AI UNDERDOGSDAILY PICK
AI UNDERDOGS
Tesseract 总读错字，让 LLM 改
Tesseract mangles text. Let an LLM fix it
Dicklesworthstone/llm_aided_ocr
Tesseract 免费又好用
但读出来的字三天两头出错
这哥们想了个办法——不换 OCR 引擎
直接把烂摊子甩给大语言模型
Tesseract is free and everywhere — but it
mangles characters constantly. Instead of retraining, this repo
just hands the messy output to an LLM
and says: proofread this
扫描件 → 粗稿 → LLM 精修
Scanned PDF → rough draft → LLM polish
流程很简单：扫描件丢给 Tesseract 出粗文本
然后交给大语言模型逐段修正
最后输出干净的 Markdown
不重建轮子，只在后面打补丁
The pipeline is clean: Tesseract churns out raw
text from scanned PDFs, then an LLM corrects
each chunk and spits out clean Markdown. You
don't rebuild the wheel — you patch it
★ SIGNAL 1
语义分块，塞进上下文窗口
Semantic chunking for context windows
聪明的地方在于分块策略
你没法把一整页文字塞进上下文窗口
所以它按语义把文本切开，逐块修正
最后再拼回去
这不是简单截断，是真的在理解内容边界
The clever bit is the chunking. You can't
stuff a whole page into a context window
so it breaks text semantically, fixes each piece
then stitches it back. Not a dumb split
— it respects content boundaries
★ SIGNAL 2
本地模型也能跑
Runs on local models too
不锁死在某个 API 上——本地大模型和云端接口都支持
扫合同、扫病历的场景
数据不过第三方不是锦上添花，是刚需
Docker 一键部署
63 个 commit
正经的 changelog
不是周末扔上来就跑的那种
Not locked to a single API — supports
both local LLMs and cloud endpoints. For anyone
scanning contracts or medical docs, data never leaving
your machine isn't a nice-to-have. Docker-ready, 63 commits
a proper changelog — this isn't a weekend
drop-and-forget
作者没留什么金句
但 63 个 commit 本身就是态度
能在 GitHub 上持续迭代这种工具的人
大概率是自己被 OCR 折磨得够呛
然后决定一劳永逸地解决它
The author didn't leave a manifesto, but 63
commits speak for themselves. Someone who keeps iterating
on an OCR fix like this probably got
burned enough times to finally do something about
it
AI UNDERDOGS
不换引擎，只加个校对员
Don't swap the engine — hire a proofreader
Dicklesworthstone/llm_aided_ocr
关注 · 每天发现更多 AI 神作
github.com/Dicklesworthstone/llm_aided_ocr