kreuzberg 3.13.2__py3-none-any.whl → 3.14.0__py3-none-any.whl

kreuzberg/_utils/_image_preprocessing.py

@@ -0,0 +1,346 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+from PIL import Image
+
+from kreuzberg._types import ExtractionConfig, ImagePreprocessingMetadata
+
+if TYPE_CHECKING:
+    from PIL.Image import Image as PILImage
+
+
+def calculate_optimal_dpi(
+    page_width: float,
+    page_height: float,
+    target_dpi: int,
+    max_dimension: int,
+    min_dpi: int = 72,
+    max_dpi: int = 600,
+) -> int:
+    """Calculate optimal DPI based on page dimensions and constraints.
+
+    Args:
+        page_width: Page width in points (1/72 inch)
+        page_height: Page height in points (1/72 inch)
+        target_dpi: Desired target DPI
+        max_dimension: Maximum allowed pixel dimension
+        min_dpi: Minimum DPI threshold
+        max_dpi: Maximum DPI threshold
+
+    Returns:
+        Optimal DPI value that keeps image within max_dimension
+    """
+    # Convert points to inches (72 points = 1 inch)
+    width_inches = page_width / 72.0
+    height_inches = page_height / 72.0
+
+    # Calculate pixel dimensions at target DPI
+    target_width_pixels = int(width_inches * target_dpi)
+    target_height_pixels = int(height_inches * target_dpi)
+
+    # Check if target DPI results in oversized image
+    max_pixel_dimension = max(target_width_pixels, target_height_pixels)
+
+    if max_pixel_dimension <= max_dimension:
+        # Target DPI is fine, clamp to min/max bounds
+        return max(min_dpi, min(target_dpi, max_dpi))
+
+    # Calculate maximum DPI that keeps within dimension constraints
+    max_dpi_for_width = max_dimension / width_inches if width_inches > 0 else max_dpi
+    max_dpi_for_height = max_dimension / height_inches if height_inches > 0 else max_dpi
+    constrained_dpi = int(min(max_dpi_for_width, max_dpi_for_height))
+
+    # Clamp to min/max bounds
+    return max(min_dpi, min(constrained_dpi, max_dpi))
+
+
+def _extract_image_dpi(image: PILImage) -> tuple[tuple[float, float], float]:
+    """Extract DPI information from image."""
+    current_dpi_info = image.info.get("dpi", (72.0, 72.0))
+    if isinstance(current_dpi_info, (list, tuple)):
+        original_dpi = (float(current_dpi_info[0]), float(current_dpi_info[1]))
+        current_dpi = float(current_dpi_info[0])  # Use horizontal DPI
+    else:
+        current_dpi = float(current_dpi_info)
+        original_dpi = (current_dpi, current_dpi)
+    return original_dpi, current_dpi
+
+
+def _should_skip_processing(
+    original_width: int,
+    original_height: int,
+    current_dpi: float,
+    config: ExtractionConfig,
+) -> bool:
+    """Check if processing should be skipped."""
+    max_current_dimension = max(original_width, original_height)
+    current_matches_target = abs(current_dpi - config.target_dpi) < 1.0
+    return not config.auto_adjust_dpi and current_matches_target and max_current_dimension <= config.max_image_dimension
+
+
+def _calculate_target_dpi(
+    original_width: int,
+    original_height: int,
+    current_dpi: float,
+    config: ExtractionConfig,
+) -> tuple[int, bool, int | None]:
+    """Calculate target DPI and whether it was auto-adjusted."""
+    calculated_dpi = None
+    if config.auto_adjust_dpi:
+        # Convert pixel dimensions to approximate point dimensions
+        # This is an approximation since we don't know the actual physical size
+        approx_width_points = original_width * 72.0 / current_dpi
+        approx_height_points = original_height * 72.0 / current_dpi
+
+        optimal_dpi = calculate_optimal_dpi(
+            approx_width_points,
+            approx_height_points,
+            config.target_dpi,
+            config.max_image_dimension,
+            config.min_dpi,
+            config.max_dpi,
+        )
+        calculated_dpi = optimal_dpi
+        auto_adjusted = optimal_dpi != config.target_dpi
+        target_dpi = optimal_dpi
+    else:
+        auto_adjusted = False
+        target_dpi = config.target_dpi
+
+    return target_dpi, auto_adjusted, calculated_dpi
+
+
+def normalize_image_dpi(
+    image: PILImage,
+    config: ExtractionConfig,
+) -> tuple[PILImage, ImagePreprocessingMetadata]:
+    """Normalize image DPI and dimensions for optimal OCR processing.
+
+    Args:
+        image: PIL Image to normalize
+        config: ExtractionConfig containing DPI settings
+
+    Returns:
+        Tuple of (normalized_image, ImagePreprocessingMetadata)
+
+    Note:
+        If auto_adjust_dpi is False, uses target_dpi directly.
+        If True, calculates optimal DPI based on image dimensions and constraints.
+    """
+    original_width, original_height = image.size
+    original_dpi, current_dpi = _extract_image_dpi(image)
+
+    # If no auto-adjustment and current DPI matches target and within limits, skip processing
+    if _should_skip_processing(original_width, original_height, current_dpi, config):
+        return image, ImagePreprocessingMetadata(
+            original_dimensions=(original_width, original_height),
+            original_dpi=original_dpi,
+            target_dpi=config.target_dpi,
+            scale_factor=1.0,
+            auto_adjusted=False,
+            final_dpi=config.target_dpi,
+            skipped_resize=True,
+        )
+
+    # Calculate target DPI
+    target_dpi, auto_adjusted, calculated_dpi = _calculate_target_dpi(
+        original_width, original_height, current_dpi, config
+    )
+
+    # Calculate scale factor based on DPI ratio
+    scale_factor = target_dpi / current_dpi
+
+    # If scale factor is very close to 1.0, skip resizing
+    if abs(scale_factor - 1.0) < 0.05:
+        return image, ImagePreprocessingMetadata(
+            original_dimensions=(original_width, original_height),
+            original_dpi=original_dpi,
+            target_dpi=config.target_dpi,
+            scale_factor=scale_factor,
+            auto_adjusted=auto_adjusted,
+            final_dpi=target_dpi,
+            calculated_dpi=calculated_dpi,
+            skipped_resize=True,
+        )
+
+    # Calculate new dimensions
+    new_width = int(original_width * scale_factor)
+    new_height = int(original_height * scale_factor)
+
+    # Ensure we don't exceed max_dimension (safety check)
+    dimension_clamped = False
+    max_new_dimension = max(new_width, new_height)
+    if max_new_dimension > config.max_image_dimension:
+        dimension_scale = config.max_image_dimension / max_new_dimension
+        new_width = int(new_width * dimension_scale)
+        new_height = int(new_height * dimension_scale)
+        scale_factor *= dimension_scale
+        dimension_clamped = True
+
+    # Resize image
+    try:
+        # Use LANCZOS for high-quality downscaling, BICUBIC for upscaling
+        # Handle different PIL versions
+        try:
+            # Modern PIL version
+            if scale_factor < 1.0:
+                resample_method = Image.Resampling.LANCZOS
+                resample_name = "LANCZOS"
+            else:
+                resample_method = Image.Resampling.BICUBIC
+                resample_name = "BICUBIC"
+        except AttributeError:
+            # Older PIL version
+            if scale_factor < 1.0:
+                resample_method = getattr(Image, "LANCZOS", 1)  # type: ignore[arg-type]
+                resample_name = "LANCZOS"
+            else:
+                resample_method = getattr(Image, "BICUBIC", 3)  # type: ignore[arg-type]
+                resample_name = "BICUBIC"
+
+        normalized_image = image.resize((new_width, new_height), resample_method)
+
+        # Update DPI info in the new image
+        normalized_image.info["dpi"] = (target_dpi, target_dpi)
+
+        return normalized_image, ImagePreprocessingMetadata(
+            original_dimensions=(original_width, original_height),
+            original_dpi=original_dpi,
+            target_dpi=config.target_dpi,
+            scale_factor=scale_factor,
+            auto_adjusted=auto_adjusted,
+            final_dpi=target_dpi,
+            new_dimensions=(new_width, new_height),
+            resample_method=resample_name,
+            dimension_clamped=dimension_clamped,
+            calculated_dpi=calculated_dpi,
+        )
+
+    except OSError as e:
+        # If resizing fails, return original image with error info
+        return image, ImagePreprocessingMetadata(
+            original_dimensions=(original_width, original_height),
+            original_dpi=original_dpi,
+            target_dpi=config.target_dpi,
+            scale_factor=scale_factor,
+            auto_adjusted=auto_adjusted,
+            final_dpi=target_dpi,
+            calculated_dpi=calculated_dpi,
+            resize_error=str(e),
+        )
+
+
+def get_dpi_adjustment_heuristics(
+    width: float,
+    height: float,
+    current_dpi: int,
+    target_dpi: int,
+    max_dimension: int,
+    content_type: str = "document",
+) -> dict[str, Any]:
+    """Get smart DPI adjustment recommendations based on content analysis.
+
+    Args:
+        width: Image width in pixels
+        height: Image height in pixels
+        current_dpi: Current DPI setting
+        target_dpi: Desired target DPI
+        max_dimension: Maximum allowed dimension
+        content_type: Type of content ("document", "photo", "mixed")
+
+    Returns:
+        Dictionary with adjustment recommendations and rationale
+    """
+    recommendations: list[str] = []
+    heuristics = {
+        "recommended_dpi": target_dpi,
+        "content_analysis": {},
+        "performance_impact": "medium",
+        "quality_impact": "medium",
+        "recommendations": recommendations,
+    }
+
+    # Calculate aspect ratio and size analysis
+    aspect_ratio = width / height if height > 0 else 1.0
+    total_pixels = width * height
+    megapixels = total_pixels / 1_000_000
+
+    heuristics["content_analysis"] = {
+        "aspect_ratio": aspect_ratio,
+        "megapixels": megapixels,
+        "is_portrait": aspect_ratio < 0.8,
+        "is_landscape": aspect_ratio > 1.2,
+        "is_large": max(width, height) > max_dimension * 0.8,
+    }
+
+    # Document-specific heuristics
+    if content_type == "document":
+        if aspect_ratio > 2.0 or aspect_ratio < 0.5:
+            # Very wide or very tall documents (like forms, receipts)
+            recommendations.append("Consider higher DPI for narrow documents")
+            if target_dpi < 200:
+                heuristics["recommended_dpi"] = min(200, target_dpi * 1.3)
+
+        if megapixels > 50:  # Very large document
+            recommendations.append("Large document detected - consider DPI reduction")
+            heuristics["performance_impact"] = "high"
+            if target_dpi > 150:
+                heuristics["recommended_dpi"] = max(120, target_dpi * 0.8)
+
+    # Memory usage estimation
+    estimated_memory_mb = (width * height * 3) / (1024 * 1024)  # RGB bytes
+    if estimated_memory_mb > 200:
+        heuristics["performance_impact"] = "high"
+        recommendations.append(f"High memory usage expected (~{estimated_memory_mb:.0f}MB)")
+
+    # Quality vs performance tradeoffs
+    scale_factor = target_dpi / current_dpi if current_dpi > 0 else 1.0
+    if scale_factor < 0.7:
+        heuristics["quality_impact"] = "high"
+        recommendations.append("Significant downscaling may reduce OCR accuracy")
+    elif scale_factor > 1.5:
+        heuristics["performance_impact"] = "high"
+        recommendations.append("Upscaling will increase processing time")
+
+    return heuristics
+
+
+def estimate_processing_time(
+    width: int,
+    height: int,
+    ocr_backend: str = "tesseract",
+) -> dict[str, float | str]:
+    """Estimate processing time based on image dimensions and OCR backend.
+
+    Args:
+        width: Image width in pixels
+        height: Image height in pixels
+        ocr_backend: OCR backend name
+
+    Returns:
+        Dictionary with time estimates in seconds
+    """
+    total_pixels = width * height
+    megapixels = total_pixels / 1_000_000
+
+    # Base processing times per megapixel (rough estimates)
+    base_times = {
+        "tesseract": 2.5,  # seconds per megapixel
+        "easyocr": 4.0,  # slower due to deep learning
+        "paddleocr": 3.5,  # moderate speed
+    }
+
+    base_time = base_times.get(ocr_backend, 3.0)
+
+    # Non-linear scaling for very large images
+    scaling_factor = 1.0 + (megapixels - 10) * 0.1 if megapixels > 10 else 1.0
+
+    estimated_time = base_time * megapixels * scaling_factor
+
+    return {
+        "estimated_seconds": estimated_time,
+        "megapixels": megapixels,
+        "backend": ocr_backend,
+        "scaling_factor": scaling_factor,
+    }

kreuzberg/_utils/_ocr_cache.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import hashlib
 import io
+from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 import anyio
@@ -9,17 +10,13 @@ import anyio
 from kreuzberg._utils._cache import get_ocr_cache
 
 if TYPE_CHECKING:
-    from pathlib import Path
-
     from PIL.Image import Image as PILImage
 
     from kreuzberg._types import ExtractionResult
 
 
 def get_file_info(path: Path) -> dict[str, Any]:
-    from pathlib import Path as PathType  # noqa: PLC0415
-
-    path_obj = PathType(path) if not isinstance(path, PathType) else path
+    path_obj = path if isinstance(path, Path) else Path(path)
 
     try:
         stat = path_obj.stat()

kreuzberg/_utils/_process_pool.py

@@ -4,7 +4,7 @@ import io
 import multiprocessing as mp
 from concurrent.futures import ProcessPoolExecutor
 from contextlib import contextmanager
-from typing import TYPE_CHECKING, Any, TypeVar
+from typing import TYPE_CHECKING, Any, TypeVar, cast
 
 import anyio
 import psutil
@@ -173,7 +173,7 @@ class ProcessPoolManager:
                     self._active_tasks -= 1
 
         async with anyio.create_task_group() as tg:
-            results: list[T] = [None] * len(arg_batches)  # type: ignore[list-item]
+            results: list[T | None] = [None] * len(arg_batches)
 
             async def run_task(idx: int, args: tuple[Any, ...]) -> None:
                 results[idx] = await submit_single(args)
@@ -181,7 +181,7 @@ class ProcessPoolManager:
             for idx, args in enumerate(arg_batches):
                 tg.start_soon(run_task, idx, args)
 
-        return results
+        return cast("list[T]", results)
 
     def get_system_info(self) -> dict[str, Any]:
         memory = psutil.virtual_memory()

kreuzberg/_utils/_table.py

@@ -89,6 +89,8 @@ def _format_table_row(row: Any, df: Any, float_col_formatting: dict[str, str]) -
                     formatted_row.append(str(int(value)))
                 else:
                     formatted_row.append(f"{value:.2f}")
+            elif isinstance(value, bool):
+                formatted_row.append(str(value).lower())
             else:
                 clean_value = str(value).strip().replace("|", "\\|")
                 formatted_row.append(clean_value)
@@ -201,7 +203,8 @@ def extract_table_structure_info(table: TableData) -> dict[str, Any]:
 
     total_cells = df.height * df.width
     if total_cells > 0:
-        empty_cells = df.null_count().sum().item()
+        null_counts = df.null_count()
+        empty_cells = sum(null_counts.row(0))
         info["empty_cells"] = empty_cells
         info["data_density"] = (total_cells - empty_cells) / total_cells
 

kreuzberg/cli.py

@@ -62,7 +62,20 @@ def format_extraction_result(result: ExtractionResult, show_metadata: bool, outp
         if show_metadata:
             output_data["metadata"] = result.metadata
         if result.tables:
-            output_data["tables"] = result.tables
+            json_tables = []
+            for table in result.tables:
+                json_table = {
+                    "page_number": table.get("page_number"),
+                    "text": table.get("text"),
+                }
+                if "df" in table and table["df"] is not None:
+                    df = table["df"]
+                    if hasattr(df, "write_csv"):
+                        json_table["data_csv"] = df.write_csv()
+                    elif hasattr(df, "to_csv"):
+                        json_table["data_csv"] = df.to_csv(index=False)
+                json_tables.append(json_table)
+            output_data["tables"] = json_tables
         if result.chunks:
             output_data["chunks"] = result.chunks
         return json.dumps(output_data, indent=2, ensure_ascii=False)
@@ -77,7 +90,11 @@ def format_extraction_result(result: ExtractionResult, show_metadata: bool, outp
         output_parts.append("\n\n--- TABLES ---")
         for i, table in enumerate(result.tables):
             output_parts.append(f"\nTable {i + 1}:")
-            output_parts.append(json.dumps(table, indent=2, ensure_ascii=False))
+            json_table = {
+                "page_number": table.get("page_number"),
+                "text": table.get("text"),
+            }
+            output_parts.append(json.dumps(json_table, indent=2, ensure_ascii=False))
 
     return "\n".join(output_parts)
 

kreuzberg/extraction.py

@@ -426,12 +426,12 @@ def batch_extract_file_sync(
     with ThreadPoolExecutor(max_workers=max_workers) as executor:
         future_to_index = {executor.submit(extract_single, fp): i for i, fp in enumerate(file_paths)}
 
-        results: list[ExtractionResult] = [None] * len(file_paths)  # type: ignore[list-item]
+        results: list[ExtractionResult | None] = [None] * len(file_paths)
         for future in as_completed(future_to_index):
             index, result = future.result()
             results[index] = result
 
-    return results
+    return cast("list[ExtractionResult]", results)
 
 
 def batch_extract_bytes_sync(
@@ -479,9 +479,9 @@ def batch_extract_bytes_sync(
     with ThreadPoolExecutor(max_workers=max_workers) as executor:
         future_to_index = {executor.submit(extract_single, (i, content)): i for i, content in enumerate(contents)}
 
-        results: list[ExtractionResult] = [None] * len(contents)  # type: ignore[list-item]
+        results: list[ExtractionResult | None] = [None] * len(contents)
         for future in as_completed(future_to_index):
             index, result = future.result()
             results[index] = result
 
-    return results
+    return cast("list[ExtractionResult]", results)

{kreuzberg-3.13.2.dist-info → kreuzberg-3.14.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: kreuzberg
-Version: 3.13.2
+Version: 3.14.0
 Summary: Document intelligence framework for Python - Extract text, metadata, and structured data from diverse file formats
 Project-URL: documentation, https://kreuzberg.dev
 Project-URL: homepage, https://github.com/Goldziher/kreuzberg
@@ -31,15 +31,15 @@ Requires-Python: >=3.10
 Requires-Dist: anyio>=4.10.0
 Requires-Dist: chardetng-py>=0.3.5
 Requires-Dist: exceptiongroup>=1.2.2; python_version < '3.11'
-Requires-Dist: html-to-markdown[lxml]>=1.9.1
-Requires-Dist: mcp>=1.13.0
+Requires-Dist: html-to-markdown[lxml]>=1.11.0
+Requires-Dist: mcp>=1.14.0
 Requires-Dist: msgspec>=0.18.0
-Requires-Dist: numpy>=1.24.0
+Requires-Dist: numpy>=2.0.0
 Requires-Dist: playa-pdf>=0.7.0
-Requires-Dist: polars>=1.33.0
+Requires-Dist: polars>=1.33.1
 Requires-Dist: psutil>=7.0.0
 Requires-Dist: pypdfium2==4.30.0
-Requires-Dist: python-calamine>=0.5.2
+Requires-Dist: python-calamine>=0.5.3
 Requires-Dist: python-pptx>=1.0.2
 Requires-Dist: typing-extensions>=4.15.0; python_version < '3.12'
 Provides-Extra: additional-extensions
@@ -55,17 +55,17 @@ Requires-Dist: keybert>=0.9.0; extra == 'all'
 Requires-Dist: litestar[opentelemetry,standard,structlog]>=2.17.0; extra == 'all'
 Requires-Dist: mailparse>=1.0.15; extra == 'all'
 Requires-Dist: paddleocr>=3.2.0; extra == 'all'
-Requires-Dist: paddlepaddle>=3.1.1; extra == 'all'
+Requires-Dist: paddlepaddle>=3.2.0; extra == 'all'
 Requires-Dist: playa-pdf[crypto]>=0.7.0; extra == 'all'
 Requires-Dist: rich>=14.1.0; extra == 'all'
-Requires-Dist: semantic-text-splitter>=0.27.0; extra == 'all'
+Requires-Dist: semantic-text-splitter>=0.28.0; extra == 'all'
 Requires-Dist: setuptools>=80.9.0; extra == 'all'
 Requires-Dist: spacy>=3.8.7; extra == 'all'
 Requires-Dist: tomli>=2.0.0; (python_version < '3.11') and extra == 'all'
 Provides-Extra: api
 Requires-Dist: litestar[opentelemetry,standard,structlog]>=2.17.0; extra == 'api'
 Provides-Extra: chunking
-Requires-Dist: semantic-text-splitter>=0.27.0; extra == 'chunking'
+Requires-Dist: semantic-text-splitter>=0.28.0; extra == 'chunking'
 Provides-Extra: cli
 Requires-Dist: click>=8.2.1; extra == 'cli'
 Requires-Dist: rich>=14.1.0; extra == 'cli'
@@ -85,7 +85,7 @@ Provides-Extra: langdetect
 Requires-Dist: fast-langdetect>=0.3.2; extra == 'langdetect'
 Provides-Extra: paddleocr
 Requires-Dist: paddleocr>=3.2.0; extra == 'paddleocr'
-Requires-Dist: paddlepaddle>=3.1.1; extra == 'paddleocr'
+Requires-Dist: paddlepaddle>=3.2.0; extra == 'paddleocr'
 Requires-Dist: setuptools>=80.9.0; extra == 'paddleocr'
 Description-Content-Type: text/markdown
 

{kreuzberg-3.13.2.dist-info → kreuzberg-3.14.0.dist-info}/RECORD

@@ -1,57 +1,58 @@
 kreuzberg/__init__.py,sha256=Oh_NTp8wf0BlvD8CSBad2A493nEWH4jTE0x8v7v1Y9w,1341
 kreuzberg/__main__.py,sha256=3cIDdzTggj2kj8uKx4WShWHmCWqdZazdM3BxUGbAuSI,104
 kreuzberg/_chunker.py,sha256=tr9_KUYTSLauFois3MsB-A-0hGcTT8hTQFrqNRTii-I,1373
-kreuzberg/_config.py,sha256=Q5oiJE1XRf8ITuYcO8LZAOB3G2zNlXz2458rgPSth-U,12257
+kreuzberg/_config.py,sha256=2LI5z9gXniqO4afrMmbZfMdhlT2701O5OlGKkrMo-bM,12385
 kreuzberg/_constants.py,sha256=Bxc8oiN-wHwnWXT9bEiJhTUcu1ygPpra5qHirAif3b4,191
 kreuzberg/_document_classification.py,sha256=Mz_s2GJGsEl7MQ-67BPoGYCZibTy9Sw0PScUZKBjKOA,5736
 kreuzberg/_entity_extraction.py,sha256=5YpPnqoJ5aiHd_sy4bN4-Ngiq79RhCV6yaUQE8joGXo,3503
-kreuzberg/_gmft.py,sha256=jKbD7V_KP9XTLjT9SBgSgE3CyDjqbRDm9BAiWV2sAC0,19542
+kreuzberg/_gmft.py,sha256=a7KDXbZM0PxyFpAIjM0xMRvxzoMo4fTQuGlFNa8uXBU,20502
 kreuzberg/_language_detection.py,sha256=T9p6aimB7QFXAQiEntIMZeH_Z62E52E8fBQ43hWuyhs,1960
 kreuzberg/_mime_types.py,sha256=kGBDSMO4XPgzUKC7iaBeChCtRQXZ9_zXq6eJydejX_k,7739
 kreuzberg/_playa.py,sha256=p4G5ymSSCbQoDeXJjH-yuVzdd4y-wKcolqDthjPtqok,11413
-kreuzberg/_registry.py,sha256=8cPpz3oZVnMwWDT2v_Q7wf-GHd5YuHmc-nkLtvPfE1I,2433
-kreuzberg/_types.py,sha256=D-2d_WG8HyByA163izGhjk7t-e4FL_N-_6UzlVso8Dg,36020
-kreuzberg/cli.py,sha256=nPH4FDW6WkoF4gtH0s4RWmxjAveJ_-Unb6fev6x0Sko,12752
+kreuzberg/_registry.py,sha256=8XYT-vPhNYMAbB5RBIUKz-1Zdg48OCnBcdVZzBq6YwY,3307
+kreuzberg/_types.py,sha256=yw8ZzCgwp8T4byh00gdSlABDtRwro6H1pemQsO5IZMQ,39132
+kreuzberg/cli.py,sha256=Ob0IfqWcaiM09pFdC6wTpdSeql0SGZDxBxfrEhJAGmo,13501
 kreuzberg/exceptions.py,sha256=PTiAZgQwcG9hXbgYg2W7sfxksFhq5_wzOFgZGnTJAoc,2991
-kreuzberg/extraction.py,sha256=jiMKiDyTf3sHyk76sMffHR-eH-_yg-DFRMuXEKufRYI,17649
+kreuzberg/extraction.py,sha256=qT-Ziw5FmMqcPT88VrglikL1RASSJCf5W7xP6L9Vi5s,17673
 kreuzberg/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 kreuzberg/_api/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-kreuzberg/_api/main.py,sha256=JALYRD0qwyoZloWk5dNNuslBtG4GlVNc0G2oADm6cAc,7578
+kreuzberg/_api/main.py,sha256=bZLaQpW8eoTFGvCGJgFodALy4rDfe9kuY1oj9OKPQpU,10792
 kreuzberg/_extractors/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 kreuzberg/_extractors/_base.py,sha256=i2FvAhRnamEtBb4a-C7pfcdWIXnkEBw0saMQu7h1_RQ,2069
 kreuzberg/_extractors/_email.py,sha256=jn_8J4BASKJ7zFHBG0PgxNe3OT4pjmEM2tTKX8y_0AE,5887
 kreuzberg/_extractors/_html.py,sha256=NyQKChNLvaSUC_5x1qTYlIQGwL4lEbgUF7BgH9ejEVY,1583
-kreuzberg/_extractors/_image.py,sha256=UqPoYfvDRX6Rd1yPhcLHJLDw6d2cUzgkqOGjh2eleJM,3301
+kreuzberg/_extractors/_image.py,sha256=lFPoxAf7_Zbx-1t8W4vU2bhHauiNGOAFbZxr_2gNUsw,3991
 kreuzberg/_extractors/_pandoc.py,sha256=-Ai4S1cXs7F6yeonb_7Y7_ZoWHn29E2oP1WlPtM-4HM,22505
-kreuzberg/_extractors/_pdf.py,sha256=Yv_c3xYzrGAjgTbwCGqbiQTDLjIUP_Pu7Z3GmMOqgqg,17865
+kreuzberg/_extractors/_pdf.py,sha256=naJ_AgtAgtGIjAqiU4_G7lgftKWhUjZDLVILSG2AyVc,18757
 kreuzberg/_extractors/_presentation.py,sha256=ULGkt7dzeA9sYSEhpAucKZmkdv9EubzeZtOjoLP3Z2E,6994
-kreuzberg/_extractors/_spread_sheet.py,sha256=x25u2M-ufxpDd7_qrjhMEz1yFftIcOISE1qwPW09Zm0,11962
+kreuzberg/_extractors/_spread_sheet.py,sha256=eBAx_OwoyRqMzmD4Z07UlOBwcXckymgvj_0o7di6thA,12715
 kreuzberg/_extractors/_structured.py,sha256=PpefI_GDrdLyUgnElrbdB-MeTMKVWium4Ckxm5Zg100,5536
 kreuzberg/_mcp/__init__.py,sha256=h6DgLFO4TMUk7_wCJ2jn2Y6IkFmfzb-Z7jX-G5UCYVc,43
-kreuzberg/_mcp/server.py,sha256=iYJG6g0u7I6mWtC4R1XlxydBrPpgnp5dGJzpm9QAZig,8438
+kreuzberg/_mcp/server.py,sha256=YPMJp6xnZ3DC32NEdX5Gqf3vwxsHZxXxUxZ6jghpv6I,5688
 kreuzberg/_ocr/__init__.py,sha256=grshVFwVQl2rMvH1hg1JNlYXjy5-Tdb_rusLD1Cselk,706
-kreuzberg/_ocr/_base.py,sha256=5ef2g8JuSaZF2sDiAmoaODHbeG4MT0LtNzbtW0n9BnU,1445
+kreuzberg/_ocr/_base.py,sha256=ZvOJvW8DtylQJZdCPk9vlVNZiBFK-dC4Oj7Kb6-mWkY,1419
 kreuzberg/_ocr/_easyocr.py,sha256=XbgpGt5tkE4xHleIGvV1cHlpOQTp43rSXBO1CyIyKTg,14599
-kreuzberg/_ocr/_paddleocr.py,sha256=58sKOHfKCHGFJNlRLrJwey8G_7xbsAAPBXB4n3hKc7k,14052
+kreuzberg/_ocr/_paddleocr.py,sha256=hfc6Zi2eSUYTVVF9y9D1P2_pLiLXPfFRoJ6QDJ6oZag,15017
 kreuzberg/_ocr/_table_extractor.py,sha256=LhBiCX8R_xR-uK1FH3ONA_vqOmqUWANZJ2HMCBLsmNY,5513
-kreuzberg/_ocr/_tesseract.py,sha256=xGML3ygY5xMN5T3YznrKDVAH_DWfaFiteFBo_-GpjCs,48931
+kreuzberg/_ocr/_tesseract.py,sha256=QEKK_PDZnNiZRgpklOgMXB-cObJy6C-HuxL6Gza5Z3c,49136
 kreuzberg/_utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-kreuzberg/_utils/_cache.py,sha256=S6Oc4TJamiuuWeJ2ABxDFbbQh4o8w38AUyZeBEc1NN8,12767
+kreuzberg/_utils/_cache.py,sha256=qeyI6rJOQlKtdHjJeOjUxx31eItak_drrNn8Cf8HbN8,13956
 kreuzberg/_utils/_device.py,sha256=UxGkSTN3Up-Zn43CSyvf8CozW2xAF05Cm01LWA2FZmg,8263
 kreuzberg/_utils/_document_cache.py,sha256=tfk9_Yc1cQkT5_uM5R1uaI4w-2SjNn7QyAd6AmWkSz8,4851
 kreuzberg/_utils/_errors.py,sha256=aQYEnp8oJ-WJVmCNo7YY-25y1KZZFEwjAmxVRfw4a_M,4920
-kreuzberg/_utils/_ocr_cache.py,sha256=8_-qmPlK2adQKsH4OO4Mlk8wmqBMl3XxkcV_NsXVyFs,3501
+kreuzberg/_utils/_image_preprocessing.py,sha256=2u0A28M07F9XlYebTG5salOUVEE3YT3m8fiR8Z2ZM8E,12326
+kreuzberg/_utils/_ocr_cache.py,sha256=uCCZfdY7EiqMhCnhNwqirFOr-Wfaobd2Ntc-F07TKec,3425
 kreuzberg/_utils/_pdf_lock.py,sha256=Ytvds30aZf3yXeZFo27ZenrhUoU-GZlR2rKEkhJ_wlk,1349
-kreuzberg/_utils/_process_pool.py,sha256=9dPMD_gBocQ5VaeCIrlSJfPXKyXNuyKaATmqOPExxiE,6723
+kreuzberg/_utils/_process_pool.py,sha256=7p8Co1w-Tvh2MUdxMcPMpvOikumrb0nN2ApQVytV-_c,6726
 kreuzberg/_utils/_quality.py,sha256=f7NbyZysyJQD8jKCNWhogvluU9A7GdEYhMsDBeMbGAA,5412
 kreuzberg/_utils/_ref.py,sha256=iOflvjTUc_F0XaL28Bd6fpvL6qkeoURGA4B77Nqky7I,840
 kreuzberg/_utils/_serialization.py,sha256=97iIgdcxdbym-BEvy0J6HAduBCUXyCGwhuEHCT_l7I4,1513
 kreuzberg/_utils/_string.py,sha256=wVyvEHByHBeu_6evmqJGv9Ml-NAwkyz60n8l-7L5Cw0,4366
 kreuzberg/_utils/_sync.py,sha256=OWiciXPTGHIxgiGoHI2AglZ1siTNT-nU_JCgHPNzzHk,2196
-kreuzberg/_utils/_table.py,sha256=R-6owHjvcvHGhem_vDsFH7S2yMHGoUUO2PFcj-Idptk,6361
+kreuzberg/_utils/_table.py,sha256=OVg6T2QnerMhVNb1juLTBSIjyjFiE5-OrUWr5NSCgnQ,6493
 kreuzberg/_utils/_tmp.py,sha256=wnOInBkcuQoxI1vBLvNv9NqbRCEu9Y03qfOjqQuAk3s,841
-kreuzberg-3.13.2.dist-info/METADATA,sha256=c1w8iB_Frnzr0DHY-X-a9rk5S9vQPICPIniPzwfvHV8,12127
-kreuzberg-3.13.2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-kreuzberg-3.13.2.dist-info/entry_points.txt,sha256=GplGhFryCP7kyAG_k-Mdahznvo2fwi73qLFg5yQfH_A,91
-kreuzberg-3.13.2.dist-info/licenses/LICENSE,sha256=-8caMvpCK8SgZ5LlRKhGCMtYDEXqTKH9X8pFEhl91_4,1066
-kreuzberg-3.13.2.dist-info/RECORD,,
+kreuzberg-3.14.0.dist-info/METADATA,sha256=68rRivXnf8n_F9lqekOydDOd8sehWpHpbbKzRup7XDc,12127
+kreuzberg-3.14.0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+kreuzberg-3.14.0.dist-info/entry_points.txt,sha256=GplGhFryCP7kyAG_k-Mdahznvo2fwi73qLFg5yQfH_A,91
+kreuzberg-3.14.0.dist-info/licenses/LICENSE,sha256=-8caMvpCK8SgZ5LlRKhGCMtYDEXqTKH9X8pFEhl91_4,1066
+kreuzberg-3.14.0.dist-info/RECORD,,