eisonAI: A Structure-First, Local-First Reading Assistant

eisonAI is an iOS/iPadOS Safari extension and native app that prioritizes structure as the entry point to reading. Users can view key points and relationships first before deciding where to dive deeper. It supports multiple inference methods: Apple Intelligence, MLX local models, and BYOK (Bring Your Own Key). Its core innovation is Cognitive Index™, aiming to reduce linear reading and cognitive load while making content structure visible.

Traditional linear reading (from start to end) faces challenges in the information age: forgetting key points' locations, losing saved content, and disorganized thoughts when writing. eisonAI's design philosophy is: Don't just remember content, remember its purpose—organizing content like a library to know where things belong and how to retrieve them.

eisonAI's core innovation is Cognitive Index™, which makes structure visible before content, lowering the cost of locating meaningful parts. It achieves four goals: 1) Reduce linear reading (no need to follow narrative order); 2) Encourage conscious thinking (focus on judgment/understanding instead of keeping context in working memory); 3) Make structure visible (see relationships and key points first then choose paths); 4) Lower cognitive load (quickly locate meaning-dense sections via structured output).

eisonAI uses AnyLanguageModel for a unified inference stack:

• Safari Popup: Runs via Apple Intelligence or BYOK.
• Native App: Supports three methods: Apple Intelligence (local AI on Apple devices), BYOK (configure OpenAI-compatible/Anthropic/Gemini/Ollama endpoints), MLX local models (download from Hugging Face MLX repo).
• Local Model Management: In Settings → AI Models → MLX Models; only supports MLX repo (not GGUF/llama.cpp).

Key features include:

1. Safari Extension: Generate summaries and structured highlights within Safari without leaving the browser.
2. MLX Model Library: Browse mlx-community models (hides overly large ones by default; supports custom Hugging Face MLX repos).
3. Long Document Support: Chunk processing; auto-switch to cloud when local model context is insufficient.
4. CloudKit Sync: Cross-device sync of library (iPhone/iPad reading progress).
5. Library & Tags: Save, tag, search processed articles to build personal knowledge base.
6. Thinking Language: Adjust model output language for multilingual scenarios.

Technical Evolution: Early versions used WebLLM/WebGPU in Safari; current version uses native bridge for popup (Apple Intelligence/BYOK), downloads MLX models from Hugging Face at runtime (not packed in extension), and queries mlx-community sorted by lastModified. Privacy: Open-sourced under PolyForm Noncommercial License 1.0.0—users can review code to see how data is handled, avoiding reliance on black-box service privacy promises.

System Requirements: iOS/iPadOS 18.0+; iPhone/iPad; Apple Intelligence needs supported devices and Foundation Models runtime; local models only from Hugging Face MLX repo. Use Cases: Research reading (extract key points from papers), news tracking (structure reports to follow the event thread), learning notes (structure textbooks for knowledge frameworks), content curation (save valuable web pages into searchable personal knowledge base).

eisonAI represents a new direction for reading tools—from linear consumption to structured exploration. With Cognitive Index™ and flexible local/cloud inference, it offers a privacy-protective and powerful reading assistant for knowledge workers. It's worth attention for users focused on personal knowledge management, local AI deployment, and reading experience innovation.