Home

AI NoteTaking — Wiki

Welcome to the project wiki. Use the sidebar to navigate between pages.

Quick links

Getting Started	Install, configure, and run the app
Architecture	System design, data flow, and design patterns
Audio Pipeline	Capture, VAD, loopback, and platform differences
AI Pipeline	WhisperX, diarization, LangGraph, and LLM providers
Database	SQLite schema, migrations, and FTS5 search
UI Guide	Compact widget, expanded view, settings popover
Sprint Roadmap	Current status and upcoming work
Contributing	How to contribute, commit conventions, issue templates

What is AI NoteTaking?

A privacy-first desktop notetaker that records meetings natively — no bots, no meeting links — and generates transcripts and structured summaries using your LLM of choice. Runs 100% locally with Ollama, or connects to OpenAI, Gemini, and Anthropic.

Built with Tauri 2 + Rust (core), React + TypeScript (frontend), and Python (AI engine).

Current release snapshot

• Development runtime: npm run tauri dev launches the Python engine from src-python/.venv/Scripts/python.exe and src-python/main.py.
• Windows packaging: the CPU-only PyInstaller sidecar is validated at 405.1 MiB against a 1536 MiB release budget.
• Structured summaries: the LangGraph workflow stores schema v2 summaries with evidence-backed decisions/action items. If a local LLM returns invalid JSON, the app now falls back to a useful transcript-derived summary instead of showing a hard failure.
• Latest benchmark highlights:
  • deterministic example fixture: weighted score 1.000;
  • QMSum two-fixture smoke run: weighted score 1.000;
  • AMI ASR IB4001: weighted score 0.924, WER 0.258, CER 0.200, latency 68.428s.
• Deployment guardrails: latest report is APPROVED for Python linting, Rust Clippy, TypeScript type-safety, RAG index audit, prompt structure, compliance/env-key checks, and MLOps pipeline structure.