document-intelligence

name	document-intelligence
version	0.2.0
description	Turn unstructured real estate documents into typed, verifiable data. Classifies uploads (utility bill, energy audit, lease, PCA, drawing, regulatory filing), extracts structured fields per document type, digitizes charts/graphs/tables from PDFs into JSON, writes results to the asset record, and RAG-indexes every document so it stays searchable. Triggers on: "extract from this bill", "parse this audit", "what does the lease say about", "pull the data out of this PCA", "digitize this chart/table", "process these documents", "batch ingest", any document upload that should become structured data.

Document Intelligence

One pipeline for every document that enters a portfolio: classify → extract → write → index. The goal is typed data with provenance, not prose summaries.

Pipeline

1. Classify

Read the document (you have native PDF/image vision). Emit: { type, confidence, language, pages, scanned: bool }

Types: utility_bill | energy_audit | lease | pca | architectural_drawing | regulatory_filing | om | rent_roll | other. If confidence < 0.8, say what it looks like and ask rather than guessing. Multi-document PDFs (e.g. a year of bills in one file): split logically and process each.

2. Extract typed fields

Use the schema for the detected type in references/extraction-schemas.md. Rules:

• Every extracted value carries { value, unit, page, verbatim_source_text } — the page reference is what makes downstream verification possible.
• Never infer a value that isn't in the document. Missing = null + listed in gaps.