from __future__ import annotations import json from pathlib import Path import pytest from ocrtech import pipeline from ocrtech.benchmark import _quality_metrics from ocrtech.engines import EngineOutput from ocrtech.pipeline import parse_document from ocrtech.schemas import BBox, Figure, Page, Table, TableCell, TextLine def test_parse_doc_exports_json_markdown_tables_and_figures(tmp_path: Path) -> None: source = tmp_path / "page.png" source.write_bytes(b"fake image bytes") sidecar = tmp_path / "page.ocr.json" sidecar.write_text( json.dumps( { "pages": [ { "width": 800, "height": 1000, "lines": [ {"text": "नेपाल सरकार", "bbox": [10, 10, 200, 30], "confidence": 0.99}, {"text": "नाम उमेर जिल्ला", "bbox": [10, 80, 400, 20], "confidence": 0.95}, {"text": "राम ३२ काठमाडौं", "bbox": [10, 110, 420, 20], "confidence": 0.94}, {"text": "Sita 29 Lalitpur", "bbox": [10, 140, 420, 20], "confidence": 0.93}, {"text": "This is English text.", "bbox": [10, 220, 300, 20], "confidence": 0.98}, ], } ], "figures": [ { "figure_id": "fig-0", "page_index": 0, "bbox": [100, 300, 200, 120], "caption": "चित्र १", "summary": "A figure placeholder.", "confidence": 0.8, } ], }, ensure_ascii=False, ), encoding="utf-8", ) document = parse_document(source, tmp_path / "out", engine_name="sidecar") assert len(document.pages) == 1 assert len(document.tables) == 1 assert len(document.figures) == 1 assert (tmp_path / "out" / "document.json").exists() assert (tmp_path / "out" / "document.md").exists() assert (tmp_path / "out" / "tables" / "p0-t0.csv").exists() assert (tmp_path / "out" / "tables" / "p0-t0.html").exists() assert (tmp_path / "out" / "figures" / "metadata.json").exists() assert (tmp_path / "out" / "quality.json").exists() exported = json.loads((tmp_path / "out" / "document.json").read_text(encoding="utf-8")) assert exported["tables"][0]["csv_path"] == str(tmp_path / "out" / "tables" / "p0-t0.csv") assert exported["tables"][0]["html_path"] == str(tmp_path / "out" / "tables" / "p0-t0.html") assert exported["tables"][0]["html"].startswith("") markdown = (tmp_path / "out" / "document.md").read_text(encoding="utf-8") assert "# नेपाल सरकार" in markdown assert "| नाम | उमेर | जिल्ला |" in markdown assert "This is English text." in markdown assert document.metadata["quality"]["passed"] def test_export_document_crops_figure_from_source_image_when_image_path_missing(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: source = tmp_path / "page.png" try: from PIL import Image except ImportError: pytest.skip("Pillow is not installed") Image.new("RGB", (120, 100), "white").save(source) class FigureEngine: def recognize(self, input_path: Path) -> EngineOutput: line = TextLine(text="[Figure: chart]", bbox=BBox(10, 10, 60, 20), confidence=0.9, page_index=0) figure = Figure(figure_id="fig-0", page_index=0, bbox=BBox(20, 30, 40, 35), summary="chart") return EngineOutput(pages=[Page(page_index=0, text_lines=[line])], figures=[figure], metadata={"engine": "figure-engine"}) monkeypatch.setattr(pipeline, "create_engine", lambda name, **kwargs: FigureEngine()) document = parse_document(source, tmp_path / "out", engine_name="sidecar") cropped = tmp_path / "out" / "figures" / "fig-0.png" assert cropped.exists() assert document.figures[0].image_path == str(cropped) assert document.figures[0].metadata["export_source"] == "source_crop" with Image.open(cropped) as image: assert image.size == (40, 35) metadata = json.loads((tmp_path / "out" / "figures" / "metadata.json").read_text(encoding="utf-8")) assert metadata[0]["image_path"] == str(cropped) def test_export_document_records_warning_when_figure_image_cannot_be_created(tmp_path: Path) -> None: source = tmp_path / "page.pdf" source.write_bytes(b"%PDF-1.4\n") sidecar = tmp_path / "page.ocr.json" sidecar.write_text( json.dumps( { "pages": [{"lines": [{"text": "[Figure: chart]", "bbox": [10, 10, 80, 20], "confidence": 0.9}]}], "figures": [{"figure_id": "fig-0", "page_index": 0, "bbox": [20, 30, 40, 35], "summary": "chart"}], } ), encoding="utf-8", ) document = parse_document(source, tmp_path / "out", engine_name="sidecar") assert not (tmp_path / "out" / "figures" / "fig-0.png").exists() warnings = document.figures[0].metadata["export_warnings"] assert warnings == ["source image format cannot be cropped: .pdf"] metadata = json.loads((tmp_path / "out" / "figures" / "metadata.json").read_text(encoding="utf-8")) assert metadata[0]["metadata"]["export_warnings"] == warnings def test_export_document_copies_relative_figure_image_from_page_bundle(tmp_path: Path) -> None: source = tmp_path / "bundle" source.mkdir() figure_image = source / "embedded-figure.png" figure_image.write_bytes(b"figure bytes") (source / "document.ocr.json").write_text( json.dumps( { "pages": [{"lines": [{"text": "[Figure: chart]", "bbox": [10, 10, 80, 20], "confidence": 0.9}]}], "figures": [ { "figure_id": "fig-0", "page_index": 0, "bbox": [20, 30, 40, 35], "image_path": "embedded-figure.png", } ], } ), encoding="utf-8", ) document = parse_document(source, tmp_path / "out", engine_name="sidecar") copied = tmp_path / "out" / "figures" / "embedded-figure.png" assert copied.read_bytes() == b"figure bytes" assert document.figures[0].image_path == str(copied) def test_export_document_crops_figure_from_page_bundle(tmp_path: Path) -> None: source = tmp_path / "bundle" source.mkdir() try: from PIL import Image except ImportError: pytest.skip("Pillow is not installed") Image.new("RGB", (120, 100), "white").save(source / "page-0000.png") Image.new("RGB", (160, 140), "white").save(source / "page-0001.png") (source / "document.ocr.json").write_text( json.dumps( { "pages": [ {"lines": [{"text": "page one", "bbox": [0, 0, 80, 20], "confidence": 0.9}]}, {"lines": [{"text": "[Figure: chart]", "bbox": [10, 10, 80, 20], "confidence": 0.9}]}, ], "figures": [{"figure_id": "fig-1", "page_index": 1, "bbox": [30, 40, 50, 45], "summary": "chart"}], } ), encoding="utf-8", ) document = parse_document(source, tmp_path / "out", engine_name="sidecar") cropped = tmp_path / "out" / "figures" / "fig-1.png" assert cropped.exists() assert document.figures[0].image_path == str(cropped) assert document.figures[0].metadata["export_source"] == "source_crop" with Image.open(cropped) as image: assert image.size == (50, 45) def test_export_document_records_warning_when_page_bundle_has_no_images(tmp_path: Path) -> None: source = tmp_path / "bundle" source.mkdir() (source / "document.ocr.json").write_text( json.dumps( { "pages": [{"lines": [{"text": "[Figure: chart]", "bbox": [10, 10, 80, 20], "confidence": 0.9}]}], "figures": [{"figure_id": "fig-0", "page_index": 0, "bbox": [20, 30, 40, 35], "summary": "chart"}], } ), encoding="utf-8", ) document = parse_document(source, tmp_path / "out", engine_name="sidecar") warnings = document.figures[0].metadata["export_warnings"] assert warnings == [f"page bundle has no croppable image files: {source}"] def test_fallback_triggers_on_low_confidence(tmp_path: Path) -> None: source = tmp_path / "page.png" source.write_bytes(b"fake image bytes") sidecar = tmp_path / "page.ocr.json" sidecar.write_text( json.dumps( { "pages": [ { "lines": [ {"text": "low confidence", "bbox": [0, 0, 100, 20], "confidence": 0.2}, ] } ] } ), encoding="utf-8", ) document = parse_document(source, tmp_path / "out", engine_name="sidecar", fallback_engine="paddleocr") assert "fallback_error" in document.metadata assert document.metadata["fallback"]["triggered"] is True assert document.metadata["fallback"]["outcome"] == "error" assert document.metadata["fallback"]["reason"] == "average_confidence_below_threshold" assert document.pages[0].blocks[0].text == "low confidence" def test_fallback_success_records_decision_and_benchmark_metrics(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: source = tmp_path / "page.png" source.write_bytes(b"fake image bytes") class PrimaryEngine: def recognize(self, input_path: Path) -> EngineOutput: line = TextLine(text="primary text", bbox=BBox(0, 0, 100, 20), confidence=0.2, page_index=0, line_id="p0-l0") return EngineOutput(pages=[Page(page_index=0, text_lines=[line])], metadata={"engine": "primary"}) class FallbackEngine: def recognize(self, input_path: Path) -> EngineOutput: line = TextLine(text="fallback text", bbox=BBox(0, 0, 120, 20), confidence=0.95, page_index=0, line_id="p0-l0") return EngineOutput(pages=[Page(page_index=0, text_lines=[line])], metadata={"engine": "fallback"}) def fake_create_engine(name: str, **kwargs: object) -> object: if name == "sidecar": return PrimaryEngine() if name == "paddleocr": return FallbackEngine() raise AssertionError(f"unexpected engine request: {name}, kwargs={kwargs}") monkeypatch.setattr(pipeline, "create_engine", fake_create_engine) document = parse_document(source, tmp_path / "out", engine_name="sidecar", fallback_engine="paddleocr") metrics = _quality_metrics(document) assert document.pages[0].blocks[0].text == "fallback text" assert document.metadata["fallback"] == { "configured": True, "triggered": True, "outcome": "success", "primary_engine": "sidecar", "fallback_engine": "paddleocr", "fallback_model_config": None, "primary_average_confidence": 0.2, "threshold": 0.8, "reason": "average_confidence_below_threshold", } assert metrics["fallback_triggered"] == 1.0 assert metrics["fallback_succeeded"] == 1.0 assert metrics["fallback_failed"] == 0.0 assert metrics["fallback_primary_average_confidence"] == 0.2 def test_fallback_not_triggered_records_reason(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: source = tmp_path / "page.png" source.write_bytes(b"fake image bytes") class PrimaryEngine: def recognize(self, input_path: Path) -> EngineOutput: line = TextLine(text="primary text", bbox=BBox(0, 0, 100, 20), confidence=0.99, page_index=0, line_id="p0-l0") return EngineOutput(pages=[Page(page_index=0, text_lines=[line])], metadata={"engine": "primary"}) def fake_create_engine(name: str, **kwargs: object) -> object: if name == "sidecar": return PrimaryEngine() raise AssertionError(f"fallback should not be constructed: {name}, kwargs={kwargs}") monkeypatch.setattr(pipeline, "create_engine", fake_create_engine) document = parse_document(source, tmp_path / "out", engine_name="sidecar", fallback_engine="paddleocr") metrics = _quality_metrics(document) assert document.metadata["fallback"]["triggered"] is False assert document.metadata["fallback"]["outcome"] == "not_triggered" assert document.metadata["fallback"]["reason"] == "confidence_above_threshold" assert metrics["fallback_triggered"] == 0.0 def test_fallback_model_config_uses_candidate_engine(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: source = tmp_path / "page.png" source.write_bytes(b"fake image bytes") fallback_model = tmp_path / "fallback-model.json" fallback_model.write_text("{}", encoding="utf-8") calls: list[tuple[str, dict[str, object]]] = [] class PrimaryEngine: def recognize(self, input_path: Path) -> EngineOutput: line = TextLine(text="primary text", bbox=BBox(0, 0, 100, 20), confidence=0.2, page_index=0, line_id="p0-l0") return EngineOutput(pages=[Page(page_index=0, text_lines=[line])], metadata={"engine": "primary"}) class FallbackEngine: def recognize(self, input_path: Path) -> EngineOutput: line = TextLine(text="model fallback", bbox=BBox(0, 0, 140, 20), confidence=0.96, page_index=0, line_id="p0-l0") return EngineOutput(pages=[Page(page_index=0, text_lines=[line])], metadata={"engine": "fallback"}) def fake_create_engine(name: str, **kwargs: object) -> object: calls.append((name, kwargs)) if name == "sidecar": return PrimaryEngine() if name == "candidate": return FallbackEngine() raise AssertionError(f"unexpected engine request: {name}, kwargs={kwargs}") monkeypatch.setattr(pipeline, "create_engine", fake_create_engine) document = parse_document(source, tmp_path / "out", engine_name="sidecar", fallback_model_config=fallback_model) assert calls == [("sidecar", {}), ("candidate", {"model_config": fallback_model})] assert document.pages[0].blocks[0].text == "model fallback" assert document.metadata["fallback"]["fallback_engine"] == "candidate" assert document.metadata["fallback"]["fallback_model_config"] == str(fallback_model) assert document.metadata["fallback"]["outcome"] == "success" def test_build_document_does_not_duplicate_engine_supplied_table(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: source = tmp_path / "page.png" source.write_bytes(b"fake image bytes") class TableEngine: def recognize(self, input_path: Path) -> EngineOutput: table_bbox = BBox(10, 20, 300, 50) table = Table( table_id="p0-t0", page_index=0, bbox=table_bbox, cells=[ TableCell(row=0, col=0, text="नाम"), TableCell(row=0, col=1, text="रकम"), TableCell(row=0, col=2, text="Remarks"), TableCell(row=1, col=0, text="किरण"), TableCell(row=1, col=1, text="२१००"), TableCell(row=1, col=2, text="Paid"), ], ) lines = [ TextLine(text="नाम रकम Remarks", bbox=BBox(10, 20, 300, 20), confidence=0.9, page_index=0), TextLine(text="किरण २१०० Paid", bbox=BBox(10, 50, 300, 20), confidence=0.9, page_index=0), ] return EngineOutput(pages=[Page(page_index=0, text_lines=lines)], tables=[table], metadata={"engine": "table-engine"}) monkeypatch.setattr(pipeline, "create_engine", lambda name, **kwargs: TableEngine()) document = parse_document(source, tmp_path / "out", engine_name="sidecar") assert len(document.tables) == 1 assert document.pages[0].blocks[0].table_id == "p0-t0" assert (tmp_path / "out" / "tables" / "p0-t0.csv").exists() def test_figure_marker_uses_nearby_ocr_damaged_caption(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: source = tmp_path / "page.png" source.write_bytes(b"fake image bytes") class FigureEngine: def recognize(self, input_path: Path) -> EngineOutput: lines = [ TextLine(text="वार्षिक प्रतिवेदन", bbox=BBox(10, 10, 200, 30), confidence=0.9, page_index=0), TextLine(text="[Figure: revenue chart]", bbox=BBox(10, 60, 260, 30), confidence=0.9, page_index=0), TextLine(text="चतुर्थ १: Revenue growth from 2021 to 2025.", bbox=BBox(10, 100, 500, 30), confidence=0.9, page_index=0), TextLine(text="निष्कर्ष: अन्तिम दुई वर्षमा वृद्धि भएको छ।", bbox=BBox(10, 150, 500, 30), confidence=0.9, page_index=0), ] return EngineOutput(pages=[Page(page_index=0, text_lines=lines)], metadata={"engine": "figure-engine"}) monkeypatch.setattr(pipeline, "create_engine", lambda name, **kwargs: FigureEngine()) document = parse_document(source, tmp_path / "out", engine_name="sidecar") assert len(document.figures) == 1 assert document.figures[0].summary == "[Figure: revenue chart]" assert document.figures[0].caption == "चतुर्थ १: Revenue growth from 2021 to 2025." assert document.pages[0].blocks[1].text == "चतुर्थ १: Revenue growth from 2021 to 2025." assert "निष्कर्ष" in document.pages[0].blocks[2].text def test_fallback_records_unexpected_runtime_failure(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None: source = tmp_path / "page.png" source.write_bytes(b"fake image bytes") class PrimaryEngine: def recognize(self, input_path: Path) -> EngineOutput: line = TextLine(text="primary text", bbox=BBox(0, 0, 100, 20), confidence=0.2, page_index=0, line_id="p0-l0") return EngineOutput(pages=[Page(page_index=0, text_lines=[line])], metadata={"engine": "primary"}) def fake_create_engine(name: str, **kwargs: object) -> object: if name == "sidecar": return PrimaryEngine() if name == "paddleocr": raise RuntimeError("paddle runtime missing") raise AssertionError(f"unexpected engine request: {name}, kwargs={kwargs}") monkeypatch.setattr(pipeline, "create_engine", fake_create_engine) document = parse_document(source, tmp_path / "out", engine_name="sidecar", fallback_engine="paddleocr") assert document.metadata["fallback_error"] == "RuntimeError: paddle runtime missing" assert document.pages[0].blocks[0].text == "primary text"