TRAIL
Beyond the model
Retrieval, tools, memory, and multimodality — the machinery bolted around the model that does most of the work.
Almost nothing you use is just a model. It is a model with search in front of it, tools beside it, a memory store behind it, and a system diagram nobody has drawn. That surrounding machinery is where most quality — and nearly all risk — actually lives.
By the end you’ll be able to explain how semantic search finds documents with no shared keywords, why a grounded answer can still cite the wrong source, where a human checkpoint belongs in a tool-calling loop, and how to draw an AI feature end to end so its failures have owners.
Begin trail →- 1UnderstandEmbeddings
Text as coordinates, so “nearby” means “similar in meaning”. It’s the basis of search without keywords — and of a specific way that search misleads.
- 2UnderstandRetrieval-augmented generation
Fetch the passages first, then answer from them. The main practical control on fabrication, and the stage most often responsible when answers are wrong.
- 3UnderstandHow a model uses tools
A model cannot send an email; it can propose one. Understanding that gap precisely is what tells you where the approval gate belongs.
- 4UnderstandMemory and state
Models are stateless, so every “it remembers me” feature is a store outside the model — which is exactly why it can be inspected, corrected, and deleted.
- 5UnderstandHow multimodal models work
Images and audio become tokens too. That explains both what these models are impressively good at and the specific detail they cannot see.
- 6UnderstandDrawing the system
The ring’s closing move: draw the whole path, mark where it can fail and who would notice. This is what turns understanding into judgment.