FlinkISO AI can help authorised users work with quality documents and generated forms. It can use a centrally managed FlinkISO AI subscription, your organisation’s cloud AI account, or models hosted on an Ollama server under your control. Only an MR or administrator account can open and change AI Setup.

Choose how AI will run

Use a separate AI server

For better application performance, do not run a production Ollama workload on the same server as FlinkISO. AI inference can consume sustained CPU, system memory, disk bandwidth and GPU memory. A separate host keeps document editing, forms, approvals and scheduled work responsive while AI is processing a long request.

The AI API URL must be reachable from the FlinkISO AI service that handles the paid request. http://127.0.0.1:11434 is valid only when Ollama and that service run on the same host. With the recommended separate AI server, use its reachable private address or a secured HTTPS endpoint. Do not expose the unauthenticated Ollama port directly to the public Internet.

Practical starting hardware

Model size, quantisation and context length determine the actual requirement. The following is a practical minimum for the example 7B–8B models in this guide; larger models need more memory.

The two example model downloads currently use about 11.2 GB before runtime overhead and working files. Increase RAM, VRAM and storage when using larger models, longer context windows, more PDF pages or concurrent requests.

Install Ollama on a separate Linux server

The following basic flow follows the official Ollama Linux guide. Run it on the dedicated AI server, not on the FlinkISO application server.

1. Install Ollama.

   curl -fsSL https://ollama.com/install.sh | sh

2. Enable and start the service.

   sudo systemctl enable --now ollama
   sudo systemctl status ollama

3. Download a text model and a vision model. These names are examples that match the fields shown later.

   ollama pull qwen3:8b
   ollama pull qwen2.5vl:7b
   ollama list

4. Verify the local API.

   curl http://127.0.0.1:11434/api/tags

The example text model is approximately 5.2 GB and the vision model approximately 6.0 GB. Model names entered in FlinkISO must exactly match the names returned by ollama list.

Allow the AI service to reach Ollama

Ollama listens on 127.0.0.1:11434 by default. On a separate Linux host, add a systemd override only when network access is required:

sudo systemctl edit ollama.service

Add:

[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"

Then reload and restart:

sudo systemctl daemon-reload
sudo systemctl restart ollama

Configure AI in FlinkISO

1. Sign in with the MR or administrator account.
2. Open the user menu and select AI Setup.
3. Select the required AI Provider.
4. For Ollama or a cloud provider, complete the provider fields described below.
5. Select Turn AI On, then choose Save AI Setup.

Higher page counts, page-pixel values and context sizes increase memory use, request size and processing time. Begin with the defaults and raise them only after testing representative documents.

Connect a cloud OpenAI-compatible provider

1. Confirm that the provider supports OpenAI-compatible streamed chat completions and JSON responses.
2. Create an API key with only the permissions required for model inference.
3. Enter the provider’s API base URL. FlinkISO appends /chat/completions unless that path is already present.
4. Enter exact text and vision model identifiers available to the account.
5. Enter the API key, retain the document-processing defaults, turn AI on and save.

Verify the connection

1. Reopen AI Setup and confirm that AI remains turned on.
2. Open a supported quality document or generated form where the AI control is available.
3. Start with a short text-only request. Confirm that a response is returned before testing a document.
4. Test a small PDF or image with the vision option and confirm that the selected vision model is used.
5. Review the proposed result before allowing a form rebuild or field change.

FlinkISO keeps the local request history and applies its normal authorisation rules. Only an MR user can rebuild a generated form from an AI proposal.

Troubleshooting