CLOUDCRUISE UNITED

LLM_VISION only

ExtractDatamodel — PDF / Canvas Viewer

Claims data is rendered on a <canvas> element, simulating a PDF viewer or canvas-based spreadsheet. There are no DOM text nodes — STATIC XPath finds nothing and LLM_DOM sees only a blank canvas tag. Only LLM_VISION (screenshot) can read the rendered content.

Ground truth (hidden from DOM — only in canvas pixels)

Claim ID	Patient	Amount	Status	Date
CLM-90210	Martinez, Rosa	$1,245.00	Approved	04/15/2026
CLM-90211	Kim, David	$892.50	Pending	04/16/2026
CLM-90212	Brown, Alice	$3,100.00	Denied	04/17/2026
CLM-90213	Singh, Raj	$567.25	Approved	04/18/2026

Why only LLM_VISION works

Canvas renders pixels — the DOM contains only <canvas> with zero text nodes
STATIC XPath: no text to extract, returns null for every field
LLM_DOM: serialized HTML just shows <canvas data-testid="claims-canvas"></canvas>
LLM_VISION: screenshot shows the full rendered table — LLM reads it like a human would
Real-world equivalent: PDF.js viewer, canvas-based EHR grids, chart widgets with labels

Workflow node config

{
  "action": "EXTRACT_DATAMODEL",
  "parameters": {
    "execution": "LLM_VISION",
    "extract_data_model": {
      "type": "object",
      "properties": {
        "claims": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "claim_id": { "type": "string" },
              "patient":  { "type": "string" },
              "amount":   { "type": "string" },
              "status":   { "type": "string" },
              "date":     { "type": "string" }
            }
          }
        }
      }
    }
  }
}