Live demo

Documents: classify, extract, correct

Drop a rental agreement, an inspection report, or an invoice. @de_canter/apogee-documents-ai classifies it against the three types the rental company declared, extracts the fields from that type's Zod schema, and returns every claim as a kernel Assertion with its own confidence. Correct a value and the human assertion supersedes the model's; the audit log keeps the correction as ground truth for evals.

or a PNG, JPEG, or PDF:

Domain

src/demo/rental/documents/taxonomy.ts: defineTaxonomy with three types and a defineExtraction per type. Nothing in the package knows what an invoice is.

Pipeline

processDocument runs classify then extract as a checkpointed pipeline; a failed stage keeps its error and can resume. Images and PDFs go to the model as media blocks; text as a document block.

Corrections

correctField confirms a human assertion at confidence 1 and supersedes the model's; the audit records who and when, and groundTruth turns corrected runs into eval cases.