Building a personal LLM system.

The asset is the knowledge

Your Markdown, PDFs, spreadsheets, sources, context, and decisions are the durable asset. The language model is the instrument that helps you use them.

The source files remain the truth. The AI index should be disposable and rebuildable.

A practical first build

• Obsidian as the source of truth
• Ollama to run a local model
• AnythingLLM Desktop for ingestion, embeddings, retrieval, and chat
• 20–50 well-understood notes in one narrow subject area
• An 8B–14B model on the current Apple Silicon Mac

How RAG works

A question is converted into an embedding. The system searches a vector database, retrieves relevant passages, and gives those passages to the language model to construct a sourced answer.

The model does not need to memorise the vault. It needs to find the right evidence when asked.

Where it becomes useful

• Māori knowledge — preserve source, imagery, genealogy, place, and cultural context
• Obsidian — connect notes, surface prior decisions, and track changing beliefs
• Strength and health — combine training, recovery, technique, pain, and spreadsheets
• Media production — support transcripts, paper edits, briefs, deliverables, and approvals
• Decision filtering — test new ideas against priorities and displacement cost
• Language learning — move from literal structure to imagery, natural meaning, then grammar

Spreadsheets need two brains

Retrieval is good for finding workouts, dates, notes, and textual context. Exact calculations, comparisons, formulas, and trend analysis belong in Python or DuckDB.

Code performs the mathematics. The LLM performs the explanation and interpretation.