Please enable JavaScript.
Coggle requires JavaScript to display documents.
安規知識庫查詢小幫手 - Coggle Diagram
安規知識庫查詢小幫手
MinerU
*.md
MinerU 解析後的主要可讀內容（段落/標題/可能含部分表格或圖註等）
2026-01-13 Tue, 是有包含 Table 文字的.
*_content_list.json
*_middle.json 的簡化版, 依序把内容攤平成一個 list.
type: text / table / image / equation
page_idx: 在第幾頁 (從 0 開始)
bbox: 該塊在頁面上的框 (normalized 0-1k)
圖表還有 img_path / caption / footnote 等資訊
*_layout.pdf
視覺化切塊的部分, 用來人工檢查切得好不好
*_model.json
模型推理原始結果 / 佈局偵測結果, 含 class, polygon, bbox. 可用來精確定位.
*_origin.pdf
原始 pdf 檔
*_middle.json
更完整的中間結果（含更細的層次：block/line/span…）
*_content_lis_v2.json
2026-01-13 Tue, 多模態版本 content list, 依照頁面區分 paragraph, image, table, text, header, list, etc.
2026-01-13 Tue, Table 有圖也有文字内容 (HTML 格式儲存), 這些文字已經包含在 .md
[DONE] V0.1 PoC, no multimodal embedding but OCR text
1) Ingest
Captioning
2026-01-12 Mon, caption (qwen2.5vl:7b >> llama3.2-vision:latest), cf. Notion
OCR
2026-01-12 Mon, OCR (qwen2.5vl:7b >> llama3.2-vision:latest), cf. Notion
2026-01-14 Wed, DeepSeek-OCR seems to be more powerful than Qwen2.5VL, cf.
DeepSeek-OCR
PLAN: 對每張圖做 "文字化索引" (image -> text)
2) Query 文字檢索
2026-01-13 Tue, image block -> md
2026-01-13 Tue, md 較適合 RAG chunk (相比 csv)
2026-01-14 Wed, PLAN: Use existing .md, and attach image's .md we generated to RAG Knowledge Base.
3) Rerank
2026-01-09 Fri, Dify 目前外網有問題無法連 rerank model. And somewhat we can't host rerank model locally (by Sheng)
4) UI
2026-01-9 Fri, can retrieve content (with text question)
Progress
2026-01-16 Fri, Released
Goal
[DONE] 2026-01-14 Wed, Let's PoC image RAG
2026-01-14 Wed, V1: can provide image link (downloadable)
2026-01-9 Fri, V0: can retrieve content (with text question)
User Stories
[Story 1] 輸出投影片; [David][重要3][緊急1] 將查詢結果整理成投影片供下載.[2026-01-16 Fri,]
[Story 2] 輸入檢查清單自動逐條檢查. [David][重要3][緊急1] As an SE, I want AI Agent 能把所有要檢查的項目是否符合, so that [2026-01-20 Tue]
[Story 3] 資料導入知識庫自動化. [Joen][重要2][緊急1][2026-01-21 Wed]
[Story 4] 知識協助 [David][重要3][緊急2] As a senior-level manager, I want junior component engineers to use AI agents to verify regulations, so they can achieve senior-level performance.[2026-01-22 Thu]
[Story 5] 指定項目檢索 [SE Jerry?][重要3][緊急3] As an SE, I want 以自然語言檢索取得特定條件下的所有安規項目, so 可以很快的得知需要的資料項目與相關圖片. [2026-01-27 Tue.]
[DONE] 2026-01-16 Fri. Let's continue to do full doc.
2026-01-21 Wed, Let's deal [Issue 2]
See survey in coggle
Chat to Slides
Issues
[Issue 1] 圖片顯示問題. [Ink][重要3][緊急1][偶發]. 圖片有找到, LLM也有輸出, 但在 Dify render 的過程被消失.[2026-01-20 Tue]
[Issue 2] 圖片資料問題; [David][重要3][緊急2][常態] 有些資料是 table 中帶有圖片, 目前 MinerU 會把整個 table 當成一張圖, 表中圖就不會獨立分出來. See PDF "data\安规标示规范_FQ-C-QI-SFT-004 A10_20250514 Final" [2026-01-21 Wed]
[Issue 3] 有 feedback 說 Qwen-2.5VL-7b OCR 還是有 hallucination? 可能需要試試更多 model. 等 Joen 補充結果資料. [Joen][重要2][緊急1][偶發][2026-01-26 Mon]
Resource
in-house LLM
qwen2.5vl:7b
llama3.2-vision:latest
V1.0 (Full doc)
Progress
2026-01-19 Mon,
Somehow Qwen2.5VL works fine, no repeating issue.
2026-01-16 Fri,
Modularize functions
Dealing with Qwen2.5VL's repeating issue.
2026-01-20 Tue,
Fix file name issue by hand.
Released.
Ongoing
2026-01-21 Wed, Doing [Issue 2]