Skip to content

Tutorials

Five tutorials, in order. Each one builds on the last and ends where the next begins. All of them run on a laptop; only the last needs a GPU, and it says exactly where.

They share one convention with the rest of this documentation: every step ends in a ✓ look for block stating what a correct result looks like. If your output disagrees with one, stop there. A mismatch is information — it tells you which step broke, which is far more useful than discovering at the end that something did.

Tutorial You will build Needs
1 Your first pipeline A folder of documents turned into an AI-ready corpus + knowledge graph, stage by stage. Nothing but the workspace. No network, no downloads, no GPU.
2 Understanding the output Fluency in every emitted artifact — and the ability to trace a graph edge back to the sentence it came from. Tutorial 1's output.
3 Retrieval A four-signal query path over the emitted files: dense, sparse, BM25 and graph traversal, fused and packed. Tutorial 1's output + one extra.
4 Bringing your own Provider A real Provider package of your own, registered as a plugin and green against the conformance suite. The workspace.
5 Production Served GPU models, staged runs, checkpoint/resume, and incremental corpus deltas. Tutorials 1–2; a GPU for the served-model sections.

Before you start

Work through the quickstart first if you have not — it installs the workspace and proves it end to end in one command. These tutorials assume you are inside a clone with uv sync already run.

They also assume the vocabulary from Concepts & Architecture:

  • a Capability is a narrow, model-agnostic interface (Parser, EntityExtractor, Embedder, …);
  • a Provider is a concrete implementation of one, shipped as its own package and discovered as a plugin;
  • a Stage binds one Provider to one node of a pipeline DAG;
  • the Runner executes that DAG; the Storage seam is where every artifact lands.

If you want the reasoning rather than the mechanics — why the pipeline extracts narrow and resolves wide, why retrieval is stateless, what the knowledge graph actually buys — read the deep dive. It is the argument these tutorials are the practice for.