The challenge
Every company keeps its knowledge in hundreds of documents, wikis, tickets and emails – in German and English, with different terms for the same thing. Newcomers have to ask their way around. People who leave take knowledge with them.
Classic search finds words, not connections. And sending sensitive company documents to a cloud AI is out of the question for many.
The solution: OntologyMiner
OntologyMiner uses a locally running language model to extract people, teams, systems, projects and how they relate. Duplicates are merged across languages, and new terms are grouped into schema proposals.
A human approves the schema – and the system learns from every decision. Step by step this produces an ontology that fits the company instead of a rigid standard model.
How it works
-
Ingest & embed
Documents are split up and indexed semantically – for search by meaning rather than keyword.
-
Extraction with a local LLM
An open-source language model recognises entities and relationships – on your own hardware, without the cloud.
-
Consolidation
Rules and an AI judge merge duplicates, even across languages.
-
Learning schema
New types are collected as proposals. Approvals in the review feed into the next runs.
-
Measured quality
A synthetic test company with an answer key makes every improvement measurable.
Outcome
- A navigable knowledge map instead of scattered folders.
- The foundation for AI assistants that answer questions with sources.
- Privacy-friendly: language model and database run entirely locally.