sigdetect 0.3.1__tar.gz → 0.5.0__tar.gz

{sigdetect-0.3.1 → sigdetect-0.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sigdetect
-Version: 0.3.1
+Version: 0.5.0
 Summary: Signature detection and role attribution for PDFs
 Author-email: BT Asmamaw <basmamaw@angeiongroup.com>
 License: MIT
@@ -10,9 +10,11 @@ Requires-Dist: pypdf>=4.0.0
 Requires-Dist: rich>=13.0
 Requires-Dist: typer>=0.12
 Requires-Dist: pydantic>=2.5
+Requires-Dist: pillow>=10.0
+Requires-Dist: python-docx>=1.1.0
+Requires-Dist: pytesseract>=0.3.10
+Requires-Dist: pymupdf>=1.23
 Requires-Dist: pyyaml>=6.0
-Provides-Extra: pymupdf
-Requires-Dist: pymupdf>=1.23; extra == "pymupdf"
 
 # CaseWorks.Automation.CaseDocumentIntake
 
@@ -95,14 +97,16 @@ sigdetect detect \
 ### Notes
 
 - The config file controls `pdf_root`, `out_dir`, `engine`, `pseudo_signatures`, `recurse_xobjects`, etc.
-- `--engine` supports **pypdf2** (default); a **pymupdf** engine placeholder exists and may be included in a future build.  
+- Engine selection is forced to **auto** (prefers PyMuPDF for geometry, falls back to PyPDF2); any configured `engine` value is overridden.  
 - `--pseudo-signatures` enables a vendor/Acro-only pseudo-signature when no actual `/Widget` is present (useful for DocuSign / Acrobat Sign receipts).  
 - `--recurse-xobjects` allows scanning Form XObjects for vendor markers and labels embedded in page resources. 
 - `--profile` selects tuned role logic:
   - `hipaa` → patient / representative / attorney
   - `retainer` → client / firm (prefers detecting two signatures)
 - `--recursive/--no-recursive` toggles whether `sigdetect detect` descends into subdirectories when hunting for PDFs (recursive by default).
-- `--crop-signatures` enables PNG crops for each detected widget (requires installing the optional `pymupdf` dependency). Use `--crop-dir` to override the destination and `--crop-dpi` to choose rendering quality.
+- Results output is disabled by default; set `write_results: true` or pass `--write-results` when you need `results.json` (for EDA).
+- Cropping (`--crop-signatures`) writes a one-image `.docx` per signature in the crop output directory (no PNG files on disk); `--crop-bytes` embeds base64 PNG data in `signatures[].crop_bytes` for in-memory use. PyMuPDF is required for crops, and `python-docx` is required for `.docx` output.
+- Wet detection runs automatically for non-e-sign PDFs when dependencies are available; missing OCR dependencies add a `ManualReview:*` hint instead of failing. PyMuPDF + Tesseract are required for wet detection.
 - If the executable is not on `PATH`, you can always fall back to `python -m sigdetect.cli ...`.
 
 ### EDA (quick aggregate stats)
@@ -113,6 +117,8 @@ sigdetect eda \
 
 ~~~
 
+`sigdetect eda` expects `results.json`; enable `write_results: true` when running detect.
+
 ---
 
 ## Library usage
@@ -136,15 +142,13 @@ result = detector.Detect(Path("/path/to/pdfs/example.pdf"))
 print(result.to_dict())
 ~~~
 
-`Detect(Path)` returns a **FileResult** dataclass; call `.to_dict()` for the JSON-friendly representation (see [Result schema](#result-schema)). Each signature entry now exposes `bounding_box` coordinates (PDF points, origin bottom-left). When PNG cropping is enabled, `crop_path` points at the generated image.
+`Detect(Path)` returns a **FileResult** dataclass; call `.to_dict()` for the JSON-friendly representation (see [Result schema](#result-schema)). Each signature entry now exposes `bounding_box` coordinates (PDF points, origin bottom-left). When cropping is enabled, `crop_path` points at the generated `.docx`. Use `Engine="auto"` if you want the single-pass defaults that prefer PyMuPDF (for geometry) when available.
 
 ---
 
 ## Library API (embed in another script)
 
-Minimal, plug-and-play API
-Import from `sigdetect.api` and get plain dicts out (JSON-ready), 
-with no I/O side effects by default:
+Minimal, plug-and-play API that returns plain dicts (JSON-ready) without side effects unless you opt into cropping. Engine selection is forced to `auto` (PyMuPDF preferred) to ensure geometry. Wet detection runs automatically for non-e-sign PDFs; pass `runWetDetection=False` to skip OCR.
 
 ~~~python
 from pathlib import Path
@@ -167,6 +171,7 @@ result = DetectPdf(
     profileName="retainer",
     includePseudoSignatures=True,
     recurseXObjects=True,
+    # runWetDetection=False,  # disable OCR-backed wet detection if desired
 )
 print(
     result["file"],
@@ -189,26 +194,17 @@ for res in ScanDirectory(
     # store in DB, print, etc.
     pass
 
-# 3) Crop PNG snippets for FileResult objects (requires PyMuPDF)
+# 3) Create DOCX crops for FileResult objects (requires PyMuPDF + python-docx)
 detector = get_detector(pdfRoot="/path/to/pdfs", profileName="hipaa")
 file_result = detector.Detect(Path("/path/to/pdfs/example.pdf"))
-crops = CropSignatureImages(
+CropSignatureImages(
     "/path/to/pdfs/example.pdf",
     file_result,
     outputDirectory="./signature_crops",
     dpi=200,
-    returnBytes=True,  # also returns in-memory PNG bytes for each crop
-    # saveToDisk=False,  # optional: skip writing PNGs to disk
 )
-
-first_crop = crops[0]
-print(first_crop.path, len(first_crop.image_bytes))
 ~~~
 
-When ``returnBytes=True`` the helper returns ``SignatureCrop`` objects containing the saved path,
-PNG bytes, and the originating signature metadata.
-Pass ``saveToDisk=False`` if you only want in-memory PNG bytes (no files on disk or ``crop_path`` updates).
-
 
 ## Result schema
 
@@ -237,7 +233,7 @@ High-level summary (per file):
       "hint": "AcroSig:sig_patient",
       "render_type": "typed",
       "bounding_box": [10.0, 10.0, 150.0, 40.0],
-      "crop_path": "signature_crops/example/sig_01_patient.png"
+      "crop_path": "signature_crops/example/sig_01_patient.docx"
     },
     {
       "page": null,
@@ -247,7 +243,7 @@ High-level summary (per file):
       "scores": { "page_label": 4, "general": 2 },
       "evidence": ["page_label:representative(parent/guardian)", "pseudo:true"],
       "hint": "VendorOrAcroOnly",
-      "render_type": "unknown",
+      "render_type": "typed",
       "bounding_box": null,
       "crop_path": null
     }
@@ -263,7 +259,8 @@ High-level summary (per file):
 - **`roles`** summarizes unique non-`unknown` roles across signatures.
 - In retainer profile, emitter prefers two signatures (client + firm), often on the same page.
 - **`signatures[].bounding_box`** reports the widget rectangle in PDF points (origin bottom-left).  
-- **`signatures[].crop_path`** is populated when PNG crops are generated (via CLI `--crop-signatures` or `CropSignatureImages`).
+- **`signatures[].crop_path`** is populated when DOCX crops are generated (via CLI `--crop-signatures` or `CropSignatureImages`).
+- **`signatures[].crop_bytes`** contains base64 PNG data when CLI `--crop-bytes` is enabled.
 
 ---
 
@@ -285,13 +282,18 @@ You can keep one config YAML per dataset, e.g.:
 # ./sample_data/config.yml (example)
 pdf_root: ./pdfs
 out_dir: ./sigdetect_out
-engine: pypdf2
+engine: auto
+write_results: false
 pseudo_signatures: true
 recurse_xobjects: true
 profile: retainer    # or: hipaa
-crop_signatures: false   # enable to write PNG crops (requires pymupdf)
+crop_signatures: false   # enable to write DOCX crops (requires pymupdf + python-docx)
 # crop_output_dir: ./signature_crops
 crop_image_dpi: 200
+detect_wet_signatures: false   # kept for compatibility; non-e-sign PDFs still trigger OCR
+wet_ocr_dpi: 200
+wet_ocr_languages: eng
+wet_precision_threshold: 0.82
 ~~~
 
 YAML files can be customized or load at runtime (see CLI `--config`, if available, or import and pass patterns into engine).
@@ -306,6 +308,7 @@ YAML files can be customized or load at runtime (see CLI `--config`, if availabl
   - Looks for client and firm labels/tokens; boosts pages with law-firm markers (LLP/LLC/PA/PC) and “By:” blocks.
   - Applies an anti-front-matter rule to reduce page-1 false positives (e.g., letterheads, firm mastheads).
   - When only vendor/Acro clues exist (no widgets), it will emit two pseudo signatures targeting likely pages.
+- **Wet detection (non-e-sign):** The CLI runs an OCR-backed pass (PyMuPDF + pytesseract/Tesseract) after e-sign detection whenever no e-sign evidence is found. It emits `RenderType="wet"` signatures for high-confidence label/stroke pairs in the lower page region. When an image-based signature is present on a page, label-only OCR candidates are suppressed unless a stroke is detected. Results are deduped to the top signature per role (dropping `unknown`). Missing OCR dependencies add a `ManualReview:*` hint instead of failing.
 
 ---
 

{sigdetect-0.3.1 → sigdetect-0.5.0}/README.md

@@ -79,14 +79,16 @@ sigdetect detect \
 ### Notes
 
 - The config file controls `pdf_root`, `out_dir`, `engine`, `pseudo_signatures`, `recurse_xobjects`, etc.
-- `--engine` supports **pypdf2** (default); a **pymupdf** engine placeholder exists and may be included in a future build.  
+- Engine selection is forced to **auto** (prefers PyMuPDF for geometry, falls back to PyPDF2); any configured `engine` value is overridden.  
 - `--pseudo-signatures` enables a vendor/Acro-only pseudo-signature when no actual `/Widget` is present (useful for DocuSign / Acrobat Sign receipts).  
 - `--recurse-xobjects` allows scanning Form XObjects for vendor markers and labels embedded in page resources. 
 - `--profile` selects tuned role logic:
   - `hipaa` → patient / representative / attorney
   - `retainer` → client / firm (prefers detecting two signatures)
 - `--recursive/--no-recursive` toggles whether `sigdetect detect` descends into subdirectories when hunting for PDFs (recursive by default).
-- `--crop-signatures` enables PNG crops for each detected widget (requires installing the optional `pymupdf` dependency). Use `--crop-dir` to override the destination and `--crop-dpi` to choose rendering quality.
+- Results output is disabled by default; set `write_results: true` or pass `--write-results` when you need `results.json` (for EDA).
+- Cropping (`--crop-signatures`) writes a one-image `.docx` per signature in the crop output directory (no PNG files on disk); `--crop-bytes` embeds base64 PNG data in `signatures[].crop_bytes` for in-memory use. PyMuPDF is required for crops, and `python-docx` is required for `.docx` output.
+- Wet detection runs automatically for non-e-sign PDFs when dependencies are available; missing OCR dependencies add a `ManualReview:*` hint instead of failing. PyMuPDF + Tesseract are required for wet detection.
 - If the executable is not on `PATH`, you can always fall back to `python -m sigdetect.cli ...`.
 
 ### EDA (quick aggregate stats)
@@ -97,6 +99,8 @@ sigdetect eda \
 
 ~~~
 
+`sigdetect eda` expects `results.json`; enable `write_results: true` when running detect.
+
 ---
 
 ## Library usage
@@ -120,15 +124,13 @@ result = detector.Detect(Path("/path/to/pdfs/example.pdf"))
 print(result.to_dict())
 ~~~
 
-`Detect(Path)` returns a **FileResult** dataclass; call `.to_dict()` for the JSON-friendly representation (see [Result schema](#result-schema)). Each signature entry now exposes `bounding_box` coordinates (PDF points, origin bottom-left). When PNG cropping is enabled, `crop_path` points at the generated image.
+`Detect(Path)` returns a **FileResult** dataclass; call `.to_dict()` for the JSON-friendly representation (see [Result schema](#result-schema)). Each signature entry now exposes `bounding_box` coordinates (PDF points, origin bottom-left). When cropping is enabled, `crop_path` points at the generated `.docx`. Use `Engine="auto"` if you want the single-pass defaults that prefer PyMuPDF (for geometry) when available.
 
 ---
 
 ## Library API (embed in another script)
 
-Minimal, plug-and-play API
-Import from `sigdetect.api` and get plain dicts out (JSON-ready), 
-with no I/O side effects by default:
+Minimal, plug-and-play API that returns plain dicts (JSON-ready) without side effects unless you opt into cropping. Engine selection is forced to `auto` (PyMuPDF preferred) to ensure geometry. Wet detection runs automatically for non-e-sign PDFs; pass `runWetDetection=False` to skip OCR.
 
 ~~~python
 from pathlib import Path
@@ -151,6 +153,7 @@ result = DetectPdf(
     profileName="retainer",
     includePseudoSignatures=True,
     recurseXObjects=True,
+    # runWetDetection=False,  # disable OCR-backed wet detection if desired
 )
 print(
     result["file"],
@@ -173,26 +176,17 @@ for res in ScanDirectory(
     # store in DB, print, etc.
     pass
 
-# 3) Crop PNG snippets for FileResult objects (requires PyMuPDF)
+# 3) Create DOCX crops for FileResult objects (requires PyMuPDF + python-docx)
 detector = get_detector(pdfRoot="/path/to/pdfs", profileName="hipaa")
 file_result = detector.Detect(Path("/path/to/pdfs/example.pdf"))
-crops = CropSignatureImages(
+CropSignatureImages(
     "/path/to/pdfs/example.pdf",
     file_result,
     outputDirectory="./signature_crops",
     dpi=200,
-    returnBytes=True,  # also returns in-memory PNG bytes for each crop
-    # saveToDisk=False,  # optional: skip writing PNGs to disk
 )
-
-first_crop = crops[0]
-print(first_crop.path, len(first_crop.image_bytes))
 ~~~
 
-When ``returnBytes=True`` the helper returns ``SignatureCrop`` objects containing the saved path,
-PNG bytes, and the originating signature metadata.
-Pass ``saveToDisk=False`` if you only want in-memory PNG bytes (no files on disk or ``crop_path`` updates).
-
 
 ## Result schema
 
@@ -221,7 +215,7 @@ High-level summary (per file):
       "hint": "AcroSig:sig_patient",
       "render_type": "typed",
       "bounding_box": [10.0, 10.0, 150.0, 40.0],
-      "crop_path": "signature_crops/example/sig_01_patient.png"
+      "crop_path": "signature_crops/example/sig_01_patient.docx"
     },
     {
       "page": null,
@@ -231,7 +225,7 @@ High-level summary (per file):
       "scores": { "page_label": 4, "general": 2 },
       "evidence": ["page_label:representative(parent/guardian)", "pseudo:true"],
       "hint": "VendorOrAcroOnly",
-      "render_type": "unknown",
+      "render_type": "typed",
       "bounding_box": null,
       "crop_path": null
     }
@@ -247,7 +241,8 @@ High-level summary (per file):
 - **`roles`** summarizes unique non-`unknown` roles across signatures.
 - In retainer profile, emitter prefers two signatures (client + firm), often on the same page.
 - **`signatures[].bounding_box`** reports the widget rectangle in PDF points (origin bottom-left).  
-- **`signatures[].crop_path`** is populated when PNG crops are generated (via CLI `--crop-signatures` or `CropSignatureImages`).
+- **`signatures[].crop_path`** is populated when DOCX crops are generated (via CLI `--crop-signatures` or `CropSignatureImages`).
+- **`signatures[].crop_bytes`** contains base64 PNG data when CLI `--crop-bytes` is enabled.
 
 ---
 
@@ -269,13 +264,18 @@ You can keep one config YAML per dataset, e.g.:
 # ./sample_data/config.yml (example)
 pdf_root: ./pdfs
 out_dir: ./sigdetect_out
-engine: pypdf2
+engine: auto
+write_results: false
 pseudo_signatures: true
 recurse_xobjects: true
 profile: retainer    # or: hipaa
-crop_signatures: false   # enable to write PNG crops (requires pymupdf)
+crop_signatures: false   # enable to write DOCX crops (requires pymupdf + python-docx)
 # crop_output_dir: ./signature_crops
 crop_image_dpi: 200
+detect_wet_signatures: false   # kept for compatibility; non-e-sign PDFs still trigger OCR
+wet_ocr_dpi: 200
+wet_ocr_languages: eng
+wet_precision_threshold: 0.82
 ~~~
 
 YAML files can be customized or load at runtime (see CLI `--config`, if available, or import and pass patterns into engine).
@@ -290,6 +290,7 @@ YAML files can be customized or load at runtime (see CLI `--config`, if availabl
   - Looks for client and firm labels/tokens; boosts pages with law-firm markers (LLP/LLC/PA/PC) and “By:” blocks.
   - Applies an anti-front-matter rule to reduce page-1 false positives (e.g., letterheads, firm mastheads).
   - When only vendor/Acro clues exist (no widgets), it will emit two pseudo signatures targeting likely pages.
+- **Wet detection (non-e-sign):** The CLI runs an OCR-backed pass (PyMuPDF + pytesseract/Tesseract) after e-sign detection whenever no e-sign evidence is found. It emits `RenderType="wet"` signatures for high-confidence label/stroke pairs in the lower page region. When an image-based signature is present on a page, label-only OCR candidates are suppressed unless a stroke is detected. Results are deduped to the top signature per role (dropping `unknown`). Missing OCR dependencies add a `ManualReview:*` hint instead of failing.
 
 ---
 

{sigdetect-0.3.1 → sigdetect-0.5.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sigdetect"
-version = "0.3.1"
+version = "0.5.0"
 description = "Signature detection and role attribution for PDFs"
 readme = "README.md"
 authors = [{ name = "BT Asmamaw", email = "basmamaw@angeiongroup.com" }]
@@ -15,12 +15,13 @@ dependencies = [
   "rich>=13.0",
   "typer>=0.12",
   "pydantic>=2.5",
+  "pillow>=10.0",
+  "python-docx>=1.1.0",
+  "pytesseract>=0.3.10",
+  "pymupdf>=1.23",
   "pyyaml>=6.0",
 ]
 
-[project.optional-dependencies]
-pymupdf = ["pymupdf>=1.23"]
-
 [project.scripts]
 sigdetect = "sigdetect.cli:app"
 

{sigdetect-0.3.1 → sigdetect-0.5.0}/src/sigdetect/__init__.py

@@ -21,4 +21,4 @@ try:
 except PackageNotFoundError:  # pragma: no cover
     __version__ = "0.0.0"
 
-DEFAULT_ENGINE = "pypdf2"
+DEFAULT_ENGINE = "auto"

{sigdetect-0.3.1 → sigdetect-0.5.0}/src/sigdetect/api.py

@@ -9,8 +9,9 @@ from typing import Any, Generator, Iterable, Iterator, Literal, overload
 from sigdetect.config import DetectConfiguration
 from sigdetect.cropping import SignatureCrop
 from sigdetect.detector import BuildDetector, Detector, FileResult, Signature
+from sigdetect.wet_detection import apply_wet_detection
 
-EngineName = Literal["pypdf2", "pypdf", "pymupdf"]
+EngineName = Literal["pypdf2", "pypdf", "pymupdf", "auto"]
 ProfileName = Literal["hipaa", "retainer"]
 
 
@@ -18,12 +19,16 @@ def DetectPdf(
     pdfPath: str | Path,
     *,
     profileName: ProfileName = "hipaa",
-    engineName: EngineName = "pypdf2",
+    engineName: EngineName = "auto",
     includePseudoSignatures: bool = True,
     recurseXObjects: bool = True,
+    runWetDetection: bool = True,
     detector: Detector | None = None,
 ) -> dict[str, Any]:
-    """Detect signature evidence and assign roles for a single PDF."""
+    """Detect signature evidence and assign roles for a single PDF.
+
+    Wet detection runs by default for non-e-sign PDFs; pass ``runWetDetection=False`` to skip OCR.
+    """
 
     resolvedPath = Path(pdfPath)
     activeDetector = detector or get_detector(
@@ -36,6 +41,10 @@ def DetectPdf(
     )
 
     result = activeDetector.Detect(resolvedPath)
+    if runWetDetection:
+        configuration = _ResolveConfiguration(activeDetector)
+        if configuration is not None:
+            apply_wet_detection(resolvedPath, configuration, result)
     return _ToPlainDictionary(result)
 
 
@@ -43,12 +52,15 @@ def get_detector(
     *,
     pdfRoot: str | Path | None = None,
     profileName: ProfileName = "hipaa",
-    engineName: EngineName = "pypdf2",
+    engineName: EngineName = "auto",
     includePseudoSignatures: bool = True,
     recurseXObjects: bool = True,
     outputDirectory: str | Path | None = None,
 ) -> Detector:
-    """Return a reusable detector instance configured with the supplied options."""
+    """Return a reusable detector instance configured with the supplied options.
+
+    Engine selection is forced to ``auto`` (prefers PyMuPDF when available).
+    """
 
     configuration = DetectConfiguration(
         PdfRoot=Path(pdfRoot) if pdfRoot is not None else Path.cwd(),
@@ -108,6 +120,7 @@ def _ToPlainValue(value: Any) -> Any:
 def DetectMany(
     pdfPaths: Iterable[str | Path],
     *,
+    runWetDetection: bool = True,
     detector: Detector | None = None,
     **kwargs: Any,
 ) -> Iterator[dict[str, Any]]:
@@ -115,17 +128,18 @@ def DetectMany(
 
     if detector is not None:
         for pdfPath in pdfPaths:
-            yield _DetectWithDetector(detector, pdfPath)
+            yield _DetectWithDetector(detector, pdfPath, runWetDetection=runWetDetection)
         return
 
     for pdfPath in pdfPaths:
-        yield DetectPdf(pdfPath, **kwargs)
+        yield DetectPdf(pdfPath, runWetDetection=runWetDetection, **kwargs)
 
 
 def ScanDirectory(
     pdfRoot: str | Path,
     *,
     globPattern: str = "**/*.pdf",
+    runWetDetection: bool = True,
     detector: Detector | None = None,
     **kwargs: Any,
 ) -> Iterator[dict[str, Any]]:
@@ -143,7 +157,7 @@ def ScanDirectory(
 
     for pdfPath in iterator:
         if pdfPath.is_file() and pdfPath.suffix.lower() == ".pdf":
-            yield DetectPdf(pdfPath, detector=detector, **kwargs)
+            yield DetectPdf(pdfPath, detector=detector, runWetDetection=runWetDetection, **kwargs)
 
 
 def ToCsvRow(result: dict[str, Any]) -> dict[str, Any]:
@@ -174,11 +188,25 @@ def Version() -> str:
         return "0.0.0-dev"
 
 
-def _DetectWithDetector(detector: Detector, pdfPath: str | Path) -> dict[str, Any]:
+def _DetectWithDetector(
+    detector: Detector, pdfPath: str | Path, *, runWetDetection: bool
+) -> dict[str, Any]:
     """Helper that runs ``detector`` and returns the plain dictionary result."""
 
     resolvedPath = Path(pdfPath)
-    return _ToPlainDictionary(detector.Detect(resolvedPath))
+    result = detector.Detect(resolvedPath)
+    if runWetDetection:
+        configuration = _ResolveConfiguration(detector)
+        if configuration is not None:
+            apply_wet_detection(resolvedPath, configuration, result)
+    return _ToPlainDictionary(result)
+
+
+def _ResolveConfiguration(detector: Detector) -> DetectConfiguration | None:
+    configuration = getattr(detector, "Configuration", None)
+    if isinstance(configuration, DetectConfiguration):
+        return configuration
+    return None
 
 
 @contextmanager
@@ -225,12 +253,15 @@ def CropSignatureImages(
     returnBytes: bool = False,
     saveToDisk: bool = True,
 ) -> list[Path] | list[SignatureCrop]:
-    """Crop detected signature regions to PNG files.
+    """Create DOCX files containing cropped signature images.
 
     Accepts either a :class:`FileResult` instance or the ``dict`` returned by
     :func:`DetectPdf`. Requires the optional ``pymupdf`` dependency.
     Set ``returnBytes=True`` to also receive in-memory PNG bytes for each crop. Set
     ``saveToDisk=False`` to skip writing PNG files while still returning in-memory data.
+    When ``saveToDisk`` is enabled, a one-image DOCX file is also written per crop. When
+    ``returnBytes`` is True and ``python-docx`` is available, the returned
+    :class:`SignatureCrop` objects include ``docx_bytes``.
     """
 
     from sigdetect.cropping import crop_signatures
@@ -273,6 +304,7 @@ def _CoerceFileResult(
                 RenderType=str(entry.get("render_type") or "unknown"),
                 BoundingBox=tuple(bbox) if bbox else None,
                 CropPath=entry.get("crop_path"),
+                CropBytes=entry.get("crop_bytes"),
             )
         )