LandingAI ADE 增加 Build Schema:用自然语言或解析后的 Markdown 自动生成文档抽取 JSON Schema
LandingAI 为 Agentic Document Extraction 推出 Build Schema:在 Parse 与 Extract 之间,根据文档 Markdown 或纯语言描述生成可编辑的结构化 schema,减轻按版式手写维护成本。
LandingAI 指出,Agentic Document Extraction(ADE)的抽取步骤依赖 JSON Schema 定义字段;按文档类型手写并维护 schema 成本高,供应商版式一变就要再改一版。
新能力 Build Schema 插在 Parse 与 Extract 之间:Parse 把文档转为 Markdown 后,Build Schema 阅读该 Markdown,或直接根据纯语言描述生成 schema。在 playground 中上传发票并描述需要供应商、到期日、总额、行项目等,即可得到含嵌套字段的结构化 schema,可在线改字段、别名与描述;再要“税额”会生成新版本并保留旧版。
文中归纳三类用法:统一多供应商版式到一套 schema、为新文档类型 refinement、仅凭提示构建 schema。目标是让 schema 从“手写维护物”变成“可描述生成物”,从而把人力从字段工程挪到审核与业务规则。
对财务共享中心、物流单证与合同数字化团队,这降低了“每来一个新模板就改抽取配置”的节奏。原帖含 kw:crypto 提示标记,内容为文档智能产品能力,与虚拟货币营销无关,故采用。配套演示见官方视频链接。
落地时仍建议人工核对生成字段的类型、必填与枚举,并把通过验证的 schema 纳入版本库;对多语言单据可先统一 Parse 输出再生成 schema,避免把 OCR 噪声直接写进字段定义。
LandingAI Build Schema for Agentic Document Extraction
Writing a JSON schema for every document type is a job in itself.
Extraction in Agentic Document Extraction (ADE) runs on a JSON schema. It defines the fields you want pulled out of a document.
Writing those schemas by hand is slow. Maintaining them is worse, because the moment a vendor changes a layout, the schema needs another pass.
Build Schema takes that off you.
It sits between Parse and Extract. Parse turns the document into Markdown. Build Schema reads that Markdown, or just a plain-language description, and writes the schema for you.
In the playground, you upload an invoice and describe what you want: vendor, due date, total, line items. It generates a structured schema, nested fields and all. You review it, edit fields, add aliases, and tighten descriptions inline.
Need one more field? Prompt for tax amount and it spins a new schema version, with your previous one kept.
It is built for three real patterns:
→ unify many vendor layouts into one schema → refine a schema for a new document type → build a schema from a prompt alone
The schema stops being something you write and maintain. It becomes something you describe.
Watch the full YT video here: https://t.co/DHfP7RSsP7