AI Library &
Knowledge Distillation
AI 書庫與
知識融煉
Ingests PDFs / EPUBs, OCRs them with a vision model, then runs a five-agent pipeline that maps structure, mines concepts, distills core mental models, and emits reusable LLM Skill modules. An in-book chat agent retrieves pages via tool calls. 匯入 PDF/EPUB 並以視覺模型 OCR,接著由五個 AI 代理接力:分析結構、萃取概念、蒸餾核心思維模型,最終生成可重複使用的 LLM Skill 知識模組。書內聊天代理透過工具呼叫即時檢索頁面。
These screenshots are from the actual app, captured locally — OCR runs on a vision model that needs a GPU, so there's no public-facing demo. What follows is the real UI and real data, not mockups. 下面的截圖來自本機實際跑起來的應用程式 — OCR 需要 GPU 跑視覺模型,所以沒有公開版 Demo。以下是真實畫面與真實資料,不是示意圖。
How It Works運作流程
01 · Import匯入
Any format goes in任何格式都能匯入
PDF, EPUB, TIFF, or a ZIP/CBZ/RAR of scanned pages — drag in a file, a folder, or point at a local path. Rendering DPI and output format are configurable per import. PDF、EPUB、TIFF,或掃描頁面壓縮成 ZIP/CBZ/RAR 都可以 —— 拖檔案、拖資料夾,或直接填本機路徑。渲染 DPI 與輸出格式每次匯入都可調整。
02 · Library書庫
Organized by reading status依閱讀狀態組織
Every OCR'd book lands on the shelf automatically, filterable by 全部 / 閱讀中 / 未開始 / 已完成 (all / reading / not started / finished), with per-book page counts and progress. 每本 OCR 完成的書會自動加入書架,可依「全部/閱讀中/未開始/已完成」篩選,並顯示每本書的頁數與閱讀進度。
03 · Read + Ask閱讀+提問
The chat agent retrieves real pages, not vibes聊天代理檢索的是真實頁面,不是憑印象回答
Reading the "Attention Is All You Need" paper here — the right-side agent calls a search_book tool that keyword-matches the OCR'd page text, retrieves the top matching pages, and grounds its explanation of scaled dot-product / multi-head attention in what's actually on page 4.
這裡在讀《Attention Is All You Need》—— 右側代理呼叫 search_book 工具,對 OCR 出來的頁面文字做關鍵字比對,抓出最相關的頁面,再根據第 4 頁實際內容解釋縮放點積注意力與多頭注意力,而不是憑空生成。
Tech Stack技術工具
Engineering Notes工程筆記
Five-Agent Distillation五代理蒸餾
Outline analyzer and chunk miner run on a lighter "flash" model tier; the wisdom distiller, document weaver, and skill generator run on a heavier "pro" tier — split by how much reasoning each stage actually needs, not uniformly. 結構分析與段落萃取用較輕的「flash」模型層級;核心思維蒸餾、文件編織、Skill 生成則用較重的「pro」層級 —— 依每個階段實際需要的推理量分級,不是統一用同一顆模型。
Keyword Retrieval, Not Vectors關鍵字檢索,不是向量
Page retrieval for in-book chat is keyword-weighted text search over cached OCR text, not an embedding index. Simpler, and for single-book context windows this size, accurate enough — no vector DB to operate. 書內聊天的頁面檢索是對快取的 OCR 文字做關鍵字加權比對,不是向量索引。更簡單,而且在單本書這種規模的檢索範圍內已經夠準,不用額外維運向量資料庫。
Provider-Agnostic LLM Layer不綁定單一 LLM 供應商
The five-agent pipeline talks to any OpenAI-compatible endpoint (URL + Bearer token), so swapping model providers is a settings change, not a code change. 五代理流程走通用的 OpenAI 相容介面(URL + Bearer Token),換模型供應商只要改設定,不用改程式碼。