IT Let’s talk

Digipa AI · 01

Your private LLM, on your documents

Open-weight models on your servers or in a European cloud, with search across company documents (RAG), permissions respected and sources always cited.

Book a 30-minute call What you get

The problem

Contracts, procedures, technical manuals, tickets: your company’s knowledge is scattered across thousands of files. An internal assistant lets you query it in natural language, always showing which document each answer comes from.

The hard part is rarely the model: it is the data, the permissions and the infrastructure. That is exactly what we do every day with DroneDB, where we index, search and share terabytes of data with multi-level permissions.

Example questions

  • “What is the returns procedure for B2B customers?”
  • “Summarise the changes between version 3 and 4 of the specification.”
  • “Which customers have an exclusivity clause about to expire?”
  • “Find the service reports on model X from the last two years.”

How it works

From question to answer, verifiably.

  1. Collection

    We connect the sources and index the documents, respecting who can see what.

  2. Question

    The user asks in natural language, from the web interface or the tools they already use.

  3. Retrieval

    The system finds the relevant passages among the documents that user is allowed to read.

  4. Answer

    The model answers and cites its sources. If it finds nothing, it says so.

  • Your data stays yoursHosting in Europe or on your own servers. No model training on your data.
  • Sources always citedEvery answer shows the document or query it comes from. If there is no source, the assistant says so.
  • A human in the loopHand-off to a person when needed, human approval for anything that commits the company.
  • Measured, not promisedA set of real questions agreed at the start and re-run after every change.

What you get

Six deliverables, no black boxes.

  1. 01

    Data and use-case audit

    Inventory of documents, sensitivity, permissions and hosting level for each collection.

  2. 02

    Model selection

    Models compared on your real questions, Italian included, with a reusable test set.

  3. 03

    Infrastructure

    On-premise GPU servers or a private European cloud, secured, with backups and monitoring.

  4. 04

    RAG pipeline

    Search that respects permissions, answers with citations and audit logs.

  5. 05

    Integration

    REST API, web interface, SSO and connectors to network folders, DMS and management software.

  6. 06

    Handover and management

    Documentation, administrator training, managed service with periodic re-evaluation.

Where it runs

Three hosting levels, your choice.

  1. Level A

    API under contract

    Commercial models via API, with a data processing agreement. The quickest to get started.

  2. Level B

    European cloud

    Models running on dedicated infrastructure in EU data centres. A balance of power and control.

  3. Level C

    On your servers

    Open-weight models installed on premise. No data leaves your perimeter.

How we start

From the first meeting to ongoing management.

  1. Discovery call

    30 minutes, free: use cases, available data, constraints.

  2. Pilot

    One or two real use cases, with an agreed test set and success criteria.

  3. Roll-out

    Integration with your systems, permissions, logs and user training.

  4. Management

    Monitoring, content updates and periodic re-evaluation of the models.

Who builds your AI

Engineers who run real data, in production.

We don’t show AI case studies we don’t have. We show what we run every day: DroneDB, our open-source platform for geospatial data, where we index, search and protect terabytes of files with multi-level permissions.

Discover DroneDB
1,300
active users on DroneDB
2,500
organisations
4.5 TB
of data managed
290★
on GitHub, MPL-2.0 code

Figures published on dronedb.app and GitHub, October 2026.

FAQ

Before we start

Is our data used to train the model?

No. Documents are indexed for search, not used to train the model. With on-premise hosting they never leave your servers.

Do we need to buy hardware?

Not necessarily: you can start in a European cloud and move on premise when volumes and requirements justify it. If needed, we give you concrete hardware recommendations.

Does it work in Italian?

Yes. Models are compared on your own questions, in Italian, before we choose.

Other Digipa AI solutions

02

AI training

Hands-on courses on the tools your team actually uses, tailored to each role and documented as an AI literacy measure under the AI Act (Art. 4).

03

AI for customer care

An AI assistant on your website, WhatsApp and helpdesk that answers from your official information and hands over to a person when needed.

04

AI for business intelligence

Natural-language questions on your management software, CRM and databases: answers with numbers, charts and the query that produced them.

Free introductory call

Let’s talk about custom AI.

A free 30-minute call with an engineer: use cases, available data and a concrete idea of timings and costs.

Book the call +39 02 9166 5401