LandingAI ADE:一次解析长重复表格并让行合计与页眉总额对齐
LandingAI 介绍 Agentic Document Extraction:整页读取长表并抽出每行字段,用与印刷合计一致作为完整性校验,示例租金卷为 184 个字段。
LandingAI 介绍 Agentic Document Extraction(ADE)在长重复表格上的表现:传统抽取常在近乎相同的数百行里漏行,ADE 声称一次读完整表并抽出各行字段,再用「抽取行合计是否等于页上印刷合计」做完整性核对。示例为一份租金卷(rent roll):单页 25 个单元共 184 个字段,返回合计与页面一致——23,760 平方英尺与月租 22,280 美元。官方称可在评论区 GitHub 链接自行跑样本。
对研发与文档智能体,这把「抽字段」推进到「可对账」。生产前仍需在扫描件、多页跨表与非英文数字格式上复测,并确认合计校验失败时的人工流程。数字来自厂商示例,不保证泛化到所有版式。
接入建议是:校验失败即阻断自动入账,只把通过对账的行写入下游;对金额与面积字段保留原文截图链接,便于审计抽查。
房地产与财务运营里的租金卷、清单表最怕静默漏行。把「合计对账」做成硬门禁后,ADE 更像质检工序而不是一次性 OCR。多页跨页合计与手写批注仍是已知难点,需样本回归。
ADE extracts long repeating tables in one pass with reconciled totals
Agentic Document Extraction (ADE) extracts long repeating tables in one pass, and the totals still reconcile.
Long tables are where extraction usually breaks. Hundreds of near-identical rows, and a single dropped line is easy to miss.
ADE reads the full table at once. It parses the page and extracts every row with its fields, however long the table runs.
The reconciliation is the check. When the extracted rows sum to the printed totals, you know the table came through complete.
On a sample rent roll, that was 184 fields off one page across 25 units. The returned totals matched the page: 23,760 square feet and $22,280 in monthly rent.
Run the sample yourself. GitHub link in the comments.