reddit · SideProject · 10-23 14:50 56 分 ← 返回列表

made a thing that highlights the actual pdf/excel source for every claim claude makes

There's been some personal tools i've been building that i think should exist natively - this is one of them. I use claude a lot sitting with spreadsheets and larger pdfs (family business docs mostly) and kept catching it "quoting" stuff that wasn't quite there. so i built claim. after every answer it checks what claude said against the files it actually read and shows you the page with the quote highlighted. green = found, amber = partial, red = not in the file. the checking doesn't use t

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：AI（如 Claude）引用 PDF/Excel 时会“编造”看似存在但实际不在文件里的原文和数字，用户核对每一条论断来源的成本极高。

目标用户：需要基于 PDF/Excel 做财务、法务、家庭生意文书分析，并核验 AI 输出的分析师、会计、律师和中小企业主。

证据：kept catching it "quoting" stuff that wasn't quite there... numbers that aren't in any file get marked "worked out" so you know what to double check. the checking doesn't use the model at all. it matches exact runs of text from the file and checks every number, units too, so $22,270M matches 22,270 on a page that says "in millions".

变现：开源核心免费引流，商业化做企业/专业版订阅：团队审计日志、OCR 与 Word/PPT 支持、合规报告一键导出、云端校验 API 按次或按席位收费。

MVP：在现有开源版上加商业化外壳：补 OCR 与 Word/PPT 解析、做“论断—来源”审计报告一键导出（PDF/CSV）、支持团队共享校验记录，桌面端按席位订阅。

竞品：Claude 原生 citations、Google NotebookLM、Perplexity 引用溯源、Vectara 等 RAG grounding 校验工具

摘要：作者用 Claude 处理电子表格和家庭生意 PDF 时，频繁发现它在“引用”并不存在的原文和数字，于是做了一个 MIT 开源工具 claim：对每个回答逐字比对源文件，绿色=找到、黄色=部分匹配、红色=文件里没有，并校验数字与单位（如 $22,270M 匹配 “in millions” 下的 22,270）。目前免费、本地运行、兼容任意 MCP 客户端，尚缺 OCR 和 Word/PPT 支持，图片卡片仅在 ghostty/kitty/wezterm 显示。这揭示了 AI 文档输出的可核验痛点，但帖子本身无明确付费证据。