The workshop uses six notebooks, numbered 00 to 05. Each one ends with a Checkpoint cell that tells you what you should have reached by that point.

The six notebooks

Notebook	Budget	What you do	Tier notes	On GitHub
00_check_setup	10 min	Check Python, memory, disk, Ollama, the chat and embedding models (on 8 GB machines it checks for nishad/apertus:4b instead of the 8B) and one smoke call each	Same notebook on both tiers	00_check_setup.ipynb
01_python_and_ollama	20 min	Chat, system prompts, temperature, streaming, the OpenAI client, provenance logging	Same notebook on both tiers	01_python_and_ollama.ipynb
02_structured_output	25 min	Extract catalog card records into a Pydantic schema, validate them, measure against a gold file	Same notebook on both tiers	02_structured_output.ipynb
03_vocabulary_search	25 min	Embeddings, cross-lingual search over STW and UNESCO labels, a sqlite-vec index	On the 8 GB tier, download the prebuilt index files	03_vocabulary_search.ipynb
04_rag	20 min	Chunk documents, retrieve, answer with numbered citations, hit the context limit on purpose	8 GB tier: 3 chunks, 4,096-token window; copy in tier-8gb/	04_rag.ipynb
05_tool_use	15 min	Let the model propose GND, Wikidata and SPARQL calls; approve, correct or refuse each one	8 GB tier: single-tool cells only; copy in tier-8gb/	05_tool_use.ipynb

The tier notes refer to the two hardware tiers: a machine with 16 GB of memory or more runs the 8B model, an 8 GB machine runs the 4B.

The budgets in the table are the times given in the notebooks, and they are meant for self-study. In the live session we shorten them to fit the 110 minutes. The Part 2 timeline shows the minutes for each segment.

The site chapters that go with notebooks 01 to 05 are Python and Ollama, structured output, embeddings, RAG and tool use.

How to run them

Follow the environment page: install uv, clone the repository, install the packages and start JupyterLab. Open 00_check_setup.ipynb first; every line should show a check mark. Notebook 03 uses two prebuilt index files that you download into runs/. The steps and checksums are on the environment page, and the files are listed on the data page.

An 8 GB machine runs the same notebooks with the 4B model: set the environment variable WORKSHOP_TIER=8gb before you start JupyterLab, or change "16gb" to "8gb" in the first cell of each notebook.

Read them without running them

Every notebook is saved on GitHub with its outputs, so you can read it there without Python. The tier-8gb/ folder holds notebooks 04 and 05 as they ran on the 8 GB tier with nishad/apertus:4b. If you have an 8 GB machine, open them to see what to expect.

The 2025 notebooks

The notebooks from the 2025 edition stay available at the repository tag swib25.

Licenses

• Code: MIT.
• Vocabularies: STW under CC BY 4.0; UNESCO Thesaurus under CC BY-SA 3.0 IGO.
• Documents and catalog cards: CC BY 4.0.
• Model weights: Apache-2.0, with the Apertus LLM Acceptable Use Policy.

Sources, licenses and retrieval dates for every data file are on the data and licenses page.