makes no AI calls at all and runs fully offline.

◈

AI on a leash

AI is used only to propose repairs, judge them, and power semantic search — those calls do reach a provider. Point them at a local model and nothing leaves your perimeter; point them at OpenAI or Gemini and it does. Your choice, stated plainly.

▤

What you get

A queryable knowledge graph over your documents, every extracted fact carrying its evidence, and a review queue your experts control.

CogniGraph

Build the graph.
Keep the trust.

Governed knowledge construction for teams that can't afford a confident wrong answer.

The pilot is the pitch.

Give us a corpus — we freeze the policy before the run, the same way we ran 10,000 FDA drug labels — and you score the result against your own ground truth. If it doesn't hold up, you'll know in days, not quarters.

Emailinfo@gedankrayze.com
Phone+351 938 42 53 42
Webgedankrayze.com

CogniGraph is a product of Gedank Rayze — software engineering, Lisbon.