Skip to content

Hardware in your server room

Knowledge base on your own hardware

Search contracts, technical documentation, and correspondence via chat when company policy prohibits sending them to the cloud. The model and index run in your rack cabinet. A single machine handles 5 to 20 concurrent simple queries within a 100 to 1,000 ms window. An additional server doubles that capacity.

In practice

Data never leaves the building

Legal, HR, and R&D hold documents that company policy forbids sending to public APIs. The solution is not abandoning semantic search, but moving it to hardware in your own server room.

We deliver a preconfigured machine with an open-source model, document index, and admin panel. You indicate the folders; we configure permissions and train your IT specialist to refresh the index independently.

The databases we build on usually contain 10–500 knowledge files, each with graphics. We match the vectorisation method to your actual operational processes.

Queries exceeding the current pool are not lost. They enter a queue of up to 1,000 items, with a timeout limit of 1 to 10 seconds set during configuration. Every response cites its source document, allowing employees to verify it directly.

You remain the data controller under GDPR. Our access is enabled only during servicing and upon your explicit request.

  • How it works

    The language model and the vector index of your files reside on a single server. We select the GPU based on dataset size: 128 GB of VRAM accommodates 70B-class quantized models. Queries, documents, and responses never leave the corporate network, and the internet connection can be unplugged.

  • Integrations

    The device connects to your existing network and reads SMB shares, email, and your current document repository. Authentication runs through Active Directory, eliminating the need to manage a separate set of accounts and passwords.

  • Who decides

    The hardware, drives, and model weights belong to you. After our collaboration ends, nothing remains to recover from us. Logs show who asked what and which document was referenced, which supports GDPR compliance audits.

Problem

Do regulations prohibit uploading these files to the cloud?

HEXART solution

HEXART on-premise RAG keeps the index, model, and documents on a single device in your server room. Queries and responses stay within the corporate network, and access logs belong to your administrator, not a vendor.

Book an infrastructure audit

Questions and answers

Frequently asked questions

How many queries can one machine handle?

Between 5 and 20 simple vector queries concurrently within a 100–1,000 ms window. Each additional server doubles this capacity. Overflow queries enter a queue of up to 1,000 items, with a wait limit of 1 to 10 seconds set during configuration.

Do documents leave the corporate network?

No. The language model and vector index reside on a single server in your server room. Queries, document contents, and responses never leave your corporate network. The internet connection can be unplugged during operation.

Is it necessary to create separate accounts and passwords?

No. Authentication runs through Active Directory, and the device reads SMB shares, email, and your existing document repository.

Who is the data controller under GDPR?

You are. Our access is enabled only during servicing and upon your explicit request. Logs show who asked what and which document was referenced, which supports GDPR compliance audits.

No-obligation call

Schedule a workshop with Paul

Thirty minutes or a full workshop, choose what suits you. Book a slot directly in the calendar and get instant confirmation.

Paul Lazniak

HEXART Founder

Contact

Let us talk about your project

We will show you a working product before we start talking about it. Let us know what you need: a film, an XR environment, an AI system, or brand identity.

Write or call

A few sentences are enough: what needs to be created, by when and for whom.

Address
Aleja Zwycięstwa 96/98, 81-451 Gdynia

Deployments catalog

View full catalogue
  • Company knowledge base for your team

    A new hire asks about a procedure and receives an answer from a specific document with a link to the source. Managers stop serving as a search engine for the department, and documented knowledge stays in the company when employees leave.

  • Local AI server (Blackwell, 128 GB)

    Sending company documents to an external API can be unacceptable legally or simply out of caution. We deploy an on-premises machine with 128 GB GPU memory running open-source language models with no cloud connection.

  • Encrypted internal messenger with AI

    Acquisition plans, salary negotiations, and HR matters circulate today on messengers whose terms of service no one has read. We deploy a dedicated channel: a server on your premises, end-to-end encryption, and an AI assistant on the same network.

  • Video campaign automation

    One shoot, followed by weeks of re-editing for every market, language, and format. We build a rendering pipeline where new variants are generated in hours, ready for marketing review.

  • AI agents for chat and phone

    A large share of support tickets are repeated questions about order status, invoices, and delivery times. An AI agent answers them directly from your documentation, resolving 30–60% of such inquiries in our deployments. It routes the rest to a human agent along with full conversation context.