Governed knowledge
Workspace-scoped chat, permission-aware retrieval, citations, abstention, document quarantine, lifecycle records, and local embeddings.
Products
In Development / Private PreviewActive Development — Testing in Progress
SovAI Governed Knowledge Agents
Governed enterprise knowledge, local AI, and evidence-aware agent workflows—with control over identity, permissions, data boundaries, models, evidence, and human approvals.
Development Preview: This product is under active development and testing. It is not currently available for production deployment or regulatory certification.
Why this product
Make governance part of the knowledge workflow.
Ordinary chatbots can leave identity, authorization, provenance, and evidence as afterthoughts. This product is being designed around the questions enterprise teams must be able to answer.
Who is allowed to retrieve each document or evidence item?
Which workspace, model, and tool was authorized for an answer?
Can access be revoked without leaving stale content in sessions, memory, or indexes?
Can answers show the supporting document, citation, and permission context?
Can human reviewers govern risks, evidence, and approvals without delegating legal decisions to software?
Can an investigation reconstruct events from versioned records, audit events, and signed receipts?
Architecture
A customer-controlled governance and runtime boundary.
The current development profile uses an authenticated web workspace and control API over isolated knowledge, model, agent, evaluation, and audit services. Internal workers and data services are not directly exposed to end users.
Users and administrators
│
▼
SovAI web workspace ────── Keycloak identity
│
▼
Authenticated control and governance API
│
├── Workspace and permission services
├── Knowledge ingestion and connector workers
├── Retrieval and quarantined-reader boundary
├── Model-policy gateway and local Ollama
├── Agent, MCP, capability, and approval gateways
├── Evaluation and evidence services
└── Audit, governance, and assurance records
│
▼
PostgreSQL/pgvector + governed object storage + local model cachesCapabilities under development
Knowledge, agents, evidence, and human governance in one cumulative product.
These capabilities describe the current engineering scope. They are not claims of production acceptance, legal compliance, or certification.
Identity and access
Keycloak-backed authentication, groups, roles, workspace bindings, model entitlements, and backend-enforced authorization.
Governed agents and tools
Versioned agents, tool-policy checks, sensitive-action approvals, MCP governance boundaries, revocation checks, and execution receipts.
Knowledge and evaluation
Governed ingestion, connector processing records, evaluation datasets and runs, lineage surfaces, corrective actions, and recalls.
Evidence and audit
Evidence-oriented results, audit events, signed records, offline package verification, and separate evidence states.
Europe Governance beta
Human-governed system inventory, preliminary classification support, risk and oversight drafts, monitoring, and dossier gap exports.
Development history
Increments 0.1–0.7 belong to the same product.
The release numbers below show cumulative development progress. They are not separate editions or separately purchasable products.
| Increment | Cumulative focus | Current interpretation |
|---|---|---|
| 0.1 | Architecture, trust invariants, and contracts | Foundation |
| 0.2 | Sovereign knowledge, local RAG, citations, and administration | Integrated |
| 0.3 | Enterprise identity, connectors, and revocation | Integrated |
| 0.4 | Governed agents, MCP, tools, and memory | Integrated |
| 0.5 | Advanced retrieval, evidence, and controlled evolution | Integrated |
| 0.6 | Operational resilience, offline trust, and supply-chain controls | Acceptance pending |
| 0.7 | Europe governance and assurance beta | Legal and production acceptance pending |
Validation boundary
Engineering tests are not production acceptance.
Passing repository and controlled-environment checks means the implemented behavior passed those stated checks. It does not imply customer acceptance, independent certification, or a regulator’s assessment.
Tested in controlled environments
Cumulative database migrations through increment 0.7, including downgrade and re-upgrade exercises with synthetic data
Automated unit, integration, security, and end-to-end scenarios using synthetic records
Tenant/workspace authorization and cross-tenant non-disclosure paths
Document, retrieval, citation, model-policy, agent, orchestration, and governance control paths
Classification self-approval denial, evidence separation, and missing-evidence handling
Dossier tamper detection and signed offline regulatory-package validation
Local Docker builds, Compose rendering, web checks, and Helm template validation
Not yet validated for production claims
Customer production deployment and operational acceptance
Independent penetration testing and customer security accreditation
Customer-scale performance, capacity, and soak testing
Production high availability, failover, and disaster-recovery acceptance
Customer-operated air-gap transfer and update procedures
Real qualified trust-service-provider interoperability
External legal review of regulatory mappings, classifications, and generated drafts
Target deployment options
Evaluation paths, not production-readiness claims.
Every deployment option remains subject to its stated validation and customer acceptance boundary.
| Deployment option | Current status |
|---|---|
| Single-host Docker Compose | Development and private-preview evaluation profile |
| Customer-local or on-premises Ollama | Implemented evaluation path; customer sizing and acceptance required |
| Private-cloud local-model endpoint | Governed registration path; provider-specific validation required |
| Kubernetes | Helm engineering artifacts available; production cluster testing pending |
| OpenShift | Template and render work available; production cluster testing pending |
| Air-gapped environment | Offline patterns available; customer operating procedure pending |
| High-availability production | Planned hardening and customer-specific design work |
Directional roadmap
From preview testing toward defined acceptance gates.
Roadmap content may change based on testing, legal review, security findings, and design-partner feedback.
Current — 0.7 preview
Development preview
Continue product testing, usability correction, and design-partner discovery on the cumulative baseline.
0.8
Design-partner hardening
Benchmark-driven capacity decisions, operability, recovery work, and fixes from controlled preview use.
0.9
Europe release candidate
Feature freeze, fuller technical-documentation support, evidence packs, and broader release-candidate assurance.
1.0 — planned
General-availability decision
A Europe GA decision only after defined acceptance gates pass; no certification claim before it is achieved.
Known limitations
What the preview must not be used to claim.
The current platform supports technical and evidence workflows; it does not replace authorized legal, governance, security, or production-acceptance decisions.
Regulatory mappings and metadata require authorized customer legal and governance review.
AI-system classifications remain preliminary and human-governed; the software does not make final legal classifications.
Evidence coverage is not a compliance score or compliance determination.
The development timestamp authority is simulated and is never represented as qualified.
German and French governance-interface text is beta and has not received qualified legal-language review.
Kubernetes, OpenShift, air-gap, high-availability, recovery, and load-testing claims remain deployment-dependent.
The platform does not submit information to authorities or create a declaration of conformity.
FAQ
Private-preview questions.
Is it available for production use?
No. It is under active development and testing and is offered only for controlled private-preview or design-partner evaluation.
Is it EU AI Act compliant or certified?
No compliance or certification claim is made. The platform supports evidence and governance workflows, while legal applicability, classification, conformity, and compliance decisions remain with authorized human reviewers.
Are increments 0.1 through 0.7 separate products?
No. They are cumulative development increments of one product: SovAI Governed Knowledge Agents.
Can I download it from the website?
Not currently. Potential evaluators can request a controlled private preview; there is no public download.
Does it require a public cloud AI provider?
The current evaluation path supports local Ollama and is designed to avoid silent public-model fallback. Other approved endpoints require explicit registration, policy, and provider-specific validation.
Controlled evaluation
Request a private preview or discuss a design partnership.
Tell us about the problem you are evaluating, your knowledge sources, data boundary, and deployment constraints. Do not submit passwords, private keys, personal data, confidential documents, regulated evidence, or production credentials.
Development Preview: SovAI Governed Knowledge Agents is under active development and testing. It is not currently available for production deployment or regulatory certification.
Preview requests are reviewed before access is confirmed. Submission does not guarantee preview access, production availability, or regulatory suitability. See the privacy notice before submitting.