Edmund AI Documentation
/DOCUMENTATION

Deployment options

Edmund AI offers flexible deployment options designed to cater to a wide range of organizational needs, from cost-effective shared cloud environments to highly controlled enterprise on-demand installations. These options ensure that Edmund AI can be integrated seamlessly into existing infrastructure, whether on-premise or in the cloud.

Quick summary of available options

Deployment type / MetricData IsolationTotal Cost of OwnershipScalingAccess to new features
Public SaaSMediumLowHighInstant
Private SaaSHighMediumHighSlightly delayed
Private cloudHighMedium–HighHighDelayed
On-premiseHighHighLowVery limited

1) Shared Cloud Deployment (Public SaaS), preferred option

This model allows for the deployment of Edmund AI within a multi-tenant cloud environment. It is an ideal solution for organizations looking for rapid provisioning, reduced infrastructure management overhead, and cost efficiency.

In a shared cloud setup, resources are pooled and shared among multiple users, while still maintaining logical separation and security for each instance of Edmund AI. This option is particularly beneficial for quick starts, development environments, or use cases with fluctuating resource demands.

Advantages:

  • Lowest TCO due to shared infrastructure and automatic scaling (scaling handled by Edmund s.r.o.), autoscaling, capacity optimization.

  • Fast releases (CI/CD, canary, feature-flags), short time-to-market for changes.

  • Higher availability and resilience (easier multi-AZ/DR), better RTO/RPO.

  • Centralized security: access control, patching, scanning, secret rotation, KMS, audit.

  • Out-of-the-box observability: centralized logs/metrics/traces, 24/7 monitoring and incident response.

  • Uniform configuration and standards, fewer “snowflake” installations.

  • Simpler compliance (ISO/SOC), easier audits due to a unified operational model.

  • No operations on the customer’s side (zero-ops).

Disadvantages:

  • Less room for deep customizations (limited deviations from the standard).

  • Shared environment (requires tenancy isolation; we mitigate “noisy neighbor” risks, but they exist).

  • Regional limitations (data residency only in the regions we offer).

  • Integration into private networks may require VPN/PrivateLink and addressing egress costs for exporting logs/data.

  • Change cadence is primarily determined by us; the customer has less control over release/maintenance schedules.

2) Private SaaS — personal dedicated instance in a separate AWS account

Alternatively, Edmund AI can be deployed as a dedicated instance hosted within a separate AWS account. This offers the benefits of cloud scalability and flexibility, combined with the dedicated resources and enhanced security of a private infrastructure, exclusively for a single organization.

Advantages:

  • Hard isolation by account (billing, IAM, limits, quotas), clear separation of costs on the customer’s invoice.

  • Data residency per region as required by the customer; option for Customer-Managed Keys (KMS CMK).

  • Better integration into the customer’s network (VPC peering/PrivateLink, IP allowlists), easier access to raw logs/metrics.

  • Compatibility with their security tools (GuardDuty, Security Hub, IAM Identity Center/SSO, SCP).

  • Fast provisioning thanks to IaC (Terraform), repeatable builds, easier auditability of changes.

  • Change cadence can be aligned with the customer’s change management (maintenance windows, blackout periods).

Disadvantages:

  • Fixed baseline operational costs of a dedicated environment (EKS control plane, NAT, gateway, logging).

  • Capacity is planned rather than auto-scaled; upscaling is by agreement.

  • Higher operational load: more accounts = more patching, secrets, IAM, audits.

  • Risk of IaC drift and configuration deviations between instances → necessity of regular checks/reconciliations.

  • DR/multi-region is more complex and costly per customer.

  • Longer lead time for establishing new regions/accounts, coordination with their domain/IdP/certs.

  • Monitoring and alerting require clear interfaces (who sees what, sharing dashboards, exporting logs).

  • Support is more expensive (SLA per account, access processes, onboarding of customer’s personnel).

3) Private cloud — hosted in the customer’s own cloud account (AWS or Azure)

Edmund AI can also be installed into the customer’s own cloud account or subscription. The platform is delivered as containers and deployed with Terraform onto the provider’s managed services (managed Kubernetes, managed PostgreSQL and Redis), so the customer needs no hardware, while data and infrastructure remain entirely under the customer’s control — in the customer’s account, region and security perimeter. This option suits organizations that already run production workloads in AWS or Azure and want Edmund AI governed by the same landing zone, identity, logging and billing as the rest of their estate.

Advantages:

  • Data sovereignty without hardware: data, infrastructure and encryption keys stay in the customer’s own account and region, under the customer’s IAM, logging and security tooling.

  • Cloud economics on the customer’s own invoice: managed services, elasticity and the customer’s existing enterprise agreements and credits.

  • Same architecture and build as the other options (containers, Kubernetes, Terraform IaC) — repeatable, auditable deployments with no snowflake installation.

  • Integration inside the customer’s network (VPC/VNet peering, private endpoints); data sources are reached without traversing the public internet.

  • Change cadence and maintenance windows aligned with the customer’s change management.

Disadvantages:

  • The customer operates the environment: account, Kubernetes cluster, managed-service patching, backups and DR. Requires DevOps competence with containers and Terraform, or extended support from Edmund agreed separately.

  • Edmund needs a defined access path for deployment and support (cross-account role or guest access), which the customer’s security process must accommodate.

  • Releases are rolled out into the customer’s environment by agreement, not continuously → delayed access to new features.

  • Fixed baseline costs of a dedicated environment (Kubernetes control plane, NAT, gateway, logging) are borne by the customer; capacity planning and upscaling are by agreement.

  • Telemetry and observability need an agreed interface (export to Edmund’s stack or the customer’s own monitoring).

  • External dependencies remain: the language model is operated by an external provider (Azure OpenAI) and the vector database is Zilliz Cloud, so outbound HTTPS to these services is required (see Networking requirements).

4) On-premise (on the customer’s own hardware)

Edmund AI can be deployed directly within an organization’s own data centers. This option provides maximum control over the underlying infrastructure, data residency, and security policies. It is often chosen by enterprises with strict regulatory mandates or those that prefer to keep all data and applications within their own physical control.

Advantages:

  • Data sovereignty / “no-cloud” policy (regulated segments, specific contractual requirements).

  • Full control over the network, access, and integration points; low latency to legacy systems/data.

  • Ability to use existing hardware and on-prem tools (SIEM, backup, monitoring).

Disadvantages:

  • High maintaining costs, long hardware delivery times, capacity planning is on the customer.

  • Heterogeneous environment → demanding support, a wide “support matrix,” risk of a “snowflake” installation.

  • Without managed cloud services, worse elasticity and automatic scaling.

  • Patching, security, backups/DR – most of the responsibility is on the customer, unless extended support from Edmund is agreed separately.

  • Slower releases (typically quarterly) and more complex testing.

  • Telemetry/observability is more difficult (VPN, firewalls, export to our stack).

  • Necessity of a portal/registry for image distribution and license enforcement (blocking after contract termination).

  • SLA is difficult to achieve without significantly expensive support and on-site interventions.

Explanation of Abbreviations

  • DR: Disaster Recovery - An organization’s method of regaining access and functionality to its IT infrastructure after a disaster.

  • IaC: Infrastructure as Code - The process of managing and provisioning computer data centers through machine-readable definition files, rather than physical hardware configuration or interactive configuration tools.

  • IdP: Identity Provider - A system entity that creates, maintains, and manages identity information for principals and provides authentication services to relying applications.

  • KMS: Key Management Service - A managed service that makes it easy for you to create and control the encryption keys used to encrypt your data.

  • KMS CMK: Customer-Managed Keys - Encryption keys in a KMS that are created, owned, and managed by the customer.

  • RPO: Recovery Point Objective - The maximum acceptable amount of data loss after an unplanned incident, expressed as an amount of time.

  • RTO: Recovery Time Objective - The targeted duration of time within which a business process must be restored after a disaster or disruption to avoid unacceptable consequences.

  • SCP: Service Control Policies - A type of organization policy in AWS that you can use to manage permissions in your organization.

  • SIEM: Security Information and Event Management - A solution that helps organizations detect, analyze, and respond to security threats before they harm business operations.

  • SLA: Service Level Agreement - A commitment between a service provider and a client.

  • SOC: System and Organization Controls - A suite of reports produced during an audit of a service organization, intended to provide information about controls relevant to security, availability, processing integrity, confidentiality, or privacy.

  • SOW: Statement of Work - A document that captures and defines all the work-related aspects of a project.

  • TCO: Total Cost of Ownership - A financial estimate intended to help buyers and owners determine the direct and indirect costs of a product or system.

Currently Supported Hosting Providers

For Private cloud installations (in the customer’s own AWS or Azure account), Edmund AI is fully supported on leading cloud platforms:

  • Amazon Web Services (AWS): Leveraging the extensive suite of AWS services, Edmund AI can be deployed on EC2 instances, utilize S3 for storage, integrate with AWS databases, and benefit from AWS’s robust networking and security features.

  • Microsoft Azure: Edmund AI is also compatible with Microsoft Azure, allowing deployments on Azure Virtual Machines, integration with Azure Storage, Azure SQL Database, and other Azure services to build a comprehensive and scalable AI solution.

These diverse deployment options ensure that Edmund AI can meet the unique infrastructure and operational demands of any enterprise.

Compute requirements

This section specifies the minimal infrastructure instance requirements necessary to deploy and operate the Edmund platform on different platform providers. Specific requirements depend on the final setup.

Azure

Min. instance typeNumber of instancesPurpose
D2ps v51Application Data (PostgreSQL)
Cloud-Hosted ServiceN/ADocument Indexing (Milvus Zilliz)
F2ps v51Cache for Redis
D4ps v51Kubernetes – Application layer
D8ps v51Kubernetes – Indexing workers

AWS

Min. instance typeNumber of instancesPurpose
db.t4g.medium1Application Data (PostgreSQL)
Cloud-Hosted ServiceN/ADocument Indexing (Milvus Zilliz)
cache.t4g.medium1Cache for Redis
c6g.xlarge1Kubernetes – Application layer
m7g.2xlarge1Kubernetes – Indexing workers

On premise

CPURAMGPUNumber of instancesComponents
4 cores16 GB1Application layer (Frontend, Backend, Backend Database, Cache)
1 core8 GBRTX 3060 8GB RAM2Indexing workers (Parallel document processing and ingestion)*
1 core4 GB4Processing workers (User requests)
8 cores32 GB1Database (Application DB – PostgreSQL; Indexing DB – Milvus, hosted locally)

* Exact configuration of CPU and RAM is dependent on the number of documents expected to be processed during initial indexing.