Service 01
Local AI & RAG
We make your knowledge searchable: text, images, video and audio, with verifiable sources and on your own hardware.
Many companies sit on a large body of knowledge: articles, manuals, minutes, quotes. A classic full-text search only finds what matches word for word. With retrieval-augmented generation (RAG) this knowledge becomes searchable by meaning, and every answer points to its source.
Not just text
RAG does not only work with text. We also unlock 2D images, videos and audio files: image archives become searchable by description, videos by their content and audio files by words, acoustic similarity or the meaning of their description. That turns a media archive into searchable knowledge, too.
More than language models
Local AI is more than a chat window. In one of our own tools we have integrated more than 40 local AI models into a single application: for images, 3D models, textures, rigging and animation. A second tool generates sound effects, voices and music. We bring this experience into client projects:
- Porting models: we convert models from PyTorch formats to ggml so that they run in C++ engines instead of a Python environment, and compare the result with the original.
- Turning research into software: we build methods from current research papers into working applications.
- Many models, one application: the models compute in their own processes, separate from the user interface. That way the graphics memory can be released completely between tasks.
- Hardware of your choice: the same application runs on the CPU and on graphics cards from Nvidia and AMD.
- Licences in view: for every model we record where it comes from and whether it may be used commercially.
Your data stays in-house
We build such systems so that the data stays in your environment: the embeddings are created locally, and index and search run on your hardware. Whether a language model phrases the answer, and where it runs, we decide together based on your data.
What we pay attention to
Check search and answer separately
Search finds evidence. A language model can phrase an answer from it. Both parts need their own quality checks.
Include questions without an answer
A good system shows when there is no evidence instead of inventing an answer.
Respect permissions
Roles and document rights also apply in search. Restricted content stays restricted.
Not just text
Images, videos and audio files can also be searched by meaning, for example a photo by its description or a sound by acoustic similarity.
Integrate with existing systems
Connection to existing editorial and document systems, automatic updates, and API and MCP interfaces for AI agents.
How a knowledge search comes together
Prepare content
Take over data from the existing system. Texts stay together with metadata such as publication, issue and page.
Index passages
Split texts into meaningful sections. Local embeddings and a full-text index are built on your hardware.
Search hybrid
Combine keywords and semantic similarity and merge the result lists with rank fusion.
Open sources
Results come back with their location. The original documents remain verifiable.
Reference
Editorial RAG for Landwirtschaftsverlag
Editors find earlier articles from several publications faster, with a focus on profi: what have we already published, when, and how often?
Does this fit your task?
Tell us about your use case and your data. We will get back to you personally.
info@masterbrain-bytes.com