金融、法律和醫療常把重要資料放在 PDF、掃描檔案和試算表裡,AI 讀錯表格中的一格,就可能讓後續分析出錯。
- 7 人新創公司 Unsiloed AI 募得 $7.5M 種子資金,要讓 AI 助理更可靠地讀取這些檔案:它用雙路視覺模型同時辨認文字內容和版面,再把更多運算用在圖表、數字等容易看錯的地方。
- 讀出的每個資料欄位也會連迴文件中的位置,系統還能在客戶自己的伺服器上隔離網路執行,方便查核資料來源並保護檔案。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
A great launch for the document layer of AI, built for finance, law and healthcare, where the record still lives in PDFs, scans and spreadsheets.
Unsiloed AI, a 7-person startup, raised $7.5M in a Felicis-led seed to make documents queryable for AI agents.
Finance, law and healthcare run on PDFs and scans, which LLMs read probabilistically, so a single misread table cell can poison downstream analysis.
Unsiloed’s dual-stream vision models read content and layout in parallel, while heat-map chunking spends extra compute on figures, numbers and charts.
Each extracted field links back to its spot on the page, and the system can run air-gapped on a customer’s own servers.