Local AI for Data Sovereignty
The BECONEX AI Suite runs in your own network.
A Windows server hosts the Tungsten AutoStore platform and processing components. The language models and their runtime environment are hosted on a GPU server as local AI to support data sovereignty requirements.
Processed documents and the extracted data move between these servers inside your infrastructure, no external AI services needed.
Deploying the BECONEX AI Suite
Windows server
Tungsten (Kofax) AutoStore
- Optical Character Recognition 1
- BECONEX AI Suite 2
- BECONEX WebView 3
- Runs the workflow
- Sends the model requests
- Stores results
- Hands documents over to target systems
- Tungsten AutoStore (see requirements)
- IIS with URL Rewrite and Application Request Routing, and Python 3.11 (64-bit, machine-wide) 3
- Certificates for secure communication between web services (optional)
GPU server
Ollama
- Language models
- Vision models
- Answers the model requests of the components
- Ollama (see requirements), version 0.31.2 or later
- GPU with enough memory for the chosen model
- Network access from the AutoStore server
Data Exchange between Components and Language Models
Connection
The AI Suite components send their request over HTTP to the model server address you configure. Timeouts and retries are configurable.
The BECONEX Converter (OCR ahead of AutoStore) uses libraries for communication with the engine.
Request
The request contains a task-specific instruction in the form of a prompt, the page images or OCR text needed for the task.
The prompt can contain additional information like the schema that fixes the format of the reply, the list of document classes, etc.
Reply
Structured data only. The component derives a confidence value from the model’s output; uncertain results go to human review.
- Split: pages where a new document starts.
- Classify: the document class, chosen from predefined values.
- Extract: the configured fields, empty where a value is not found.
Where results go
Back into the workflow on the Tungsten AutoStore server. The models on the GPU server answer requests; they do not route or archive documents themselves.
How the Local AI Setup supports Data Sovereignty
Your servers hold the data
Documents, extracted data, intermediate results, and review decisions are stored on servers you operate; they stay within your infrastructure.
Review data is not publicly reachable
The review interface’s data folder sits outside the web server’s public directories, and its backend accepts connections from the local server only, through the web server.
Signed-in reviewers
Access to the review interface requires sign-in with your Microsoft Entra ID and a reviewer role; the sign-in exchanges identity data, not documents.
Your enterprise demands take the lead
Whether a deployment meets your requirements depends on your overall system, network policies, and documented procedures. The solution adjusts itself to your organization, not the other way around.
Frequently asked Questions
In the standard deployment, text recognition runs locally in your network and the language models run on your GPU server. No external AI service is called.
Ollama, running on a GPU server in your network. The components address it by a configurable server address.
Yes. The AI Suite connects to the GPU server whose address you configure. We define the sizing with you, based on the model and your document volume.