← Back to the case

Inspectable walkthrough

Turn a Visual Document Into Agent-Readable Work Material

One PDF is preserved as original page renders, flat text extraction, and a document map that restores reading order and relationships.

  • Status: runnable demo using a synthetic PDF
  • Product: Agent Computer · Document Ingester
  • Input: four-page AI customer-support pilot review

Why Text Extraction Was Not Enough

Meaning in a visual document does not live in words alone. Column order, arrow direction, the table row attached to a callout, and visual hierarchy all carry information. Flat extraction preserves text while collapsing those relationships.

Document Ingester does not try to create one perfect replacement. It separates a fast map for understanding from the original page used for exact verification, then links them.

The Source Page

Synthetic PDF page with a three-column flow and an exception path
Synthetic PDF page with a three-column flow and an exception path

The page shows Input → Agent Work → Human Review, an exception path that bypasses the normal draft flow, and a No auto-send boundary attached to human review.

What Flat Extraction Lost

The text remains:

01 INPUT Collect the request and evidence
02 AGENT WORK Classify, retrieve, and draft
03 HUMAN REVIEW Decide the next action
No evidence or high risk Draft held Human review
No auto-send

The extraction no longer tells us:

  • that the three columns progress left to right
  • which bullets belong to each stage
  • that the exception path starts in processing and ends in human review
  • that No auto-send is the operating boundary of the review stage

The Document Map Restores Relationships

The map is not a prettier text export. It states the structure an agent should understand first.

Page 2 · How one request becomes a response candidate

Primary flow
1. INPUT: customer request + account state + approved policy
2. AGENT WORK: classify + retrieve evidence + draft
3. HUMAN REVIEW: approve / revise / escalate

Exception
No evidence or high risk
→ hold response candidate
→ HUMAN REVIEW

Boundary
Never send automatically

Page text supports search and quotation. The document map supports hierarchy and relationship understanding. The original render remains available for exact wording and visual verification.

How the Three Layers Work Together

QuestionRead firstVerify against
What decision does the document make?document maporiginal pages 1 and 4
Which items belong to a stage?page note and maporiginal page
What is the exact number or phrase?extracted textoriginal render
What does an arrow or callout connect?maporiginal render
Synthetic PDF page with 120 records and linked callouts
Synthetic PDF page with 120 records and linked callouts

Demo Boundary

The company, customer requests, and 120 records in the input PDF are synthetic. The actual transformation outputs are the page renders, extracted text, page notes, and document map. This demo does not claim perfect relationship recovery for every PDF.

Back to the Agent Computer case

Back to the caseAgent Computer