Next steps
Ideas for continuing after the workshop. Start small, keep a person in the loop to review what the model produces, and check every output against your own data.
Library tasks to try next
Subject indexing against GND or LCSH, with review. Ask the model to propose subject headings for a title and abstract. Then use search to match each proposal to a real heading, and have a cataloger accept or reject it. Builds on notebook 03 (vocabulary search) and notebook 02 (structured output). For indexing in your everyday work, look at Annif first.
Authority reconciliation. Match names in a spreadsheet to GND records with the lobid reconciliation API, using OpenRefine as the interface. A model can help by putting messy name strings into a consistent form before matching. Builds on notebook 05 (tool use).
Record quality assurance. Ask the model to point out suspicious values, such as an ISBN in the wrong field or a date in a form that software cannot read. Have it return the issues as a validated list. Builds on notebook 02.
OCR post-correction. Give the model a passage of OCR text that contains errors and ask for a corrected version. Then compare the two with a diff tool before you accept anything. The tool shows exactly what changed. Builds on notebook 01 (Python and Ollama).
Multilingual catalog text. Extract fields from German, French or Italian cards, or translate subject notes, and keep the original beside the result. Builds on notebook 02.
SPARQL drafting with human review. Let the model propose a query from a plain-language question, show it to the person who asked, and run it only after they approve. Builds on notebook 05.
Tools to look at
- Annif for automated subject indexing with your own vocabulary.
- The lobid reconciliation service for matching against GND from OpenRefine.
- Skosmos for publishing and browsing SKOS vocabularies, as the STW thesaurus does.
SWIB26 sessions on the same day
- The keynote on critical AI literacies
- The CUNY talk on a human-in-the-loop approach to authority work
- The talk on open LLMs for keyword assignment
- The AI policy community session
See the SWIB26 program for times.
Communities
Reading list
- The Apertus technical report, arXiv 2509.14233
- The Open Source Initiative’s Open Source AI Definition
- The European Commission’s template for the public summary of training content under the AI Act
- Simon Willison’s writing on embeddings