doc-agent / tests /test_export.py
Sanidhya Ravi
Turn the TrOCR demo into an agentic document pipeline
5eb40dc
Raw History Blame Contribute Delete
5.71 kB
"""Every artifact must reopen cleanly in the library that owns its format."""
from __future__ import annotations
import json
import zipfile
from pathlib import Path
import pytest
from PIL import Image, ImageDraw
from docagent.export import build_bundle, build_manifest, writers
from docagent.state import Block, Box, DocKind, Document, Field, Page, Table
@pytest.fixture
def doc(tmp_path) -> Document:
image = Image.new("RGB", (600, 260), "white")
draw = ImageDraw.Draw(image)
draw.text((20, 20), "ACME INVOICE 2026-114", fill="black")
draw.text((20, 60), "Widget 2 9.99", fill="black")
page1 = Page(
index=0, image=image, source="invoice.pdf", kind=DocKind.INVOICE,
kind_confidence=0.91, backend="glm_ocr",
markdown="# ACME Invoice\n\n| Item | Qty |\n| - | - |\n| Widget | 2 |\n",
)
page1.tables = [Table(header=["Item", "Qty"], rows=[["Widget", "2"]],
page_no=1, caption="Line items")]
page1.fields = [Field("Invoice number", "2026-114", 1, 0.95)]
page1.blocks = [
Block("ACME INVOICE 2026-114", Box(20, 18, 300, 18)),
Block("Widget 2 9.99", Box(20, 58, 200, 18)),
]
page2 = Page(
index=1, source="notes.docx", kind=DocKind.PROSE, kind_confidence=0.62,
backend="docx", has_text_layer=True,
markdown="## Notes\n\n- first point\n\n1. numbered\n\nPlain paragraph.",
)
document = Document(
pages=[page1, page2], sources=["invoice.pdf", "notes.docx"],
instruction="line items to a spreadsheet",
summary="Invoice 2026-114 from ACME.", workdir=str(tmp_path),
)
return document
def test_xlsx_reopens_with_expected_sheets_and_rows(doc, tmp_path):
import openpyxl
path = writers.write_xlsx(doc, tmp_path)
workbook = openpyxl.load_workbook(path)
assert "Line items" in workbook.sheetnames
assert "Fields" in workbook.sheetnames
assert "Manifest" in workbook.sheetnames
sheet = workbook["Line items"]
assert [c.value for c in sheet[1]] == ["Item", "Qty"]
assert [c.value for c in sheet[2]] == ["Widget", "2"]
def test_xlsx_sheet_names_are_excel_legal(tmp_path):
import openpyxl
document = Document(workdir=str(tmp_path))
page = Page(index=0, markdown="x", has_text_layer=True)
page.tables = [
Table(header=["a"], rows=[["1"]], page_no=1, caption="bad/name:[with]*chars?"),
Table(header=["a"], rows=[["2"]], page_no=1, caption="bad/name:[with]*chars?"),
]
document.pages = [page]
workbook = openpyxl.load_workbook(writers.write_xlsx(document, tmp_path))
for name in workbook.sheetnames:
assert len(name) <= 31
assert not set(name) & set("[]:*?/\\")
assert len(set(workbook.sheetnames)) == len(workbook.sheetnames)
def test_docx_preserves_tables_and_headings(doc, tmp_path):
import docx
document = docx.Document(writers.write_docx(doc, tmp_path))
assert len(document.tables) == 1
assert [c.text for c in document.tables[0].rows[0].cells] == ["Item", "Qty"]
text = "\n".join(p.text for p in document.paragraphs)
assert "first point" in text
# Table rows became a real Word table, so the raw pipe text is gone.
assert "| Widget |" not in text
def test_pdf_report_opens_and_has_pages(doc, tmp_path):
import pikepdf
with pikepdf.open(writers.write_pdf_report(doc, tmp_path)) as pdf:
assert len(pdf.pages) >= 1
def test_searchable_pdf_text_matches_the_ocr_text(doc, tmp_path):
import pypdfium2 as pdfium
path = writers.write_searchable_pdf(doc, tmp_path)
pdf = pdfium.PdfDocument(path)
try:
textpage = pdf[0].get_textpage()
extracted = textpage.get_text_range() or ""
textpage.close()
finally:
pdf.close()
assert "ACME INVOICE 2026-114" in extracted
assert "Widget" in extracted
def test_csv_written_per_table(doc, tmp_path):
paths = writers.write_csv(doc, tmp_path)
assert len(paths) == 1
assert Path(paths[0]).read_text(encoding="utf-8-sig").splitlines()[0] == "Item,Qty"
def test_fields_json_contains_only_fields(doc, tmp_path):
payload = json.loads(Path(writers.write_fields_json(doc, tmp_path)).read_text("utf-8"))
assert [f["key"] for f in payload["fields"]] == ["Invoice number"]
assert "pages" not in payload
def test_bundle_contains_every_artifact_plus_manifest(doc, tmp_path):
zip_path = build_bundle(doc, tmp_path)
with zipfile.ZipFile(zip_path) as archive:
names = set(archive.namelist())
assert {"document.md", "document.json", "manifest.json", "tables.xlsx",
"summary.pdf", "searchable.pdf", "document.docx"} <= names
assert not doc.warnings
def test_manifest_flags_the_low_confidence_page(doc):
manifest = build_manifest(doc)
assert manifest["needs_review"] is True
assert any("page 2" in c for c in manifest["concerns"])
assert manifest["min_page_confidence"] == pytest.approx(0.62)
def test_manifest_is_clean_when_everything_is_confident(doc):
doc.pages[1].kind_confidence = 0.95
manifest = build_manifest(doc)
assert manifest["needs_review"] is False
assert manifest["concerns"] == []
def test_a_failing_writer_does_not_abort_the_bundle(doc, tmp_path, monkeypatch):
def boom(*args, **kwargs):
raise RuntimeError("simulated failure")
monkeypatch.setattr(writers, "write_xlsx", boom)
zip_path = build_bundle(doc, tmp_path)
with zipfile.ZipFile(zip_path) as archive:
names = set(archive.namelist())
assert "tables.xlsx" not in names
assert "document.md" in names # the rest still got written
assert any("could not write xlsx" in w for w in doc.warnings)