Private knowledge and AI hub for your business

Company documents, RAG and your own models — on-premise, with an air-gap option.

We design and deploy an environment where employees use company knowledge and AI models on infrastructure controlled by your organisation. Local document processing, RAG, access controls and model operations form one service. When the task requires it, we add LoRA fine-tuning. An air-gapped option is available for isolated environments.

Discuss the scope

What you receive

An operational environment within the agreed scope, architecture and data-flow documentation, access configuration, model versions and licences, quality evaluation results, and update and recovery procedures.

When is this service useful?

For businesses that want to search procedures, technical documentation, contracts or internal material while controlling access and data flows. We scope the work around specific users, documents and tasks.

Service scope

Company knowledge and RAG

Document preparation, OCR within the agreed scope, embeddings, indexing and retrieval that supplies context to the model. Answers with source references, retrieval evaluation and permission tests.

Models on your infrastructure

Model selection, CPU/GPU configuration, a local API and an employee interface. Hardware sizing follows a pilot covering context length, concurrent conversations and target response time.

LoRA and task adaptation

Optional LoRA fine-tuning and, where the hardware and model support it, QLoRA. Data preparation, separate training and test sets, and comparisons with the base model, prompting and RAG.

Access and integrations

Roles and source access aligned with user permissions. SSO, AD or LDAP integration is scoped after checking your systems. We also test attempts to access restricted documents.

Isolation and air-gap option

Data-flow mapping, local dependencies and outbound connection controls. For air-gap deployments, we agree offline imports of documents, models and updates, artifact verification, local logs and backups.

Operations and handover

Monitoring, configuration versioning, update and recovery procedures, and administrator documentation. Maintenance and support are agreed separately.

Deployment options

On-premise knowledge hub

A local model and RAG as the starting point for working with company documents.

  • Source preparation and indexing
  • Source references and access controls
  • Quality and performance pilot

Knowledge hub with LoRA

Task-specific model adaptation as an extension when evaluation justifies training.

  • Dataset and separate test set
  • Adapter training and baseline comparison
  • Versioning and a model rollback plan

Isolated air-gap option

Deployment without an internet connection in the operational environment, using controlled offline transfers.

  • Local OCR, embeddings, retrieval and model
  • Offline import and update procedures
  • Isolation, permission and recovery tests

RAG and LoRA serve different purposes

RAG supplies current document excerpts when the model answers. LoRA adapts behaviour to a task, such as an output format or classification. A changing document collection usually does not need to be stored in model weights. We choose training after comparing results, rather than making it a required stage.

What does air-gap mean in this deployment?

Running a model locally does not establish an air-gap. We review the entire agreed chain: document imports, OCR, embeddings, any reranker, the model, authentication, logs and backups. External APIs are replaced with local components. Acceptance includes tests for external connections and a documented offline data-transfer procedure.

Infrastructure scope

The service covers the IT environment for company knowledge and AI. Server selection or procurement, networking, migrations and physical server-room requirements are scoped after discovery. A new physical data centre is not assumed to be necessary.

How we work

  1. Data and requirements discovery

    Agree tasks, sources, permissions, isolation requirements and acceptance criteria. Check model licences and the right to use the data.

  2. Pilot and approach comparison

    Use representative questions to measure answer quality, citations, abstention when sources are missing and response time. Compare prompting, RAG and, if needed, LoRA.

  3. Deployment and acceptance tests

    Deploy the agreed option. Test document access, load, data flows and backup restoration. For air-gap deployments, also verify isolation and offline updates.

  4. Handover and maintenance

    Deliver configuration, model and licence inventory, evaluation results, and administrator and user instructions. Agree incident handling and further development.

Agree scope and price before work starts

We start by agreeing the scope of discovery or a pilot. The quote separates implementation, hardware, licences, any training and maintenance. An air-gap deployment also accounts for local dependencies and offline updates; there is no single price for every environment.

A private deployment does not guarantee error-free answers or automatic GDPR compliance. Quality criteria, data protection and user responsibilities are defined for the specific use case.

FAQ

Do company documents have to leave our server?

We can design local processing of documents and conversations. We also review OCR, embeddings, logs, backups and other dependencies. Permitted external connections are defined before deployment; the operational environment has no internet connection in the air-gap option.

Does RAG require model training?

No. RAG retrieves document excerpts and supplies them to the model at answer time. LoRA is an optional way to adapt model behaviour to a specific task when evaluation confirms a benefit.

How does on-premise differ from air-gap?

On-premise means deployment on company infrastructure, which may have internet access. Air-gap requires an isolated operational environment and controlled offline transfers. We verify the agreed scope through tests that also cover system dependencies.

What hardware do we need?

Sizing depends on the model, context length, document collection and concurrent users. The pilot measures memory use and latency before we recommend a CPU/GPU configuration.

Can you connect company SSO or Active Directory?

We scope integration after checking existing systems, protocols and permissions. This may include SSO, AD or LDAP. We test whether users can obtain content from documents outside their access rights.

Who controls the data and trained models?

Data location, administrative access and delivered artifacts are agreed in the contract. Rights to use base models and adapters depend on their licences and agreements covering training data. We document these conditions before training.

Related services

Discuss the scope

Contact

Let’s define the first step for your project.

Briefly describe your goal or problem. You do not need a complete specification. This is a no-obligation enquiry.

Describe your project