Local AI with Ollama

Waymark works fully in app-only mode with no AI at all. Local AI is an opt-in enhancement that runs entirely on your machine through Ollama. Waymark never downloads a model for you and never calls a remote API.

How setup stays safe

waymark doctor          # capability profile + Ollama detection
waymark setup           # preview a safe local recommendation
waymark setup --apply   # write config.json with those defaults (no downloads)
waymark setup models    # compare recommended vs installed models

waymark setup models prints the exact ollama pull ... commands for any missing model — it does not run them. You stay in control of every download.

waymark models list     # detect Ollama + installed models
waymark models check    # compare your config's models against what's installed

AI-assisted capture

Once a chat model is configured, capture can ask it to structure a memory card:

waymark capture --local-ai "Ask the local model to draft this memory."
waymark capture --local-ai --yes "Save a confirmed local-AI draft in scripts."

Before any model call, Waymark checks Ollama readiness and shows the precise ollama pull ... command if the configured chat model is missing. The generated card is always shown for confirmation before it is saved — pass --yes only for scripted, non-interactive capture. If config or Ollama is unavailable, capture falls back to deterministic app-only drafting.

Semantic search

Semantic retrieval is opt-in and built on explicitly generated embeddings:

waymark embeddings status                       # what's indexed
waymark embeddings backfill                     # preview only (no model calls)
waymark embeddings backfill --apply --limit 10  # generate embeddings via Ollama
waymark ask "a fuzzy topic" --semantic          # hybrid keyword + tag + vector

Default ask uses keyword + tag matching. --semantic adds vector matches over the embeddings you've generated, and still renders cited source cards.

Always grounded

Even with local AI enabled, every answer cites saved memories or imported sources. Waymark separates saved memory from AI interpretation in its UI copy, and treats all model output as a draft until you confirm it.