Skip to main content
graph.swiss
Technology

GraphRAG — Where Knowledge Graphs Meet AI

Most AI tools answer questions from documents. We answer them from a knowledge graph — millions of structured entities and relationships that give every answer context, provenance, and depth. The result: intelligence that reasons across connections, not just keywords.

What is GraphRAG?

GraphRAG (Graph Retrieval-Augmented Generation) extends classical RAG with structured graph traversal. Instead of retrieving isolated text chunks, our AI navigates a Neo4j knowledge graph — following ownership chains, board interlocks, and event timelines to assemble context-aware answers grounded in real entity relationships.

When you ask “Who controls Bergtal Holding AG?”, classical RAG searches text. GraphRAG walks the actual ownership graph — through holding companies, beneficial owners, and cross-border structures — returning a verifiable answer with every hop linked to its primary source.

The result is reasoning that scales beyond keyword matching: multi-hop questions, network analysis, and historical research that classical RAG cannot answer accurately.

Classical RAG

Retrieves text chunks. Limited to keyword similarity. Cannot follow relationships.

Vector RAG

Semantic search via embeddings. Better than keywords but still flat — no structured reasoning.

GraphRAG (Ours)

Multi-hop traversal across structured entities. Reasons through ownership, board, and event networks.

Five Architectural Pillars

Every layer is engineered for accuracy, provenance, and scale. From OCR ingestion to graph traversal, each component is independently observable and replaceable.

GraphRAG Engine

Retrieval-Augmented Generation with graph traversal. Multi-hop reasoning across entity relationships, ownership chains, and temporal events.

Neo4j Knowledge Graph

Millions of nodes (companies, people, addresses, events) with hundreds of millions of typed relationships. Native graph database for sub-second traversal.

Proprietary OCR & NER

Custom-trained OCR processes 140+ years of multilingual gazette archives. LLM-based NER extracts entities across German, French, Italian, English.

Real-Time Ingestion

Continuous pipeline from official registries with sub-24-hour latency. Every new SHAB publication, formation, or mutation indexed automatically.

Multi-Source Enrichment

Composable enrichment layers — Creditreform credit data, internet search, Intelligence X OSINT. Plugged in per use case.

Specialised AI Agents

Industry-specific agents trained on workflow data. KYC, due diligence, market intelligence, equity research — each with its own retrieval strategy and prompt architecture.

Anatomy of a GraphRAG Query

When you ask Swiss Graph a question, here's what happens behind the scenes — every step traceable, every result auditable.

01

Intent Parsing

LLM classifies the query, extracts entities, and selects the appropriate AI agent.

02

Graph Retrieval

Cypher queries traverse Neo4j to find relevant entities and their N-hop neighbours.

03

Context Assembly

Structured graph data + unstructured gazette text + enrichment layers merged into prompt.

04

AI Reasoning

Specialised agent reasons over assembled context with chain-of-thought.

05

Provenance

Every claim linked to source nodes — original SHAB images, registry entries, or external feeds.

The Knowledge Graph in Action

A small slice of the Swiss Graph — companies, people, addresses, and the relationships between them. Drag, expand, and explore. The full graph contains millions of entities.

Data at Scale

Scale without sacrificing depth. Every entity in the graph carries its full historical and relational context.

750K+

Swiss Companies

Active + dissolved since 1883

8M+

Decision Makers

Board members, directors, owners

156K+

Global Securities

96 exchanges, real-time

140+

Years of History

SHAB archives back to 1883

50B+

Data Points

Updated continuously

<24h

Update Latency

New SHAB filings to graph

4

Languages

DE, FR, IT, EN multilingual NER

99.9%

OCR Accuracy

Proprietary engine on gazettes

Why GraphRAG Wins

Multi-Hop Reasoning

Answer questions that require following 2, 3, or more relationships. UBO chains, board interlocks, beneficial ownership across jurisdictions — all natively traversable.

Verifiable Provenance

Every fact links back to its primary source — original gazette page, registry extract, or data feed. No hallucinated facts. No untraceable claims.

Cross-Border Intelligence

Swiss Graph connects to World Graph — so ownership chains, financial flows, and corporate networks span every supported jurisdiction.

Specialised Agents

Industry-specific AI agents with their own retrieval strategies, prompt architectures, and validation rules. Banking ≠ journalism ≠ M&A.

Multilingual Native

NER, OCR, and reasoning across German, French, Italian, and English. No translation layer — agents reason directly in source languages.

Historical Depth

Track corporate evolution across 140+ years. Name changes, restructurings, ownership transfers — the full lifecycle of every Swiss entity.

Security & Deployment

Enterprise-grade infrastructure. Swiss data hosting, on-premise deployment options, and zero-data retention guarantees.

Zero-Data Retention

Your queries and uploaded documents are never stored or used for AI training. Processed in-memory and deleted.

SOC 2 Type II

Audited security controls across all infrastructure, with annual third-party penetration testing.

Swiss Data Hosting

All processing and storage in Swiss data centers under Swiss federal data protection law.

On-Premise Available

Full platform deployable inside your VPC or air-gapped environment for maximum data sovereignty.

API & SSO

RESTful API access with rate limits per plan. SAML/OIDC SSO integration. Webhook event delivery for real-time alerts.

Custom AI Models

Fine-tuned industry-specific LLMs running in isolated environments per enterprise client. No data leakage between tenants.

Ready to See GraphRAG in Action?

Try the platform on real Swiss company data, or talk to our team about custom deployments and data feeds.