beginner pathway
Understand local inference, embeddings, retrieval and evaluation before trusting model output.
Local AI is a system, not just a model file. The endpoint, model, prompt, context, retrieval sources and fallback rules decide what users actually experience.
Draw the path from browser question to local model answer and label every trust boundary.
What must never happen silently in local-only mode?
A local-only path must not silently contact a hosted model provider.
Embeddings turn text into vectors that can find semantically related passages. They are useful only when every chunk retains source, license, version and deletion policy.
Write two queries that should retrieve the same lesson using different words.
Why must a vector chunk retain provenance?
Without provenance, retrieved text cannot be cited, refreshed, deleted or rights-reviewed.
A model is not promoted because it sounds impressive. It is promoted for a workload after it loads, responds, cites evidence, follows schemas and passes held-out checks within resource limits.
Write one acceptance test for a tutor answer and one for a tool-calling answer.
What should happen to an untested model assignment?
Untested capability must be labeled honestly and kept out of routing decisions.
Create a small evaluation card for one local model task, including evidence, threshold and fallback behavior.