RAG Workflow: Claude + Obsidian Vault Integration

Chunk the vault. Split every markdown note into ~500 token pieces, keep the file path and heading as metadata. Embed locally. Run a local embedding model (nomic-embed, BGE, or MiniLM) over every chunk. No Claude call here, this part is free and runs on your machine. Store the vectors. Drop them into a local vector DB (Chroma, LanceDB, or sqlite-vec). Now 10M tokens of notes is a searchable index, not a prompt. Retrieve per question. Embed the question, pull the top 5 to 10 closest chunks by similarity. Call Claude. Send only the question plus those chunks through the API. Claude answers grounded in your own notes. Wrap it. Local web app on localhost, render the vault as a 3D force graph. Why the Claude usage drops: A naive build dumps whole notes into every prompt, so you pay for thousands of tokens you never needed. RAG sends only the few chunks that match, so input tokens per call collapse. The search itself runs on the local embedding model, so finding the answer never touches Claude. Prompt caching reuses the fixed system prompt at a fraction of the input cost. Air gestures over the graph: Run MediaPipe hand tracking in the browser, it returns 21 landmarks per hand per frame. Map gestures to the graph: thumb tip to index tip distance under a threshold is a pinch that selects a node. Open palm moving is rotate, two hands spreading apart is zoom. Render the graph with 3d-force-graph on Three.js, feed the landmark deltas into the orbit controls and node picking. For a real touchscreen, swap MediaPipe for pointer and touch events: pinch to zoom, drag to rotate, tap to select.