Update .gitignore and enhance README with new project plans

- Added `RAG_data/` to .gitignore to prevent tracking of specific data files.
- Expanded the README to include detailed plans for document type-specific chunk templates, improvements in document understanding, OCR accuracy, and layout detection design.
- Removed several outdated user manuals and documents to streamline the repository.
This commit is contained in:
陈辅元
2026-07-21 11:02:58 +08:00
parent 8466ed2fbe
commit 920c0b4a54
9 changed files with 5 additions and 1 deletions
+1
View File
@@ -13,6 +13,7 @@ storage/uploads/
storage/results/
storage/assets/
output/
RAG_data/
# IDE / local agent / task notes (keep local only)
.idea/
+4 -1
View File
@@ -428,7 +428,10 @@ hits, n = recall_chunks("login password", result.chunks, top_k=5)
## 后续计划
- [ ] PDF OCR / 视觉描述增强
- [ ] **按文档类型提供专用 Chunk 模板**:在现有通用、按行和标识符切分的基础上,为论文、书籍、法律法规、演示文稿、问答和表格等文档提供结构感知的模板,并保留章节路径、页码、内容类型等元数据。
- [ ] **参考 RAGFlow DeepDoc 完善文档理解层**:在切分前统一完成 OCR、版面分析、表格结构识别、图片/图注关联和文档结构恢复,输出标准化 Block / Document Tree,供不同 Chunk 模板复用。
- [ ] **完善 OCR 精准识别**:增强扫描 PDF、文档内嵌图片和复杂背景图片的文字识别,补充方向校正、语言检测、置信度过滤与识别结果回填,并建立可量化的准确率评测集。
- [ ] **引入 Layout Detection 设计**:识别标题、正文、列表、表格、图片、图注、页眉和页脚等版面区域,结合坐标恢复多栏阅读顺序,并保证图片、表格与上下文在切片中的相对位置。
- [ ] 向量召回(Embedding)
- [ ] wps 格式支持
Binary file not shown.
Binary file not shown.