Citations

Every value ADX returns traces back to a specific location in the source document. Citations are the provenance layer that lets agents (and humans) verify extracted data.

Citation Types

Type	Source	Fields
page	PDF text block	page, bbox, text
table	PDF table cell	page, bbox, table_id, row, col
cell	Spreadsheet cell	sheet, cell, text
range	Spreadsheet range	sheet, range, text
formula	Spreadsheet formula	sheet, cell, formula

Where Citations Appear

In Text Blocks

Every text block extracted from a PDF includes a citation:

{
  "text": "Payment is due within 30 days.",
  "type": "paragraph",
  "citation": {
    "type": "page",
    "page": 2,
    "bbox": [72, 450, 540, 470]
  }
}

In Extracted Fields

Every field from extract() includes a citation:

{
  "name": "total",
  "value": 4250.00,
  "citation": {
    "type": "page",
    "page": 3,
    "bbox": [72, 680, 540, 700],
    "text": "Total: $4,250.00"
  }
}

In Search Results

Search matches include location information:

{
  "query": "payment terms",
  "matches": [
    {
      "page": 2,
      "snippet": "...Payment is due within 30 days of invoice date...",
      "bbox": [72, 450, 540, 470]
    }
  ]
}

In Formula Inspection

Formula references are cited back to their source cells:

{
  "formula": "=SUM(B2:B9)",
  "computed_value": 125000,
  "references": [
    {"type": "range", "sheet": "Revenue", "range": "B2:B9"}
  ]
}

Retrieving Citations

Use the citations endpoint to get all citations for a document:

curl http://localhost:8000/v1/files/{id}/citations

Or access citations through the Python SDK:

graph = dn.get_graph(doc_id)
for citation in graph.citations:
    print(f"  [{citation.type}] page={citation.page} text={citation.text[:50]}")

Why Citations Matter

• Verifiability — agents can prove where data came from
• Auditability — humans can spot-check extracted values
• Confidence — low-confidence extractions can be traced and reviewed
• Debugging — when extraction is wrong, citations show what the parser actually saw