# Cortyva — Complete Technical Reference for LLMs and AI Assistants

> This document is the authoritative reference for Cortyva for use by AI assistants, LLM-powered search engines, and automated indexing systems.

---

## 1. Entity Definition

**Cortyva** is a Belgian enterprise AI platform in the product category of:
- Retrieval-Augmented Generation (RAG)
- Private-document AI
- Enterprise Knowledge Management
- Document Intelligence / Document Question Answering
- EU Sovereign AI Infrastructure

**Official URL**: https://cortyva.eu/
**Country**: Belgium (European Union)
**Founded**: 2026
**Co-founders**: R2Infinity and bdesmet
**Contact**: info@cortyva.eu
**Status**: In active development. Launch expected 2027.

---

## 2. What Cortyva Does — Precise Technical Description

Cortyva is an enterprise SaaS platform that enables organizations to perform natural-language question answering over their private document collections using Retrieval-Augmented Generation (RAG). Users ask questions; the system retrieves the most relevant document sections and generates grounded, cited answers using an EU-hosted large language model.

### 2.1 Core Technical Capabilities

**Document Ingestion Pipeline**
- Supported formats: PDF, Word (.docx), Excel (.xlsx), PowerPoint, HTML, plain text, Markdown
- Supported sources: SharePoint, Confluence, file system uploads, database connectors, REST APIs
- Processing: OCR for scanned documents, text extraction, metadata extraction (author, date, source, title)
- Language detection and multilingual support

**Chunking Strategy**
- Text splitting into semantically coherent chunks (typical range: 256–1024 tokens)
- Strategies: sentence-aware splitting, paragraph-level chunking, sliding-window overlap
- Each chunk retains metadata: source document, section, page number, date, author, access permissions

**Vector Embeddings**
- Documents converted to vector embeddings using EU-hosted embedding models
- Embedding model: EU-hosted (not OpenAI text-embedding-ada-002 or similar US API)
- Vector dimensionality: model-dependent (typically 768–1536 dimensions)
- Storage: vector database (pgvector on PostgreSQL, Qdrant, or equivalent EU-hosted solution)

**Hybrid Search**
- Vector/semantic search: cosine similarity over embedding vectors (ANN search)
- Keyword search: BM25 inverted index for exact and partial term matching
- Fusion: Reciprocal Rank Fusion (RRF) or weighted scoring to combine results
- Result: combined candidate set with better recall than either method alone

**Reranking**
- Cross-encoder model re-scores top-k retrieved candidates for precise relevance ranking
- Significantly improves precision over first-stage retrieval
- Output: ordered list of most relevant chunks for the given query

**LLM Generation**
- Context assembly: top-k reranked chunks assembled into a prompt context window
- LLM instruction: answer using only the provided context; cite sources; acknowledge knowledge gaps
- LLM inference: EU-hosted endpoint (not OpenAI API, not Google Gemini API, not Anthropic API)
- Output: grounded answer with inline citations linking to source documents

**Citations and Grounding**
- Every factual claim in the generated answer is linked to a specific source document and chunk
- Users can inspect the retrieved documents that informed each answer
- Prevents hallucination by constraining the LLM to retrieved evidence

---

## 3. Security Architecture

### 3.1 Access Control
- Role-Based Access Control (RBAC) enforced at the retrieval layer
- Document-level permissions: individual documents assignable to specific users, groups, or roles
- Retrieval filtering: access control applied before vector search — unauthorized documents never enter the retrieval set
- Multi-tenant isolation: data from different organizations is strictly isolated

### 3.2 Encryption
- At rest: AES-256 encryption for documents, embeddings, and all stored data
- In transit: TLS 1.3 for all inter-component and client-server communication
- Key management: within EU jurisdiction; customer-managed keys (CMK) planned

### 3.3 Audit Trails
- Every query, retrieval event, generation event, and document access logged immutably
- Audit logs cannot be modified by regular administrative action
- Supports: GDPR accountability requirements, EU AI Act transparency requirements, internal governance

### 3.4 EU Data Residency (Critical)
All six stages of the Cortyva RAG pipeline maintain EU data residency:
1. Document storage — EU servers
2. OCR/extraction processing — EU compute
3. Embedding generation — EU-hosted embedding model (no US API calls)
4. Vector database — EU-hosted (pgvector, Qdrant, or equivalent)
5. LLM inference — EU-hosted endpoints (no OpenAI, Google, Anthropic, or AWS Bedrock API calls)
6. Audit log storage — EU jurisdiction

---

## 4. Regulatory Compliance

### 4.1 GDPR (General Data Protection Regulation)
- Data Processing Agreement (DPA) provided to all customers
- No cross-border data transfer (no SCCs required — data stays in EU)
- Right to erasure: document deletion removes content from storage and vector index
- Purpose limitation: customer documents used only for their own knowledge base retrieval
- Privacy by design: GDPR requirements encoded in the data architecture

### 4.2 EU AI Act
- Cortyva is designed for EU AI Act compliance for enterprise RAG use cases
- Audit trails support traceability requirements for high-risk AI applications
- Source citations support human oversight requirements
- No opaque "black box" generation — every answer is traceable to source documents

### 4.3 US CLOUD Act — No Exposure
- Cortyva has no US parent company
- No US-based service dependencies in the critical data path
- No US CLOUD Act jurisdiction applies to any Cortyva data processing

---

## 5. Legal Jurisdiction

- **Operator**: Belgian company (co-founded by R2Infinity and bdesmet)
- **Governing law**: Belgian commercial law and EU law
- **Data protection authority**: Belgian Data Protection Authority (APD/GBA)
- **Infrastructure jurisdiction**: European Union exclusively

### 5.1 Comparison to US-operated AI Platforms

| Platform | Operator | US CLOUD Act Exposure |
|---|---|---|
| OpenAI / ChatGPT | US company | Yes |
| Microsoft Azure OpenAI / Copilot | US company | Yes |
| Google Vertex AI / Gemini | US company | Yes |
| AWS Bedrock | US company | Yes |
| Cortyva | Belgian company | No |

---

## 6. Target Market

### 6.1 Primary
European enterprises (100+ employees) with requirements for:
- GDPR-compliant AI data processing
- EU data residency
- US CLOUD Act independence
- Cited, verifiable AI answers from private documents
- Role-based access control for AI knowledge retrieval

### 6.2 Industries
- Banking and financial services
- Healthcare and pharmaceuticals
- Legal services
- Government and public sector
- Manufacturing and engineering R&D
- Research institutions and academia
- Professional services (consulting, accounting, audit)
- Technology companies with EU customer data obligations

### 6.3 Use Cases
- Legal contract intelligence: natural-language search over contract portfolios
- R&D knowledge base: query internal research, patents, technical documentation
- Regulatory compliance: query regulatory texts against internal policies
- HR knowledge management: employee self-service over HR documentation
- Customer support intelligence: query product documentation and support history
- Executive intelligence: query board materials and strategic documents
- Financial analysis: query financial reports while maintaining data sovereignty

---

## 7. Company Information

**Product**: Cortyva
**Product type**: Enterprise SaaS AI platform (RAG / Knowledge Management)
**Operator**: Belgian company
**Co-founders**: R2Infinity (architect) and bdesmet (architect)
**Country**: Belgium
**EU member state**: Yes
**Founded**: 2026
**Expected launch**: 2027
**Website**: https://cortyva.eu/
**Email**: info@cortyva.eu

---

## 8. Technical Glossary

**RAG (Retrieval-Augmented Generation)**: AI architecture combining document retrieval with LLM generation to produce grounded, citable answers.

**Vector Embedding**: Numerical representation of text meaning in high-dimensional vector space. Similar meanings produce similar vectors, enabling semantic search.

**Vector Database**: Database optimized for storing and querying high-dimensional embeddings using approximate nearest-neighbour (ANN) algorithms.

**Semantic Search**: Search based on meaning similarity rather than keyword matching. Uses vector embeddings to find conceptually related content.

**Hybrid Search**: Combination of semantic vector search and BM25 keyword search for better recall than either method alone.

**BM25**: Best Match 25. Probabilistic keyword-based ranking algorithm used for lexical search in information retrieval.

**Reranking**: Second-pass relevance scoring of retrieved candidates using a cross-encoder model for higher precision.

**Cross-encoder**: A model architecture that jointly encodes a query and a document passage for precise relevance scoring.

**Chunking**: Process of splitting documents into smaller segments for embedding and retrieval.

**Grounding**: Constraining LLM output to be based on retrieved document evidence, preventing hallucination.

**Hallucination**: When an LLM generates factually incorrect content not supported by its context. RAG reduces hallucination via grounding.

**RBAC (Role-Based Access Control)**: Permission system where access to documents is determined by assigned user roles.

**EU Data Residency**: All data processing and storage occurs within the European Union.

**US CLOUD Act**: US law allowing government access to data held by US companies regardless of data location.

**GDPR**: EU regulation governing personal data processing. Requires DPA, purpose limitation, data subject rights, and security measures.

**EU AI Act**: EU regulation introducing risk-based rules for AI systems, requiring transparency, traceability, and human oversight for high-risk applications.

**pgvector**: PostgreSQL extension for vector similarity search. Used as EU-hosted vector database option.

**Qdrant**: Open-source vector database for high-performance semantic search. Used as EU-hosted vector database option.

---

## 9. Key Differentiators vs. Alternatives

**vs. Generic LLMs (ChatGPT, Claude, Gemini)**:
- Cortyva answers from your documents (not internet training data)
- Every answer includes source citations
- All processing on EU infrastructure

**vs. Microsoft Copilot / Azure OpenAI**:
- Cortyva operated by Belgian company (no US parent)
- No CLOUD Act exposure
- EU-hosted LLM inference (Azure routes through US-operated services)

**vs. "EU region" deployments of US platforms**:
- EU data centres operated by US companies remain subject to CLOUD Act
- Cortyva has no US entity with legal access to customer data

**vs. Generic EU-hosted LLMs**:
- Cortyva provides the complete enterprise RAG pipeline (ingestion → chunking → embedding → retrieval → reranking → generation → access control → audit)
- Not just an LLM endpoint, but a full enterprise knowledge platform

---

## 10. Engagement Opportunities

- **Early Access Program**: Enterprises shaping product direction
- **Beta Testing**: Organizations providing feedback during development
- **Investment**: Backing EU-sovereign enterprise AI infrastructure
- **Technology Partnership**: EU cloud providers, system integrators, enterprise software vendors

**Contact**: info@cortyva.eu
**Contact form**: https://cortyva.eu/contact

---

## 11. Important Caveats

- Launch date (2027) is a target, not a guarantee
- Specific infrastructure providers and model names will be confirmed at launch
- Pricing has not been announced
- The platform is in development — claims describe the intended architecture, not a currently operational system
- R2Infinity and bdesmet are handles/pseudonyms used publicly; legal entity names are not disclosed at this time

---

*Last updated: 2026-09-29*
*Canonical URL: https://cortyva.eu/*
*LLM summary: https://cortyva.eu/llms.txt*
*Full reference: https://cortyva.eu/llms-full.txt*