formadata.

Local intelligence for your documents.

formadata reads, organises and answers your documents with AI. Everything runs on your own hardware, in your own building. Without your approval, your data never leaves it.

AI for your documents, and the data stays in the building.

Contracts, manuals, expert reports, minutes: your documents hold your company's knowledge. AI finds and condenses it in seconds. Many companies still do without, because the well-known services send every page to data centres outside the EU.

  • Data protection

    Personal data stays under your control. Nothing is transferred to third countries. A cloud provider only processes data if you allow an external model from the EU.

  • Trade secrets

    Quotes, engineering documents and contracts are processed and stored on your servers. No provider trains models on your content.

  • Works council and compliance

    You know where every file lives and which model has read it. That makes approval by the works council, data protection and IT easier.

Data flow with formadata Documents, the formadata server, the models and the database are all inside your building. Outside, Mistral in the EU is optional and needs approval. There is no connection to a US cloud. Your building Documents PDF documents formadata Server on your network Models Language models via Ollama Database Postgres with pgvector Mistral (EU) optional, needs approval US cloud no connection
Everything inside the dashed line runs on your premises. An external model from the EU can be switched on when an admin allows it for a workspace.

How it works

Four steps from a pile of documents to an answer with sources. Every step runs on your hardware.

  1. 1

    Upload

    PDF documents go into the library, organised in collections.

    Your server
  2. 2

    Read

    Docling recognises text, tables and figures, on a Mac using its own GPU.

    Your hardware
  3. 3

    Index

    Table of contents, sections and summaries are built automatically. Passages are stored as vectors in Postgres.

    Your database
  4. 4

    Ask

    You ask in your own words and get answers with sources. Tables appear exactly as in the original.

    Your model

What formadata does

Library

Organise documents in collections, share them with colleagues and control access with roles.

Structure

Table of contents, sections and summaries for every document, long manuals included.

Search and answers

Ask across whole collections. The matching passages are retrieved, reranked and turned into an answer with sources.

Exhibits

Tables and figures from the sources appear unchanged next to the answer, so you can check the numbers directly.

Pipelines

Chain processing steps as blocks. Every result is versioned and shows which data and which model produced it.

Chat in development

A conversation with your documents that keeps its history and shows the sources for every answer.

Every answer can be checked

formadata names the place in the document that each statement comes from and highlights the passage in the original.

Example · question to a collection

How often does the HX-200's air filter need changing?

Every 2,000 operating hours, and at least once a year 1. In dusty environments, the interval is halved 2.

Table 7.2 · Maintenance intervals (from the source, unchanged)
PartIntervalAt the latest
Air filter2,000 hyearly
Oil filter4,000 hevery 2 years
V-belt8,000 hvisual check yearly
  1. 1 HX-200 maintenance manual, section 7.3, p. 42
  2. 2 HX-200 maintenance manual, section 7.4, p. 43
HX-200 maintenance manual · p. 42

7.3 Air filter

The air filter protects the compressor and bearings from particles. Replace the filter every 2,000 operating hours, and no later than after twelve months. Use original filters only (part number 4471-02).

Apps built on your collections in development

A collection and a pipeline become an app with its own page: an assistant for the maintenance manuals, one for the contracts, one for the standards. Colleagues open the app and ask. The machinery behind it stays out of sight.

  • Chat

    HX series maintenance

    Collection
    HX series manuals
    Pipeline
    Questions and answers

    “Which oil goes into the compressor?”

  • Chat

    Contract desk

    Collection
    Purchasing framework contracts
    Pipeline
    Questions and answers

    “What notice period applies to supplier Berger?”

  • Chat

    Standards office

    Collection
    DIN and ISO, engineering
    Pipeline
    Questions and answers, thorough

    “Which IP rating does the outdoor housing need?”

  • One pipeline, many apps

    The pipeline decides what an app may set: the collection, the number of passages, the answer depth. Everything else stays as it was built. An improvement to the pipeline reaches every app.

  • Conversations stay private

    Every question runs as an ordinary run on your hardware. Histories belong to the person asking and are shared with the workspace only on request.

  • Access as in the library

    An app shows only what the person asking may see in the library. Whoever sets up an app needs edit rights on its collections.

Places from documents on a map planned

Contracts name sites, surveys name parcels, minutes name construction sites. formadata finds the places mentioned in your documents, gives them coordinates and shows them on a map. Every point leads back to the passage.

  • Recognise places

    The language model finds places, addresses and facilities in the text, even when they only appear in passing.

  • Coordinates without the internet

    Coordinates come from a gazetteer on your server. No address leaves the building, not even for geocoding.

  • A map with sources

    Map tiles are stored locally too. A click on a point opens the documents that mention the place, with the passage highlighted.

Running it on your hardware

formadata runs on a Mac in the office or on a Linux server in your data centre. All models are open and run locally, the language models through Ollama.

Mac with Apple Silicon

Installed natively. Docling reads on the GPU, about 5 pages per second on an M1 Pro.

  • No containers needed
  • Models through Ollama for macOS

Linux server

With Docker Compose, optionally with NVIDIA GPUs for larger models.

  • Postgres with pgvector
  • Ollama in its own container
Current selection. Every model can be swapped per customer, including for one of your own.
TaskModelOrigin and license
Reading documentsDoclingIBM Research, MIT
Writing answersQwen3 or Mistral Small, depending on hardwareAlibaba or Mistral AI, both Apache 2.0
Understanding page imagesQwen2.5-VLAlibaba, Apache 2.0
Embedding passagesbge-m3BAAI, MIT
Reranking resultsbge-reranker-v2-m3BAAI, Apache 2.0

All weights live on your server. A local model downloads nothing and sends nothing, wherever it was developed. External models stay switched off. A workspace admin can allow a model from Mistral in the EU for that workspace.

Let's talk about your documents.

Tell us which documents you have and what you would like to ask them. We will show you how formadata would run at your company.

Email
mail@marvinthiel.com