@ai-coach/adapter-ollama

The local sparring partner. Everything the coach asks you during sparring, studying and synthesizing comes from a Gemma model running on your own machine through Ollama. Your notes never leave the computer.

What it does

Class	Its job	Used by
OllamaSparringPartner	Writes the opening challenge and the follow-up questions, and decides the Tier 1 verdict. Each gap it raises is checked against what you actually said.	coach spar
OllamaStudyPartner	Gives a guiding question per gap, a clue when you ask for help, and checks whether your new explanation closes the gap.	coach study
OllamaStyleClassifier	Reads a note and picks the coaching style that suits it, when you haven't chosen one.	spar, study
OllamaSynthesisPartner	Groups your marks into ideas, asks gap questions, drafts core assertions from your words, and suggests buzzwords.	coach synthesize
OllamaClient	The shared connection to Ollama, with retries and checks on what comes back.	all of the above

All the prompts are in prompts.ts, so tuning how the coach talks starts there.

Setting it up

ollama pull gemma3

Then set the model in coach.config.json:

{ "ollama": { "model": "gemma3:latest", "contextWindow": 16384 } }

Setting	Default	Notes
model	(required)	Any name from ollama list.
host	http://127.0.0.1:11434	Must be your own machine (see below).
contextWindow	16384	How many tokens Gemma reads at once. Raise it for long chapters.

Good to know

• It refuses to go online. The host must be http on 127.0.0.1, localhost or ::1. Anything else stops with Tier 1 runs offline…. This keeps the promise that your notes stay local.
• Your notes are treated as data, never as instructions. Note text, transcripts and marks are wrapped in tags with a random marker, so a note saying "ignore your instructions" has no effect.
• The coach doesn't quote your note back to you. If a question copies the note's wording, it's rewritten. If that keeps happening, the note is skipped with ChallengeLeakError.
• Too long? If a note won't fit in the context window, you get a clear ContextTooSmallError rather than a silently cut-off note.
• Small models make mistakes. Every reply is checked: right shape, a real question where a question is expected, no filler. A bad reply is retried once, and the use case falls back gracefully if it's still unusable.

Tests

npx nx test @ai-coach/adapter-ollama

The tests use a tiny local HTTP server in place of Ollama, so they don't need Ollama or a model.