kreuzberg 3.11.4__py3-none-any.whl → 3.13.0__py3-none-any.whl

kreuzberg/_ocr/_tesseract.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import csv
 import hashlib
 import io
 import os
@@ -7,26 +8,33 @@ import re
 import subprocess
 import sys
 import tempfile
-from dataclasses import dataclass
-from enum import Enum
+from io import StringIO
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, ClassVar, Final
 
 import anyio
+import html_to_markdown
+import polars as pl
 from anyio import Path as AsyncPath
 from anyio import run_process
+from bs4 import BeautifulSoup
+from bs4.element import Tag
 from PIL import Image
+from PIL.Image import Image as PILImage
 from typing_extensions import Self
 
-from kreuzberg._mime_types import PLAIN_TEXT_MIME_TYPE
+from kreuzberg._mime_types import HTML_MIME_TYPE, MARKDOWN_MIME_TYPE, PLAIN_TEXT_MIME_TYPE
 from kreuzberg._ocr._base import OCRBackend
-from kreuzberg._types import ExtractionResult
+from kreuzberg._ocr._table_extractor import extract_words, reconstruct_table, to_markdown
+from kreuzberg._types import ExtractionResult, HTMLToMarkdownConfig, PSMMode, TableData, TesseractConfig
+from kreuzberg._utils._cache import get_ocr_cache
 from kreuzberg._utils._string import normalize_spaces
 from kreuzberg._utils._sync import run_sync
 from kreuzberg._utils._tmp import create_temp_file
 from kreuzberg.exceptions import MissingDependencyError, OCRError, ValidationError
 
 if TYPE_CHECKING:
+    from bs4.element import Tag
     from PIL.Image import Image as PILImage
 
 try:  # pragma: no cover
@@ -168,68 +176,6 @@ TESSERACT_SUPPORTED_LANGUAGE_CODES: Final[set[str]] = {
 MINIMAL_SUPPORTED_TESSERACT_VERSION: Final[int] = 5
 
 
-class PSMMode(Enum):
-    """Enum for Tesseract Page Segmentation Modes (PSM) with human-readable values."""
-
-    OSD_ONLY = 0
-    """Orientation and script detection only."""
-    AUTO_OSD = 1
-    """Automatic page segmentation with orientation and script detection."""
-    AUTO_ONLY = 2
-    """Automatic page segmentation without OSD."""
-    AUTO = 3
-    """Fully automatic page segmentation (default)."""
-    SINGLE_COLUMN = 4
-    """Assume a single column of text."""
-    SINGLE_BLOCK_VERTICAL = 5
-    """Assume a single uniform block of vertically aligned text."""
-    SINGLE_BLOCK = 6
-    """Assume a single uniform block of text."""
-    SINGLE_LINE = 7
-    """Treat the image as a single text line."""
-    SINGLE_WORD = 8
-    """Treat the image as a single word."""
-    CIRCLE_WORD = 9
-    """Treat the image as a single word in a circle."""
-    SINGLE_CHAR = 10
-    """Treat the image as a single character."""
-
-
-@dataclass(unsafe_hash=True, frozen=True, slots=True)
-class TesseractConfig:
-    """Configuration options for Tesseract OCR engine."""
-
-    classify_use_pre_adapted_templates: bool = True
-    """Whether to use pre-adapted templates during classification to improve recognition accuracy."""
-    language: str = "eng"
-    """Language code to use for OCR.
-    Examples:
-            -   'eng' for English
-            -   'deu' for German
-            -    multiple languages combined with '+', e.g. 'eng+deu')
-    """
-    language_model_ngram_on: bool = False
-    """Enable or disable the use of n-gram-based language models for improved text recognition.
-
-    Default is False for optimal performance on modern documents. Enable for degraded or historical text."""
-    psm: PSMMode = PSMMode.AUTO
-    """Page segmentation mode (PSM) to guide Tesseract on how to segment the image (e.g., single block, single line)."""
-    tessedit_dont_blkrej_good_wds: bool = True
-    """If True, prevents block rejection of words identified as good, improving text output quality."""
-    tessedit_dont_rowrej_good_wds: bool = True
-    """If True, prevents row rejection of words identified as good, avoiding unnecessary omissions."""
-    tessedit_enable_dict_correction: bool = True
-    """Enable or disable dictionary-based correction for recognized text to improve word accuracy."""
-    tessedit_char_whitelist: str = ""
-    """Whitelist of characters that Tesseract is allowed to recognize. Empty string means no restriction."""
-    tessedit_use_primary_params_model: bool = True
-    """If True, forces the use of the primary parameters model for text recognition."""
-    textord_space_size_is_variable: bool = True
-    """Allow variable spacing between words, useful for text with irregular spacing."""
-    thresholding_method: bool = False
-    """Enable or disable specific thresholding methods during image preprocessing for better OCR accuracy."""
-
-
 class TesseractBackend(OCRBackend[TesseractConfig]):
     _version_checked: ClassVar[bool] = False
 
@@ -238,10 +184,14 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
         image: PILImage,
         **kwargs: Unpack[TesseractConfig],
     ) -> ExtractionResult:
-        from kreuzberg._utils._cache import get_ocr_cache  # noqa: PLC0415
+        use_cache = kwargs.pop("use_cache", True)
+
+        save_image = image
+        if image.mode not in ("RGB", "RGBA", "L", "LA", "P", "1"):
+            save_image = image.convert("RGB")
 
         image_buffer = io.BytesIO()
-        await run_sync(image.save, image_buffer, format="PNG")
+        await run_sync(save_image.save, image_buffer, format="PNG")
         image_content = image_buffer.getvalue()
 
         cache_kwargs = {
@@ -250,7 +200,40 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
             "ocr_config": str(sorted(kwargs.items())),
         }
 
+        if use_cache:
+            cached_result = await self._handle_cache_lookup(cache_kwargs)
+            if cached_result:
+                return cached_result
+
+        ocr_cache = get_ocr_cache()
+        try:
+            await self._validate_tesseract_version()
+            image_path, unlink = await create_temp_file(".png")
+
+            try:
+                await run_sync(save_image.save, str(image_path), format="PNG")
+            except OSError as e:
+                if "cannot write mode" not in str(e):
+                    raise
+                save_image = image.convert("RGB")
+                await run_sync(save_image.save, str(image_path), format="PNG")
+            try:
+                result = await self.process_file(image_path, **kwargs)
+
+                if use_cache:
+                    await ocr_cache.aset(result, **cache_kwargs)
+
+                return result
+            finally:
+                await unlink()
+        finally:
+            if use_cache:
+                ocr_cache.mark_complete(**cache_kwargs)
+
+    async def _handle_cache_lookup(self, cache_kwargs: dict[str, Any]) -> ExtractionResult | None:
+        """Handle cache lookup before processing."""
         ocr_cache = get_ocr_cache()
+
         cached_result = await ocr_cache.aget(**cache_kwargs)
         if cached_result is not None:
             return cached_result
@@ -258,49 +241,123 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
         if ocr_cache.is_processing(**cache_kwargs):
             event = ocr_cache.mark_processing(**cache_kwargs)
             await anyio.to_thread.run_sync(event.wait)
-
-            # Try cache again after waiting for other process to complete  # ~keep
             cached_result = await ocr_cache.aget(**cache_kwargs)
             if cached_result is not None:
                 return cached_result
 
         ocr_cache.mark_processing(**cache_kwargs)
+        return None
+
+    def _prepare_tesseract_run_config(self, **kwargs: Any) -> dict[str, Any]:
+        """Prepare configuration for a Tesseract run."""
+        language = self._validate_language_code(kwargs.pop("language", "eng"))
+        psm = kwargs.pop("psm", PSMMode.AUTO)
+        output_format = kwargs.pop("output_format", "markdown")
+        enable_table_detection = kwargs.pop("enable_table_detection", False)
+
+        if enable_table_detection and output_format == "text":
+            output_format = "tsv"
+
+        if output_format == "markdown":
+            tesseract_format = "hocr"
+            ext = ".hocr"
+        elif output_format == "tsv":
+            tesseract_format = "tsv"
+            ext = ".tsv"
+        elif output_format == "hocr":
+            tesseract_format = "hocr"
+            ext = ".hocr"
+        else:
+            tesseract_format = "text"
+            ext = ".txt"
+
+        return {
+            "language": language,
+            "psm": psm,
+            "output_format": output_format,
+            "enable_table_detection": enable_table_detection,
+            "tesseract_format": tesseract_format,
+            "ext": ext,
+            "remaining_kwargs": kwargs,
+        }
+
+    async def _execute_tesseract(self, path: Path, output_base: str, run_config: dict[str, Any]) -> None:
+        """Build and execute the Tesseract command."""
+        command = [
+            "tesseract",
+            str(path),
+            output_base,
+            "-l",
+            run_config["language"],
+            "--psm",
+            str(run_config["psm"].value),
+            "--oem",
+            "1",
+            "--loglevel",
+            "OFF",
+        ]
+
+        if run_config["tesseract_format"] != "text":
+            command.append(run_config["tesseract_format"])
+
+        for kwarg, value in run_config["remaining_kwargs"].items():
+            if kwarg.startswith("table_"):
+                continue
+            if isinstance(value, bool):
+                command.extend(["-c", f"{kwarg}={1 if value else 0}"])
+            else:
+                command.extend(["-c", f"{kwarg}={value}"])
+
+        env: dict[str, Any] | None = None
+        if sys.platform.startswith("linux"):
+            env = {"OMP_THREAD_LIMIT": "1"}
 
         try:
-            await self._validate_tesseract_version()
-            image_path, unlink = await create_temp_file(".png")
-            await run_sync(image.save, str(image_path), format="PNG")
-            try:
-                result = await self.process_file(image_path, **kwargs)
+            result = await run_process(command, env=env)
+            if not result.returncode == 0:
+                raise OCRError(
+                    "OCR failed with a non-0 return code.",
+                    context={"error": result.stderr.decode() if isinstance(result.stderr, bytes) else result.stderr},
+                )
+        except subprocess.CalledProcessError as e:
+            error_msg = e.stderr.decode("utf-8") if e.stderr else str(e)
+            raise OCRError(
+                f"Failed to OCR using tesseract: {error_msg}",
+                context={"command": command, "returncode": e.returncode, "error": error_msg},
+            ) from e
 
-                await ocr_cache.aset(result, **cache_kwargs)
+    async def _process_tesseract_output(self, output: str, run_config: dict[str, Any]) -> ExtractionResult:
+        """Process the raw output from Tesseract based on the requested format."""
+        output_format = run_config["output_format"]
+        enable_table_detection = run_config["enable_table_detection"]
+        kwargs = run_config["remaining_kwargs"]
+
+        if output_format == "markdown":
+            return await self._process_hocr_to_markdown(output, enable_table_detection=enable_table_detection, **kwargs)
+        if output_format == "tsv" and enable_table_detection:
+            return await self._process_tsv_output(
+                output,
+                table_column_threshold=kwargs.get("table_column_threshold", 20),
+                table_row_threshold_ratio=kwargs.get("table_row_threshold_ratio", 0.5),
+                table_min_confidence=kwargs.get("table_min_confidence", 30.0),
+            )
+        if output_format == "tsv":
+            return self._extract_text_from_tsv(output)
+        if output_format == "hocr":
+            return ExtractionResult(content=output, mime_type=HTML_MIME_TYPE, metadata={}, chunks=[])
 
-                return result
-            finally:
-                await unlink()
-        finally:
-            ocr_cache.mark_complete(**cache_kwargs)
+        return ExtractionResult(
+            content=normalize_spaces(output), mime_type=PLAIN_TEXT_MIME_TYPE, metadata={}, chunks=[]
+        )
 
-    async def process_file(
-        self,
-        path: Path,
-        **kwargs: Unpack[TesseractConfig],
-    ) -> ExtractionResult:
-        from kreuzberg._utils._cache import get_ocr_cache  # noqa: PLC0415
+    async def process_file(self, path: Path, **kwargs: Unpack[TesseractConfig]) -> ExtractionResult:
+        use_cache = kwargs.pop("use_cache", True)
 
         try:
             stat = path.stat()
-            file_info = {
-                "path": str(path.resolve()),
-                "size": stat.st_size,
-                "mtime": stat.st_mtime,
-            }
+            file_info = {"path": str(path.resolve()), "size": stat.st_size, "mtime": stat.st_mtime}
         except OSError:
-            file_info = {
-                "path": str(path),
-                "size": 0,
-                "mtime": 0,
-            }
+            file_info = {"path": str(path), "size": 0, "mtime": 0}
 
         cache_kwargs = {
             "file_info": str(sorted(file_info.items())),
@@ -308,71 +365,37 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
             "ocr_config": str(sorted(kwargs.items())),
         }
 
-        ocr_cache = get_ocr_cache()
-        cached_result = await ocr_cache.aget(**cache_kwargs)
-        if cached_result is not None:
-            return cached_result
-
-        if ocr_cache.is_processing(**cache_kwargs):
-            event = ocr_cache.mark_processing(**cache_kwargs)
-            await anyio.to_thread.run_sync(event.wait)
-
-            # Try cache again after waiting for other process to complete  # ~keep
-            cached_result = await ocr_cache.aget(**cache_kwargs)
-            if cached_result is not None:
+        if use_cache:
+            cached_result = await self._handle_cache_lookup(cache_kwargs)
+            if cached_result:
                 return cached_result
 
-        ocr_cache.mark_processing(**cache_kwargs)
-
+        ocr_cache = get_ocr_cache()
         try:
             await self._validate_tesseract_version()
-            output_path, unlink = await create_temp_file(".txt")
-            language = self._validate_language_code(kwargs.pop("language", "eng"))
-            psm = kwargs.pop("psm", PSMMode.AUTO)
-            try:
-                output_base = str(output_path).replace(".txt", "")
-                command = [
-                    "tesseract",
-                    str(path),
-                    output_base,
-                    "-l",
-                    language,
-                    "--psm",
-                    str(psm.value),
-                    "--oem",
-                    "1",
-                    "--loglevel",
-                    "OFF",
-                ]
-                for kwarg, value in kwargs.items():
-                    if isinstance(value, bool):
-                        command.extend(["-c", f"{kwarg}={1 if value else 0}"])
-                    else:
-                        # Handle string parameters (like tessedit_char_whitelist)
-                        command.extend(["-c", f"{kwarg}={value}"])
-
-                env: dict[str, Any] | None = None
-                if sys.platform.startswith("linux"):
-                    env = {"OMP_THREAD_LIMIT": "1"}
 
-                result = await run_process(command, env=env)
+            run_config = self._prepare_tesseract_run_config(**kwargs)
+            output_path, unlink = await create_temp_file(run_config["ext"])
 
-                if not result.returncode == 0:
-                    raise OCRError(
-                        "OCR failed with a non-0 return code.",
-                        context={
-                            "error": result.stderr.decode() if isinstance(result.stderr, bytes) else result.stderr
-                        },
-                    )
+            try:
+                output_base = str(output_path).replace(run_config["ext"], "")
+                await self._execute_tesseract(path, output_base, run_config)
 
                 output = await AsyncPath(output_path).read_text("utf-8")
-                extraction_result = ExtractionResult(
-                    content=normalize_spaces(output), mime_type=PLAIN_TEXT_MIME_TYPE, metadata={}, chunks=[]
-                )
-
-                final_cache_kwargs = cache_kwargs.copy()
-                final_cache_kwargs["ocr_config"] = str(sorted({**kwargs, "language": language, "psm": psm}.items()))
-                await ocr_cache.aset(extraction_result, **final_cache_kwargs)
+                extraction_result = await self._process_tesseract_output(output, run_config)
+
+                if use_cache:
+                    final_cache_kwargs = cache_kwargs.copy()
+                    final_cache_kwargs["ocr_config"] = str(
+                        sorted(
+                            {
+                                **run_config["remaining_kwargs"],
+                                "language": run_config["language"],
+                                "psm": run_config["psm"],
+                            }.items()
+                        )
+                    )
+                    await ocr_cache.aset(extraction_result, **final_cache_kwargs)
 
                 return extraction_result
             except (RuntimeError, OSError) as e:
@@ -380,7 +403,562 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
             finally:
                 await unlink()
         finally:
-            ocr_cache.mark_complete(**cache_kwargs)
+            if use_cache:
+                ocr_cache.mark_complete(**cache_kwargs)
+
+    async def _process_tsv_output(
+        self,
+        tsv_content: str,
+        table_column_threshold: int = 20,
+        table_row_threshold_ratio: float = 0.5,
+        table_min_confidence: float = 30.0,
+    ) -> ExtractionResult:
+        """Process TSV output and extract tables if detected.
+
+        Args:
+            tsv_content: Raw TSV output from Tesseract.
+            table_column_threshold: Pixel threshold for column clustering.
+            table_row_threshold_ratio: Row threshold as ratio of mean text height.
+            table_min_confidence: Minimum confidence score to include a word.
+
+        Returns:
+            ExtractionResult with extracted content and tables.
+        """
+        text_result = self._extract_text_from_tsv(tsv_content)
+
+        try:
+            if (
+                (words := extract_words(tsv_content, min_confidence=table_min_confidence))
+                and (
+                    table_data := reconstruct_table(
+                        words,
+                        column_threshold=table_column_threshold,
+                        row_threshold_ratio=table_row_threshold_ratio,
+                    )
+                )
+                and len(table_data) > 1
+            ):
+                markdown = to_markdown(table_data)
+
+                try:
+                    df = await run_sync(pl.DataFrame, table_data[1:], schema=table_data[0])
+                except (ImportError, IndexError):
+                    df = None
+
+                table: TableData = {"text": markdown, "df": df, "page_number": 1, "cropped_image": None}  # type: ignore[typeddict-item]
+
+                return ExtractionResult(
+                    content=text_result.content,
+                    mime_type=text_result.mime_type,
+                    metadata=text_result.metadata,
+                    tables=[table],
+                    chunks=text_result.chunks,
+                )
+        except (ValueError, KeyError, ImportError):
+            pass
+
+        return text_result
+
+    def _extract_text_from_tsv(self, tsv_content: str) -> ExtractionResult:
+        """Extract plain text from TSV output.
+
+        Args:
+            tsv_content: Raw TSV output from Tesseract.
+
+        Returns:
+            ExtractionResult with extracted text.
+        """
+        try:
+            reader = csv.DictReader(StringIO(tsv_content), delimiter="\t")
+
+            lines: dict[tuple[int, int, int, int], list[tuple[int, str]]] = {}
+
+            for row in reader:
+                if row.get("level") == "5" and row.get("text", "").strip():
+                    line_key = (int(row["page_num"]), int(row["block_num"]), int(row["par_num"]), int(row["line_num"]))
+
+                    if line_key not in lines:
+                        lines[line_key] = []
+
+                    lines[line_key].append((int(row["left"]), row["text"]))
+
+            text_parts: list[str] = []
+            last_block = -1
+            last_para = -1
+
+            for line_key in sorted(lines.keys()):
+                page_num, block_num, par_num, line_num = line_key
+
+                if block_num != last_block:
+                    if text_parts:  # ~keep
+                        text_parts.append("\n\n")
+                    last_block = block_num
+                    last_para = par_num
+                elif par_num != last_para:
+                    text_parts.append("\n\n")
+                    last_para = par_num
+
+                words = sorted(lines[line_key], key=lambda x: x[0])
+                line_text = " ".join(word[1] for word in words)
+                text_parts.append(line_text)
+                text_parts.append("\n")
+
+            content = "".join(text_parts).strip()
+
+        except (ValueError, KeyError):
+            content = ""
+            for line in tsv_content.split("\n")[1:]:  # ~keep skip header
+                parts = line.split("\t")
+                if len(parts) > 11 and parts[11].strip():  # ~keep text is in column 11
+                    content += parts[11] + " "
+            content = content.strip()
+
+        return ExtractionResult(
+            content=normalize_spaces(content), mime_type=PLAIN_TEXT_MIME_TYPE, metadata={}, chunks=[]
+        )
+
+    async def _process_hocr_to_markdown(
+        self,
+        hocr_content: str,
+        enable_table_detection: bool = False,
+        html_to_markdown_config: HTMLToMarkdownConfig | None = None,
+        table_column_threshold: int = 20,
+        table_row_threshold_ratio: float = 0.5,
+        table_min_confidence: float = 30.0,
+        **_kwargs: Any,
+    ) -> ExtractionResult:
+        """Convert hOCR content to Markdown with table detection.
+
+        Args:
+            hocr_content: Raw hOCR HTML/XML content from Tesseract.
+            enable_table_detection: Whether to detect and format tables.
+            html_to_markdown_config: Configuration for HTML to Markdown conversion.
+            table_column_threshold: Pixel threshold for column clustering.
+            table_row_threshold_ratio: Row threshold as ratio of mean text height.
+            table_min_confidence: Minimum confidence score to include a word.
+            **kwargs: Additional configuration options.
+
+        Returns:
+            ExtractionResult with Markdown content and detected tables.
+        """
+        config = html_to_markdown_config or HTMLToMarkdownConfig(
+            escape_asterisks=False,
+            escape_underscores=False,
+            extract_metadata=False,
+            strip="meta title",
+        )
+
+        tables: list[TableData] = []
+        if enable_table_detection:
+            soup = BeautifulSoup(hocr_content, "lxml")
+            tables = await self._extract_tables_from_hocr(
+                soup,
+                table_column_threshold,
+                table_row_threshold_ratio,
+                table_min_confidence,
+            )
+
+        hocr_converters = self._create_hocr_converters(tables)
+
+        all_converters = dict(hocr_converters)
+        if config.custom_converters:
+            all_converters.update(config.custom_converters)
+
+        config_dict = config.to_dict()
+        config_dict["custom_converters"] = all_converters
+
+        try:
+            markdown_content = html_to_markdown.convert_to_markdown(hocr_content, **config_dict)
+            markdown_content = normalize_spaces(markdown_content)
+        except (ValueError, TypeError, AttributeError):
+            try:
+                soup = BeautifulSoup(hocr_content, "lxml")
+                words = soup.find_all("span", class_="ocrx_word")
+                text_parts = []
+                for word in words:
+                    text = word.get_text().strip()
+                    if text:
+                        text_parts.append(text)
+
+                if text_parts:
+                    markdown_content = " ".join(text_parts)
+                else:
+                    markdown_content = soup.get_text().strip() or "[No text detected]"
+
+                markdown_content = normalize_spaces(markdown_content)
+            except (ValueError, TypeError, AttributeError):
+                markdown_content = "[OCR processing failed]"
+
+        if tables:
+            table_sections = []
+            for i, table in enumerate(tables):
+                table_sections.append(f"\n## Table {i + 1}\n\n{table['text']}\n")
+
+            if markdown_content.strip():
+                final_content = f"{markdown_content}\n{''.join(table_sections)}"
+            else:
+                final_content = "".join(table_sections).strip()
+        else:
+            final_content = markdown_content
+
+        return ExtractionResult(
+            content=final_content,
+            mime_type=MARKDOWN_MIME_TYPE,
+            metadata={"source_format": "hocr", "tables_detected": len(tables)},
+            chunks=[],
+            tables=tables,
+        )
+
+    def _create_basic_converters(self) -> dict[str, Any]:
+        """Create basic converters for individual hOCR elements."""
+
+        def ocrx_word_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Custom converter for hOCR word elements - adds spaces between words."""
+            del tag
+            return f"{text.strip()} "
+
+        def ocr_line_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Custom converter for hOCR line elements - handles line breaks."""
+            del tag
+            return f"{text.strip()}\n"
+
+        def ocr_par_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Custom converter for hOCR paragraph elements - handles paragraph breaks."""
+            del tag
+            content = text.strip()
+            if not content:
+                return ""
+            return f"{content}\n\n"
+
+        def ocr_carea_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Custom converter for hOCR content area elements."""
+            del tag
+            content = text.strip()
+            if not content:
+                return ""
+            return f"{content}\n\n"
+
+        def ocr_page_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Custom converter for hOCR page elements."""
+            del tag
+            return text.strip()
+
+        def ocr_separator_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Custom converter for hOCR separator elements - convert to horizontal rules."""
+            del tag, text
+            return "---\n"
+
+        def ocr_photo_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Custom converter for hOCR photo/image elements - indicate image presence."""
+            del text
+            title = tag.get("title", "")
+            if isinstance(title, str):
+                bbox_match = re.search(r"bbox (\d+) (\d+) (\d+) (\d+)", title)
+                if bbox_match:
+                    x0, y0, x1, y1 = bbox_match.groups()
+                    width = int(x1) - int(x0)
+                    height = int(y1) - int(y0)
+                    return f"*[Image region: {width}x{height} pixels]*\n\n"
+            return "*[Image detected]*\n\n"
+
+        return {
+            "ocrx_word": ocrx_word_converter,
+            "ocr_line": ocr_line_converter,
+            "ocr_par": ocr_par_converter,
+            "ocr_carea": ocr_carea_converter,
+            "ocr_page": ocr_page_converter,
+            "ocr_separator": ocr_separator_converter,
+            "ocr_photo": ocr_photo_converter,
+        }
+
+    def _create_hocr_converters(self, _tables: list[TableData]) -> dict[str, Any]:
+        """Create custom converters for hOCR elements that preserve spacing.
+
+        Args:
+            tables: List of detected tables (not used for filtering, tables added separately).
+
+        Returns:
+            Dictionary mapping HTML tags to converter functions.
+        """
+        basic_converters = self._create_basic_converters()
+
+        def generic_div_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Generic converter for div elements based on class."""
+            class_attr = tag.get("class", "")
+            if isinstance(class_attr, list):
+                class_attr = " ".join(class_attr)
+            elif not isinstance(class_attr, str):
+                class_attr = ""
+
+            for class_name in ["ocr_separator", "ocr_photo", "ocr_page", "ocr_carea"]:
+                if class_name in class_attr:
+                    converter_result = basic_converters[class_name](tag=tag, text=text, **_conv_kwargs)
+                    return str(converter_result)
+            return text
+
+        def generic_span_converter(*, tag: Tag, text: str, **_conv_kwargs: Any) -> str:
+            """Generic converter for span elements based on class."""
+            class_attr = tag.get("class", "")
+            if isinstance(class_attr, list):
+                class_attr = " ".join(class_attr)
+            elif not isinstance(class_attr, str):
+                class_attr = ""
+
+            for class_name in ["ocrx_word", "ocr_line"]:
+                if class_name in class_attr:
+                    converter_result = basic_converters[class_name](tag=tag, text=text, **_conv_kwargs)
+                    return str(converter_result)
+            return f"{text.strip()} "
+
+        return {
+            "span": generic_span_converter,
+            "div": generic_div_converter,
+            "p": basic_converters["ocr_par"],
+        }
+
+    def _process_hocr_to_markdown_sync(self, hocr_content: str, config: TesseractConfig) -> ExtractionResult:
+        """Synchronously process hOCR content to markdown format.
+
+        Args:
+            hocr_content: Raw hOCR content as string
+            config: Tesseract configuration object
+
+        Returns:
+            ExtractionResult with markdown content
+        """
+        tables: list[TableData] = []
+
+        if config.enable_table_detection:
+            pass
+
+        try:
+            converters = self._create_hocr_converters(tables)
+
+            html_config = HTMLToMarkdownConfig(
+                custom_converters=converters,
+                escape_asterisks=False,
+                escape_underscores=False,
+                extract_metadata=False,
+                strip="meta title",
+            )
+
+            markdown_content = html_to_markdown.convert_to_markdown(
+                hocr_content,
+                **html_config.to_dict(),
+            )
+
+            markdown_content = normalize_spaces(markdown_content)
+
+        except (ValueError, TypeError, AttributeError):
+            try:
+                soup = BeautifulSoup(hocr_content, "lxml")
+                words = soup.find_all("span", class_="ocrx_word")
+                text_parts = []
+                for word in words:
+                    text = word.get_text().strip()
+                    if text:
+                        text_parts.append(text)
+
+                if text_parts:
+                    markdown_content = " ".join(text_parts)
+                else:
+                    markdown_content = soup.get_text().strip() or "[No text detected]"
+
+                markdown_content = normalize_spaces(markdown_content)
+            except (ValueError, TypeError, AttributeError):
+                markdown_content = "[OCR processing failed]"
+
+        if tables:
+            table_sections = []
+            for i, table in enumerate(tables):
+                table_sections.append(f"\n## Table {i + 1}\n\n{table['text']}\n")
+
+            if markdown_content.strip():
+                final_content = f"{markdown_content}\n{''.join(table_sections)}"
+            else:
+                final_content = "".join(table_sections).strip()
+        else:
+            final_content = markdown_content
+
+        return ExtractionResult(
+            content=final_content,
+            mime_type=MARKDOWN_MIME_TYPE,
+            metadata={"source_format": "hocr", "tables_detected": len(tables)},
+            chunks=[],
+            tables=tables,
+        )
+
+    def _process_tsv_output_sync(
+        self,
+        tsv_content: str,
+        table_column_threshold: int = 20,
+        table_row_threshold_ratio: float = 0.5,
+        table_min_confidence: float = 30.0,
+    ) -> ExtractionResult:
+        """Synchronously process TSV output and extract tables if detected.
+
+        Args:
+            tsv_content: Raw TSV output from Tesseract.
+            table_column_threshold: Pixel threshold for column clustering.
+            table_row_threshold_ratio: Row threshold as ratio of mean text height.
+            table_min_confidence: Minimum confidence score to include a word.
+
+        Returns:
+            ExtractionResult with extracted content and tables.
+        """
+        text_result = self._extract_text_from_tsv(tsv_content)
+
+        try:
+            if (
+                (words := extract_words(tsv_content, min_confidence=table_min_confidence))
+                and (
+                    table_data := reconstruct_table(
+                        words,
+                        column_threshold=table_column_threshold,
+                        row_threshold_ratio=table_row_threshold_ratio,
+                    )
+                )
+                and len(table_data) > 1
+            ):
+                markdown = to_markdown(table_data)
+
+                try:
+                    df = pl.DataFrame(table_data[1:], schema=table_data[0])
+                except (ImportError, IndexError):
+                    df = None
+
+                table: TableData = {"text": markdown, "df": df, "page_number": 1, "cropped_image": None}  # type: ignore[typeddict-item]
+
+                return ExtractionResult(
+                    content=text_result.content,
+                    mime_type=text_result.mime_type,
+                    metadata=text_result.metadata,
+                    tables=[table],
+                    chunks=text_result.chunks,
+                )
+        except (ValueError, KeyError, ImportError):
+            pass
+
+        return text_result
+
+    async def _extract_tables_from_hocr(
+        self,
+        soup: Any,
+        column_threshold: int = 20,
+        row_threshold_ratio: float = 0.5,
+        min_confidence: float = 30.0,
+    ) -> list[TableData]:
+        """Extract tables from hOCR structure using coordinate analysis.
+
+        Args:
+            soup: Parsed hOCR BeautifulSoup object.
+            column_threshold: Pixel threshold for column clustering.
+            row_threshold_ratio: Row threshold as ratio of mean text height.
+            min_confidence: Minimum confidence score to include a word.
+
+        Returns:
+            List of detected tables as TableData objects.
+        """
+        tsv_data = await self._hocr_to_tsv_data(soup, min_confidence)
+
+        if not tsv_data:
+            return []
+
+        if not (words := extract_words(tsv_data, min_confidence=min_confidence)):
+            return []
+
+        tables: list[TableData] = []
+        try:
+            table_data = reconstruct_table(
+                words,
+                column_threshold=column_threshold,
+                row_threshold_ratio=row_threshold_ratio,
+            )
+            if table_data and len(table_data) > 1:  # ~keep At least header + one data row
+                markdown = to_markdown(table_data)
+
+                min_x = min(w["left"] for w in words)
+                max_x = max(w["left"] + w["width"] for w in words)
+                min_y = min(w["top"] for w in words)
+                max_y = max(w["top"] + w["height"] for w in words)
+
+                try:
+                    df = await run_sync(pl.DataFrame, table_data[1:], schema=table_data[0])
+                except (ImportError, IndexError):
+                    df = None
+
+                dummy_image = Image.new("RGB", (1, 1), "white")
+
+                table: TableData = {
+                    "text": markdown,
+                    "df": df,
+                    "page_number": 1,
+                    "cropped_image": dummy_image,
+                    "metadata": {"bbox": (min_x, min_y, max_x, max_y)},
+                }  # type: ignore[typeddict-unknown-key]
+                tables.append(table)
+        except (ValueError, KeyError, ImportError):
+            pass
+
+        return tables
+
+    async def _hocr_to_tsv_data(self, soup: Any, min_confidence: float) -> str:
+        """Convert hOCR structure to TSV format for table extraction.
+
+        Args:
+            soup: Parsed hOCR BeautifulSoup object.
+            min_confidence: Minimum confidence score to include.
+
+        Returns:
+            TSV formatted string compatible with table extractor.
+        """
+        tsv_lines = ["level\tpage_num\tblock_num\tpar_num\tline_num\tword_num\tleft\ttop\twidth\theight\tconf\ttext"]
+
+        words = soup.find_all("span", class_="ocrx_word")
+        word_num = 1
+
+        for word in words:
+            title = word.get("title", "")
+            text = word.get_text().strip()
+
+            if not text:
+                continue
+
+            bbox_match = re.search(r"bbox (\d+) (\d+) (\d+) (\d+)", title)
+            if not bbox_match:
+                continue
+
+            x0, y0, x1, y1 = map(int, bbox_match.groups())
+
+            conf_match = re.search(r"x_wconf (\d+)", title)
+            confidence = float(conf_match.group(1)) if conf_match else 100.0
+
+            if confidence < min_confidence:
+                continue
+
+            line = word.find_parent(class_="ocr_line")
+            par = word.find_parent(class_="ocr_par")
+            block = word.find_parent(class_="ocr_carea")
+
+            tsv_line = f"5\t1\t{block.get('id', '1').split('_')[-1] if block else 1}\t{par.get('id', '1').split('_')[-1] if par else 1}\t{line.get('id', '1').split('_')[-1] if line else 1}\t{word_num}\t{x0}\t{y0}\t{x1 - x0}\t{y1 - y0}\t{confidence}\t{text}"
+            tsv_lines.append(tsv_line)
+            word_num += 1
+
+        return "\n".join(tsv_lines)
+
+    def _identify_table_regions(self, words: list[dict[str, Any]]) -> list[list[dict[str, Any]]]:
+        """Identify potential table regions from word coordinates.
+
+        Args:
+            words: List of word dictionaries with coordinates.
+
+        Returns:
+            List of word groups representing potential tables.
+        """
+        if not words:
+            return []
+
+        return [words]
 
     @classmethod
     async def _validate_tesseract_version(cls) -> None:
@@ -394,8 +972,14 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
                 return
 
             command = ["tesseract", "--version"]
-            result = await run_process(command)
-            version_match = re.search(r"tesseract\s+v?(\d+)\.\d+\.\d+", result.stdout.decode())
+            env = {"OMP_THREAD_LIMIT": "1"} if sys.platform.startswith("linux") else None
+            try:
+                result = await run_process(command, env=env)
+            except (subprocess.CalledProcessError, FileNotFoundError) as e:
+                raise MissingDependencyError(
+                    "Tesseract version 5 is a required system dependency. Please install it on your system and make sure its available in $PATH."
+                ) from e
+            version_match = re.search(r"tesseract\s+v?(\d+)\.\d+\.\d+", result.stdout.decode("utf-8"))
             if not version_match or int(version_match.group(1)) < MINIMAL_SUPPORTED_TESSERACT_VERSION:
                 raise MissingDependencyError(
                     "Tesseract version 5 is a required system dependency. Please install it on your system and make sure its available in $PATH."
@@ -407,33 +991,10 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
                 "Tesseract version 5 is a required system dependency. Please install it on your system and make sure its available in $PATH."
             ) from e
 
-    def process_image_sync(
-        self,
-        image: PILImage,
-        **kwargs: Unpack[TesseractConfig],
-    ) -> ExtractionResult:
-        """Synchronously process an image and extract its text and metadata.
-
-        Args:
-            image: An instance of PIL.Image representing the input image.
-            **kwargs: Any kwargs related to the given backend
-
-        Returns:
-            The extraction result object
-        """
-        from kreuzberg._utils._cache import get_ocr_cache  # noqa: PLC0415
-
-        image_buffer = io.BytesIO()
-        image.save(image_buffer, format="PNG")
-        image_content = image_buffer.getvalue()
-
-        cache_kwargs = {
-            "image_hash": hashlib.sha256(image_content).hexdigest()[:16],
-            "ocr_backend": "tesseract",
-            "ocr_config": str(sorted(kwargs.items())),
-        }
-
+    def _handle_cache_lookup_sync(self, cache_kwargs: dict[str, Any]) -> ExtractionResult | None:
+        """Handle cache lookup before processing (sync)."""
         ocr_cache = get_ocr_cache()
+
         cached_result = ocr_cache.get(**cache_kwargs)
         if cached_result is not None:
             return cached_result
@@ -441,46 +1002,113 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
         if ocr_cache.is_processing(**cache_kwargs):
             event = ocr_cache.mark_processing(**cache_kwargs)
             event.wait()
-
-            # Try cache again after waiting for other process to complete
             cached_result = ocr_cache.get(**cache_kwargs)
             if cached_result is not None:
                 return cached_result
 
         ocr_cache.mark_processing(**cache_kwargs)
+        return None
+
+    def _execute_tesseract_sync(self, command: list[str]) -> None:
+        """Run tesseract command synchronously."""
+        env = os.environ.copy()
+        if sys.platform.startswith("linux"):
+            env["OMP_THREAD_LIMIT"] = "1"
+
+        try:
+            subprocess.run(
+                command,
+                check=True,
+                env=env,
+                capture_output=True,
+                text=True,
+                timeout=30,
+                encoding="utf-8",
+            )
+        except subprocess.CalledProcessError as e:
+            error_msg = e.stderr if e.stderr else str(e)
+            raise OCRError(
+                f"Failed to OCR using tesseract: {error_msg}",
+                context={"command": command, "returncode": e.returncode, "error": error_msg},
+            ) from e
+        except subprocess.TimeoutExpired as e:
+            raise OCRError(
+                "Tesseract timed out during processing.",
+                context={"command": command, "timeout": 30},
+            ) from e
+
+    def _process_tesseract_output_sync(self, output: str, run_config: dict[str, Any]) -> ExtractionResult:
+        """Process the raw output from Tesseract based on the requested format (sync)."""
+        output_format = run_config["output_format"]
+        enable_table_detection = run_config["enable_table_detection"]
+        kwargs = run_config["remaining_kwargs"]
+        config = TesseractConfig(**kwargs)
+
+        if output_format == "markdown":
+            return self._process_hocr_to_markdown_sync(output, config)
+        if output_format == "tsv" and enable_table_detection:
+            return self._process_tsv_output_sync(
+                output,
+                table_column_threshold=config.table_column_threshold,
+                table_row_threshold_ratio=config.table_row_threshold_ratio,
+                table_min_confidence=config.table_min_confidence,
+            )
+        if output_format == "tsv":
+            return self._extract_text_from_tsv(output)
+        if output_format == "hocr":
+            return ExtractionResult(content=output, mime_type=HTML_MIME_TYPE, metadata={}, chunks=[])
 
+        return ExtractionResult(
+            content=normalize_spaces(output), mime_type=PLAIN_TEXT_MIME_TYPE, metadata={}, chunks=[]
+        )
+
+    def process_image_sync(self, image: PILImage, **kwargs: Unpack[TesseractConfig]) -> ExtractionResult:
+        """Synchronously process an image and extract its text and metadata."""
+        use_cache = kwargs.pop("use_cache", True)
+
+        save_image = image
+        if image.mode not in ("RGB", "RGBA", "L", "LA", "P", "1"):
+            save_image = image.convert("RGB")
+
+        image_buffer = io.BytesIO()
+        save_image.save(image_buffer, format="PNG")
+        image_content = image_buffer.getvalue()
+
+        cache_kwargs = {
+            "image_hash": hashlib.sha256(image_content).hexdigest()[:16],
+            "ocr_backend": "tesseract",
+            "ocr_config": str(sorted(kwargs.items())),
+        }
+
+        if use_cache:
+            cached_result = self._handle_cache_lookup_sync(cache_kwargs)
+            if cached_result:
+                return cached_result
+
+        ocr_cache = get_ocr_cache()
         try:
             self._validate_tesseract_version_sync()
             with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_file:
                 image_path = Path(tmp_file.name)
-                image.save(str(image_path), format="PNG")
+                save_image.save(str(image_path), format="PNG")
             try:
-                result = self.process_file_sync(image_path, **kwargs)
+                kwargs_with_cache = {**kwargs, "use_cache": use_cache}
+                result = self.process_file_sync(image_path, **kwargs_with_cache)
 
-                ocr_cache.set(result, **cache_kwargs)
+                if use_cache:
+                    ocr_cache.set(result, **cache_kwargs)
 
                 return result
             finally:
                 if image_path.exists():
                     image_path.unlink()
         finally:
-            ocr_cache.mark_complete(**cache_kwargs)
+            if use_cache:
+                ocr_cache.mark_complete(**cache_kwargs)
 
-    def process_file_sync(
-        self,
-        path: Path,
-        **kwargs: Unpack[TesseractConfig],
-    ) -> ExtractionResult:
-        """Synchronously process a file and extract its text and metadata.
-
-        Args:
-            path: A Path object representing the file to be processed.
-            **kwargs: Any kwargs related to the given backend
-
-        Returns:
-            The extraction result object
-        """
-        from kreuzberg._utils._cache import get_ocr_cache  # noqa: PLC0415
+    def process_file_sync(self, path: Path, **kwargs: Unpack[TesseractConfig]) -> ExtractionResult:
+        """Synchronously process a file and extract its text and metadata."""
+        use_cache = kwargs.pop("use_cache", True)
 
         file_info = self._get_file_info(path)
 
@@ -490,53 +1118,74 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
             "ocr_config": str(sorted(kwargs.items())),
         }
 
-        ocr_cache = get_ocr_cache()
-        cached_result = ocr_cache.get(**cache_kwargs)
-        if cached_result is not None:
-            return cached_result
-
-        if ocr_cache.is_processing(**cache_kwargs):
-            event = ocr_cache.mark_processing(**cache_kwargs)
-            event.wait()
-
-            # Try cache again after waiting for other process to complete
-            cached_result = ocr_cache.get(**cache_kwargs)
-            if cached_result is not None:
+        if use_cache:
+            cached_result = self._handle_cache_lookup_sync(cache_kwargs)
+            if cached_result:
                 return cached_result
 
-        ocr_cache.mark_processing(**cache_kwargs)
-
+        ocr_cache = get_ocr_cache()
         try:
             self._validate_tesseract_version_sync()
-            with tempfile.NamedTemporaryFile(suffix=".txt", delete=False) as tmp_file:
-                output_base = tmp_file.name.replace(".txt", "")
-            language = self._validate_language_code(kwargs.pop("language", "eng"))
-            psm = kwargs.pop("psm", PSMMode.AUTO)
+
+            run_config = self._prepare_tesseract_run_config(**kwargs)
+
+            temp_fd, temp_path = tempfile.mkstemp(suffix=run_config["ext"])
+            os.close(temp_fd)
+            Path(temp_path).unlink()
+            output_base = temp_path.replace(run_config["ext"], "")
+
             try:
-                command = self._build_tesseract_command(path, output_base, language, psm, **kwargs)
-                self._run_tesseract_sync(command)
+                command = self._build_tesseract_command(
+                    path,
+                    output_base,
+                    run_config["language"],
+                    run_config["psm"],
+                    run_config["tesseract_format"],
+                    **run_config["remaining_kwargs"],
+                )
+                self._execute_tesseract_sync(command)
+
+                output_path = Path(f"{output_base}{run_config['ext']}")
+                if not output_path.exists():
+                    return ExtractionResult(
+                        content="[OCR processing failed]",
+                        mime_type=PLAIN_TEXT_MIME_TYPE,
+                        metadata={
+                            "source_format": run_config["tesseract_format"],
+                            "error": f"{run_config['ext']} file not generated",
+                        },
+                        chunks=[],
+                        tables=[],
+                    )
 
-                output_path = Path(output_base + ".txt")
                 with output_path.open(encoding="utf-8") as f:
                     output = f.read()
-                extraction_result = ExtractionResult(
-                    content=normalize_spaces(output), mime_type=PLAIN_TEXT_MIME_TYPE, metadata={}, chunks=[]
-                )
 
-                final_cache_kwargs = cache_kwargs.copy()
-                final_cache_kwargs["ocr_config"] = str(sorted({**kwargs, "language": language, "psm": psm}.items()))
-                ocr_cache.set(extraction_result, **final_cache_kwargs)
+                extraction_result = self._process_tesseract_output_sync(output, run_config)
+
+                if use_cache:
+                    final_cache_kwargs = cache_kwargs.copy()
+                    final_cache_kwargs["ocr_config"] = str(
+                        sorted(
+                            {
+                                **run_config["remaining_kwargs"],
+                                "language": run_config["language"],
+                                "psm": run_config["psm"],
+                            }.items()
+                        )
+                    )
+                    ocr_cache.set(extraction_result, **final_cache_kwargs)
 
                 return extraction_result
-            except (RuntimeError, OSError) as e:
-                raise OCRError(f"Failed to OCR using tesseract: {e}") from e
             finally:
-                for ext in [".txt"]:
-                    temp_file = Path(output_base + ext)
-                    if temp_file.exists():
-                        temp_file.unlink()
+                for cleanup_ext in [".txt", ".hocr", ".tsv"]:
+                    cleanup_path = Path(f"{output_base}{cleanup_ext}")
+                    cleanup_path.unlink(missing_ok=True)
+        except Exception as e:
+            raise OCRError(f"Failed to OCR using tesseract: {e}") from e
         finally:
-            ocr_cache.mark_complete(**cache_kwargs)
+            if use_cache:
+                ocr_cache.mark_complete(**cache_kwargs)
 
     def _get_file_info(self, path: Path) -> dict[str, Any]:
         """Get file information for caching."""
@@ -555,7 +1204,7 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
             }
 
     def _build_tesseract_command(
-        self, path: Path, output_base: str, language: str, psm: PSMMode, **kwargs: Any
+        self, path: Path, output_base: str, language: str, psm: PSMMode, output_format: str = "text", **kwargs: Any
     ) -> list[str]:
         """Build tesseract command with all parameters."""
         command = [
@@ -571,34 +1220,19 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
             "--loglevel",
             "OFF",
         ]
+
+        if output_format != "text":
+            command.append(output_format)
+
         for kwarg, value in kwargs.items():
+            if kwarg.startswith("table_"):
+                continue
             if isinstance(value, bool):
                 command.extend(["-c", f"{kwarg}={1 if value else 0}"])
             else:
                 command.extend(["-c", f"{kwarg}={value}"])
         return command
 
-    def _run_tesseract_sync(self, command: list[str]) -> None:
-        """Run tesseract command synchronously."""
-        env = os.environ.copy()
-        if sys.platform.startswith("linux"):
-            env["OMP_THREAD_LIMIT"] = "1"
-
-        result = subprocess.run(
-            command,
-            check=False,
-            env=env,
-            capture_output=True,
-            text=True,
-            timeout=30,
-        )
-
-        if result.returncode != 0:
-            raise OCRError(
-                "OCR failed with a non-0 return code.",
-                context={"error": result.stderr},
-            )
-
     @classmethod
     def _validate_tesseract_version_sync(cls) -> None:
         """Synchronously validate that Tesseract is installed and is version 5 or above.
@@ -611,7 +1245,12 @@ class TesseractBackend(OCRBackend[TesseractConfig]):
                 return
 
             command = ["tesseract", "--version"]
-            result = subprocess.run(command, capture_output=True, text=True, check=False)
+            try:
+                result = subprocess.run(command, capture_output=True, text=True, check=True, encoding="utf-8")
+            except (subprocess.CalledProcessError, FileNotFoundError) as e:
+                raise MissingDependencyError(
+                    "Tesseract version 5 is a required system dependency. Please install it on your system and make sure its available in $PATH."
+                ) from e
             version_match = re.search(r"tesseract\s+v?(\d+)\.\d+\.\d+", result.stdout)
             if not version_match or int(version_match.group(1)) < MINIMAL_SUPPORTED_TESSERACT_VERSION:
                 raise MissingDependencyError(
@@ -721,6 +1360,7 @@ def _process_image_with_tesseract(
                 capture_output=True,
                 text=True,
                 timeout=30,
+                encoding="utf-8",
             )
 
             if result.returncode != 0:
@@ -769,9 +1409,11 @@ def _process_image_bytes_with_tesseract(
         OCR result as dictionary.
     """
     try:
-        with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_image:
-            with Image.open(io.BytesIO(image_bytes)) as image:
-                image.save(tmp_image.name, format="PNG")
+        with (
+            tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_image,
+            Image.open(io.BytesIO(image_bytes)) as image,
+        ):
+            image.save(tmp_image.name, format="PNG")
             image_path = tmp_image.name
 
         try: