Stack

Why Ollama, Chroma, and on-device capture for v1 — and when to switch.

Why Inpedia v1 uses these tools, what else was considered, and when to switch.

Full rationale in architecture decisions.

Stack

v1 components

ComponentADRv1 choice
Local LLM + embeddingsADR-005Ollama
Vector storeADR-006Chroma
Capture + OCRADR-007iPhone Notes scan → Mac text (manual); Apple on-device OCR

Decision matrix

When to revisit v1

Stick with v1 if…Revisit if…
**Ollama** — privacy and zero marginal cost matter more than speedAnswers are too slow or weak; you need a stronger local model or hybrid cloud for hard queries only
**Chroma** — single Mac, single user, simple persistYou need multi-device sync, heavy metadata filters, or scale beyond ~100k chunks
**Notes + Apple OCR** — captures stay on-device, good enough for printHandwriting quality is poor; you want automated iPhone → Mac pipeline or batch OCR

Explicitly deferred

Not chosen for v1 (by design)

  • OpenAI / Anthropic APIs — cost, privacy, corpus leaves machine (ADR-004)

  • Cloud OCR (Google Vision, Adobe Scan cloud) — images leave device before Inpedia exists

  • Confluence / Notion AI — SaaS wiki rent, not local RAG

  • Qdrant / Pinecone hosted — ops overhead or cloud corpus for v1 tracer