Commit Graph

3 Commits

Author SHA1 Message Date
panda 520c8b19d0 fix: 五轮修正失败根因修复 - 评分公式去掉field_coverage权重, namespace无条件检查, OCR自动发现文档类型 2026-05-24 22:44:37 +08:00
panda bd5bfbac2d fix: band-level windowed refine_layout + programmatic map_fields to prevent 91.5% content loss
Root cause: LLM receiving full 34k-char JRXML would regenerate from scratch
instead of modifying coordinates in-place, shrinking output to ~3k chars.

Solution (programmatic node control, not prompt engineering):

- New agent/jrxml_windower.py: decompose JRXML into header (never sent to
  LLM) + individual bands. Split bands >4000 chars at element boundaries.
  Reassemble with element count validation (>10% change = rollback).

- Rewrite refine_layout: per-band windowed LLM processing (~2-4k chars
  each). LLM cannot "reimagine" the entire report.

- Rewrite map_fields: 100% programmatic regex $F{field_N} -> real name
  replacement. Zero LLM calls, zero content loss.

- _sanitize_field_name: non-ASCII chars escaped to _uXXXX_ format for
  valid JRXML identifiers.

- Tests: 48 new unit tests (windower 28 + map_fields 20). All passing.
  Full suite 385 tests, zero regressions.
2026-05-24 08:55:38 +08:00
panda 70614dff5e feat: comprehensive v2 upgrade — streaming, error KB, file upload, layout analysis
Major changes:
- Streaming: LLM统一 _BaseLLM 接口 (invoke + stream), generate/modify/correct
  节点使用 get_stream_writer() 实现逐字输出, UI 节点平铺展开自动折叠
- Prompt外部化: 7个prompt拆分到 prompts/*.md, loader.py 支持热重载
- 错误自增长: backend/error_kb.py — 指纹去重 + ChromaDB持久化,
  correct_jrxml→validate 通过时自动入库, retrieve同时搜索错误KB
- 文件上传: backend/file_parser.py — PDF/DOCX/图片/文本解析,
  侧边栏多文件上传, 文本自动注入下一条消息
- A4模板识别: backend/layout_analyzer.py — 三种模式(完整A4/行片段修改/行片段新建),
  PaddleOCR元素提取 + 行分组 + JRXML section匹配
- 会话历史下载: jrxml_versions版本追踪 + 侧边栏历史版本下载按钮
- 预览修复: route_after_save跳过预览/导出意图的验证循环
- Ctrl+C修复: JS注入拦截Streamlit裸c键清缓存

Docs: CLAUDE.md (完整项目文档), ROADMAP.md (改进路线图)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-19 15:02:53 +08:00