Knowledge Pipeline

Local-first personal knowledge system: extract from web sources and Apple Notes into an Obsidian vault, ask questions over it with local AI (LM Studio + Gemma), and turn notes into blog posts published via Vercel (and Medium drafts).

Design rule: no external database. The vault's markdown is the only source of truth. Vectors live in a rebuildable file cache (numpy npz + manifest); keyword search (BM25) and the [[wikilink]] graph are rebuilt in memory. Delete .cache/ any time.

web / YouTube / papers ──┐
Apple Notes (plugin) ────┼──> Obsidian vault (markdown + wikilinks)
agent "remember" ────────┘         │
                                   ├──> hybrid RAG (numpy cosine + BM25 + graph expansion)
                                   │        └──> QA with citations (Gemma via LM Studio)
                                   └──> blog drafts (3-pass: draft, rules edit, clarity edit)
                                            └──> Vercel deploy hook / Medium draft

Run

# Backend (port 8600)
cd backend && python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
cp ../.env.example .env   # adjust vault path, models
.venv/bin/uvicorn app.main:app --reload --port 8600

# Index + ask
curl -X POST localhost:8600/index/refresh
curl -X POST localhost:8600/ask -H 'content-type: application/json' -d '{"question":"..."}'

# Frontend shell (port 5273, proxies /api -> 8600)
cd frontend && npm install && npm run dev

# Tests
cd backend && .venv/bin/python -m pytest -q

Requires LM Studio serving on localhost:1234 with a Gemma 3 chat model and an embedding model loaded (see .env.example for names).

Layout

• backend/app/ - FastAPI + agent tool loop; rag/ (retrieval), ingest/ (web, YouTube, papers, Apple Notes), memory/ (vault-backed agent memory), publish/ (blog + Medium)
• blog/ - RULES.md (style + hard rules) and posts/; site is nextjs-obsidian-blog-kit (see blog/README.md)
• frontend/ - Vite + React shell (Ask / Search / Graph / Ingest / Blog / Settings)
• .claude/agents/self-documenter.md - assigned doc-sync agent; ADRs in docs/adr/
• STATUS.md - live checklist the portfolio dashboard reads

Reference designs: chat-with-your-data accelerator (structure), ObsidianRAG (hybrid + graph retrieval, minus ChromaDB), Obsidian Importer (Apple Notes), knowledge-manager (Zettelkasten ingestion), nextjs-obsidian-blog-kit (publishing).