Notely

AI Subsystem Overview ​

Notely features a 13-domain modular, local-first AI platform designed around private data control. Markdown files remain the absolute source of truth, parsed and indexed into offline-first databases to fuel assistant reasoning.

Capabilities at a Glance ​

1. Master Flow Orchestrator (AIFlow.js) & 13-Domain Architecture ​

  • All LLM queries flow through AIFlow.js, executing a 5-stage pipeline across 13 decoupled domain module facades (compaction, planner, personas, prompts, context, graph, embeddings, memory, executor, tools, grounding, formatter, testing).

2. Zero-Latency Context Compaction (ai/compaction/) ​

  • Automatically compacts long chat sessions (>4 messages) into an Executive Memory Summary + recent 4 turns verbatim, slashing LLM input tokens by ~75-80% with 0ms overhead.

3. SQLite Knowledge Graph ​

  • Outbound relations, tags, and CTE traversals mapped into ai-graph.db.
  • Visualized interactively in the sidebar.

4. Local Embedding Indexer ​

  • High-performance ai-embeddings.db storing note chunk vectors.
  • Runs entirely offline using a local ONNX runtime for BGE-small-en-v1.5 embeddings, or falls back to HuggingFace APIs.
  • Background Index Worker priority queues processing note changes debounced.

5. Persona Registry (Markdown Source of Truth) ​

  • All personas (builtin and custom) use .md files with YAML frontmatter as their authoritative Source of Truth.
  • SQLite (personas.db) acts strictly as an index registry. Custom personas persist as formatted .md files to disk (appData/personas/*.md).
  • Customize instructions, descriptions, metadata, and preset avatar icons (🤖, 💻, 🧠, etc.).

6. Diagnostics, Flow Telemetry & Trace Logs ​

  • Professional AI Health panel to verify subsystem initialization.
  • Flow Telemetry tab displaying 5-stage timeline cards, system prompt viewer (Copy/Expand), tool calls, compaction stats, and latency breakdown.