Masterbrain Bytes

Project ยท 2025

Editorial RAG

Editors find earlier articles from several publications faster, with source and page reference.

Over the years, a specialist publisher has published a great many articles across several publications. When planning new pieces, the newsroom wants to know quickly what has already appeared on a topic, even if different terms were used back then.

For Landwirtschaftsverlag we built an editorial search that combines keywords and meaning. The embeddings are created locally, and every hit points to publication, issue and page. That keeps every answer verifiable.

How the editorial search works

  1. Prepare content

    Take over publication data from the editorial system. Texts and metadata stay together.

  2. Index passages

    Split texts into sections. Local embeddings and a full-text index are built in LanceDB.

  3. Search hybrid

    Combine keywords and semantic similarity and merge the result lists with rank fusion.

  4. Open sources

    Results with publication, issue and page reference. The original articles remain verifiable.

In the newsroom's daily work

  • Several publications

    One shared access to the archive, with a particular focus on profi.

  • Editorial planning

    What have we already published? When and how often have we covered a topic?

  • Connected to existing systems

    Connection to the editorial system, automatic updates, and API and MCP interfaces.

  • Search and answer kept apart

    Search finds evidence. A language model can phrase an answer from it. Both parts are checked separately.

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