Benchmark questions
Together with your staff we gather between fifty and one hundred genuine questions and link each to the file containing its answer. This benchmark sets the scope and is rerun after every change to check quality.
A plastics moulder near Velenje has fifteen years of paperwork: customer specifications, work instructions, complaint reports, tool drawings and audit minutes. When a young process engineer asks how sink marks were solved on a similar part three years ago, the only person who knows is on holiday this week. A search of the shared drive returns two hundred hits, half of them out of date. Pasting those files into a public chatbot is not an option, since they contain confidential data belonging to automotive customers. What works is an AI environment the company owns outright: the model lives in the firm's own cloud tenancy, the files never leave it, and every reply points to the exact document and page behind it. We put such environments together with European hosts like Hetzner, OVHcloud and IONOS, or inside EU regions of Microsoft Azure, AWS or Google Cloud. Compute is hired by the hour rather than purchased, so there is no hardware to buy. Set-up happens entirely remotely, and the provider contract is in your name.
Design starts from the questions people will ask and works back. Servers are chosen only after we know who asks what and which files hold the answers.
Together with your staff we gather between fifty and one hundred genuine questions and link each to the file containing its answer. This benchmark sets the scope and is rerun after every change to check quality.
European hosts and hyperscaler EU regions are weighed on cost, how easy it is to get GPUs, ready-made model services and data handling terms. Options range from a fully managed service through an open model on hired GPUs to a combination.
OCR for scanned PDFs, turning tables into readable text, cutting long files into meaningful sections and labelling them by customer, product and date. If an instruction exists in several revisions, the current one is marked and earlier ones are dropped or plainly labelled as superseded.
Someone barred from a SharePoint or file share folder receives nothing from that folder through the AI either. Users log in with Entra ID or the identity service you already rely on.
Ask questions from a chat pane in Teams or an intranet page. Each reply lists file name, page and date beneath it, with a link that opens the source.
Private endpoints only, encrypted storage, keys kept in a vault inside your tenancy, and an audit trail of queries with a fixed retention period.
Budget ceilings, warnings, automatic power-down after hours and a monthly consumption summary. We also write up the data flow technically for your processing records and, if needed, an impact assessment.
A first working version over a subset of your files normally takes a few weeks. Further growth follows document volumes and what users tell us.
Benchmark questions, a list of document sources and a couple of architecture options, each with its expected cost.
Files belonging to a single team or product line are indexed, then scored against the benchmark.
For a fortnight several staff rely on it during normal work. Poor or partial replies are logged, and document preparation is tuned accordingly.
We keep maintaining it, or your IT team takes over with infrastructure-as-code and complete documentation.
AI sees everything a user can see, including what they should not. Perhaps a salaries or HR folder was opened to the whole company years ago and nobody stumbled on it. An AI search will surface it within a second and quote from it. Before go-live we therefore audit access on every folder destined for the index and narrow any that are too generous. The model is not at fault; it simply reveals permissions that were already wrong.
Where personal data is involved, keeping processing in the EU is the most straightforward option, as it sidesteps the rules on transfers to third countries. We also look at whether a provider based outside the EU could still move data in certain circumstances and note the result. For files with no personal data, customer contracts are often the bigger constraint.
In most cases no. At question time the system pulls the relevant passages from the index and grounds the reply in them. It costs less and needs no retraining whenever a file is updated. Fine-tuning is only worth considering for highly specialised wording or formats.
You could, by running open models on a GPU server hired inside your tenancy, so no text is passed to another AI company. The price is costlier GPU time, upkeep that falls to you and models that may trail the leading commercial ones. We compare both routes on your benchmark.
No. Infrastructure is defined in code, and document processing and the index are not tied to any host. A different model or region is scored on your benchmark, and we move only when it proves better or cheaper.
Building the environment is charged as a project or at €50 per hour plus VAT. Before launch you receive monthly running-cost forecasts for a few usage patterns; cloud charges go straight from the provider to you. Buying hardware is unnecessary.
Which files should become searchable, how many people will use it and what are your expectations on where data is kept? We will come back with suitable options.
Enquiry received
An engineer will write back no later than the following working day. Faults that leave your staff unable to work are always looked at first.
No town by that name here. Check the letters, or settle for the nearest larger place. Our engineers work only over the internet, so Kočevje and Koper get precisely the same treatment.