ALL STORIES.

Grounding connects a response to evidence that can be inspected. It is a relationship between a claim and a source.

An evaluation set turns ‘this model seems good’ into a question you can investigate with repeatable examples.

A plausible answer can be unsupported. Hallucination is a reliability problem to design around, not a flaw solved by asking for confidence.

Structured output gives software a predictable shape to work with. Correct structure and correct information are separate tests.

Tool calling lets a model request a defined operation. The application remains responsible for executing and checking it.

System instructions describe the assistant’s role and operating boundaries. They work alongside permissions and application logic.

A useful prompt resembles a good work brief: a clear task, relevant context and a definition of success.

Fine-tuning changes model behaviour through additional training. It works best when the target behaviour can be demonstrated and evaluated.

Retrieval-augmented generation gives a model selected source material at answer time. The retrieval step deserves as much attention as the wording.

Vector search helps find related ideas. Its value depends on the documents, filters and ranking around it.

Embeddings represent items as numerical vectors so software can compare patterns of similarity.

A large context window gives a model more material to consider. It does not provide perfect recall or automatic prioritisation.