NodeFlow (Hybrid Knowledge Orchestrator)

NodeFlow bridges the gap between chaotic brainstorming and structured knowledge execution. It is a local-first desktop application and orchestrator that converts spoken thoughts into a validated, human-approved knowledge graph. Built on top of Tauri v2, Rust, and React, NodeFlow continuously synchronizes with your local Obsidian vault, leaving your knowledge as plain Markdown files while keeping your GPU free for local inference. The Problem It Solves Waste & Opacity: Sending every minor prompt to cloud frontier models is expensive and opaque. NodeFlow routes tasks dynamically between local models (Ollama) and cloud APIs, accounting for exact token usage and financial cost. Hallucinated Actions: Traditional AI agents execute destructive changes autonomously. NodeFlow enforces a strict Human-In-The-Loop (HITL) pipeline where proposals must pass programmatic graph validation and explicit human approval. Amnesia & Lock-in: Chat interfaces lose context, and SaaS tools lock your data in proprietary databases. NodeFlow stores everything directly in your vault as Markdown + YAML. Core Features & Voice Loop Real-time Voice Execution: Built for the AssemblyAI Voice Agent Hackathon, NodeFlow opens a secure WebSocket stream with server-side JWT authentication. Dictated commands are parsed into precise graph operations (create, link, focus, condense, delegate). Strict Graph Guardrails: The engine validates returned IDs and field types against the real, live vocabulary of your graph before displaying a proposal. If a model invents an ID, it is blocked on the server. Hybrid Engine & Smart Caching: Features a two-tier cache (exact contract hash + semantic cosine similarity) that avoids recomputing duplicate prompts, saving tens of thousands of tokens while serving accurate cost metrics. Lossless Condensation: Complex subgraphs can be collapsed into macro-nodes with full lineage tracking, preventing visual pollution without losing underlying connections.