pyxtxt 0.3.2__tar.gz → 0.3.4__tar.gz

{pyxtxt-0.3.2/src/pyxtxt.egg-info → pyxtxt-0.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyxtxt
-Version: 0.3.2
+Version: 0.3.4
 Summary: A Python library for extracting text from different types of files (PDF, DOCX, PPTX, XLSX, ODT, ecc.).
 Author-email: Giuseppe Levi <giuseppe.levi@gmail.com>
 License: MIT License
@@ -308,6 +308,45 @@ image_response = requests.get("https://example.com/document.png")
 text = xtxt(image_response.content)
 ```
 
+### AI OCR Confidence Scoring (NEW)
+
+⚠️ **IMPORTANT**: AI-powered OCR is experimental technology that may produce errors, hallucinations, or misinterpretations. Always validate results for critical applications.
+
+The OCR-Ollama system includes confidence scoring to help identify unreliable results:
+
+```python
+from pyxtxt import xtxt_image_with_confidence, set_ollama_config
+
+# Configure confidence threshold (0.0-1.0, default: 0.7)
+set_ollama_config(confidence_threshold=0.8)  # More restrictive
+
+# Get text with confidence score
+text, confidence = xtxt_image_with_confidence("document.png", mode="ocr")
+print(f"Confidence: {confidence:.2f} ({confidence*100:.1f}%)")
+print(f"Text: {text}")
+
+# Check if reliable
+if confidence < 0.7:
+    print("⚠️ Low confidence - result may be unreliable")
+    print("Consider using traditional OCR or manual verification")
+else:
+    print("✅ Good confidence - result likely reliable")
+```
+
+#### Confidence Scoring Features
+
+- **Hallucination Detection**: Penalizes historical/cultural references that often indicate AI misinterpretation
+- **Pattern Recognition**: Rewards structured text (numbers, punctuation, proper formatting)
+- **Uncertainty Detection**: Flags vague language ("appears to be", "seems like", etc.)
+- **Quality Assessment**: Considers content length, repetition, and coherence
+
+#### Common Hallucination Patterns (Automatically Detected)
+- **Historical Content**: "ancient", "medieval", "Egyptian papyrus", "hieroglyphs"
+- **Artistic Interpretations**: "painting", "artwork", "masterpiece", "Renaissance"  
+- **Fantasy Content**: "mystical", "magical", "dragon", "wizard"
+- **Scientific Misinterpretation**: "fossil", "geological formation", "crystal structure"
+- **Vague Language**: "unclear", "difficult to read", "appears to be"
+
 ### Command-Line OCR Example
 
 A complete example script for command-line usage is available:
@@ -324,6 +363,11 @@ with open("ocr_example.py", "wb") as f:
 # python ocr_example.py document.png
 # python ocr_example.py chart.jpg --mode=describe --lang=italian --style=detailed
 # python ocr_example.py diagram.png --mode=describe --length=long --temp=0.3
+
+# NEW: Confidence scoring examples
+# python ocr_example.py suspicious.png --show-confidence
+# python ocr_example.py medical.png --confidence=0.9 --show-confidence
+# python ocr_example.py diagram.png --confidence=0.5 --mode=describe --show-confidence
 ```
 
 The script supports:
@@ -333,6 +377,8 @@ The script supports:
 - **Style control**: descriptive, technical, simple, detailed
 - **Length control**: short, medium, long captions
 - **Temperature**: Adjust LLM creativity (0.0-1.0)
+- **Confidence scoring**: Set threshold and display confidence scores
+- **Quality filtering**: Automatically reject low-confidence results
 
 ### Show Available Formats
 ```python
@@ -376,6 +422,7 @@ text = xtxt(attachment_bytes)
 
 ## ⚠️ Known Limitations
 
+### General Limitations
 - **Legacy file detection**: When using raw streams without filenames, legacy files (.doc, .xls, .ppt) may not be correctly detected due to identical file signatures in libmagic
 - **Filename hints recommended**: When available, providing original filenames improves detection accuracy
 - **MSWrite .doc files**: Require `antiword` installation:
@@ -383,6 +430,41 @@ text = xtxt(attachment_bytes)
   sudo apt-get update && sudo apt-get install antiword
   ```
 
+### 🤖 AI-Powered Features - Important Warnings
+
+**⚠️ EXPERIMENTAL TECHNOLOGY**: AI-powered features (OCR-Ollama, audio transcription) are based on machine learning models and may produce:
+
+#### Potential Issues:
+- **Hallucinations**: AI may "see" or "hear" content that isn't actually present
+- **Misinterpretations**: Complex images may be incorrectly identified (e.g., X-ray images mistaken for historical artifacts)
+- **Language Errors**: Transcription accuracy depends on audio quality, accents, and background noise
+- **Context Confusion**: AI may apply inappropriate cultural/historical context to technical content
+- **Model Dependence**: Results vary significantly between different AI models (gemma3, llava, whisper versions)
+- **Bias and Inconsistency**: Models may exhibit cultural, linguistic, or domain-specific biases
+
+#### Critical Applications Warning:
+**🚨 DO NOT USE for critical applications** such as:
+- Medical diagnosis or medical image interpretation
+- Legal document analysis requiring perfect accuracy  
+- Financial data extraction where errors have monetary impact
+- Security/safety systems where false positives/negatives are dangerous
+- Academic research requiring citation-quality accuracy
+
+#### Best Practices:
+- **Always validate AI results** against source material when accuracy matters
+- **Use confidence scoring** to identify potentially unreliable results
+- **Cross-reference** with traditional OCR/transcription tools for important content
+- **Human review** recommended for any production use case
+- **Test thoroughly** with your specific content types and use cases
+- **Fallback options**: Keep traditional OCR (EasyOCR) available as backup
+
+#### Recommended Use Cases:
+✅ Content discovery and initial text extraction  
+✅ Batch processing of low-stakes content  
+✅ Development and prototyping workflows  
+✅ Personal document organization  
+✅ Educational and learning projects
+
 ## 📖 Full Examples
 
 ### Accessing Examples After Installation
@@ -427,7 +509,13 @@ Pull requests, issues, and feedback are warmly welcome! 🚀
 - ✅ **NEW**: Caption length control (short/medium/long)
 - ✅ **NEW**: Temperature and token limit configuration
 - ✅ **NEW**: Command-line OCR example script with full parameter support
+- ✅ **NEW**: **Confidence Scoring System** - Advanced AI hallucination detection
+- ✅ **NEW**: `xtxt_image_with_confidence()` - Returns text + confidence score
+- ✅ **NEW**: Automatic rejection of low-confidence results with configurable thresholds
+- ✅ **NEW**: 60+ hallucination patterns detected (historical, artistic, fantasy, scientific)
+- ✅ **NEW**: `--show-confidence` and `--confidence` CLI parameters
 - ✅ **ENHANCED**: OCR-Ollama mode with both text extraction and image description
+- ✅ **ENHANCED**: Comprehensive AI safety warnings and best practices documentation
 - ✅ Support for gemma3:4b, gemma3:12b, gemma3:27b, llava:7b, llava:13b models
 
 ### v0.2.4 

{pyxtxt-0.3.2 → pyxtxt-0.3.4}/README.md

@@ -204,6 +204,45 @@ image_response = requests.get("https://example.com/document.png")
 text = xtxt(image_response.content)
 ```
 
+### AI OCR Confidence Scoring (NEW)
+
+⚠️ **IMPORTANT**: AI-powered OCR is experimental technology that may produce errors, hallucinations, or misinterpretations. Always validate results for critical applications.
+
+The OCR-Ollama system includes confidence scoring to help identify unreliable results:
+
+```python
+from pyxtxt import xtxt_image_with_confidence, set_ollama_config
+
+# Configure confidence threshold (0.0-1.0, default: 0.7)
+set_ollama_config(confidence_threshold=0.8)  # More restrictive
+
+# Get text with confidence score
+text, confidence = xtxt_image_with_confidence("document.png", mode="ocr")
+print(f"Confidence: {confidence:.2f} ({confidence*100:.1f}%)")
+print(f"Text: {text}")
+
+# Check if reliable
+if confidence < 0.7:
+    print("⚠️ Low confidence - result may be unreliable")
+    print("Consider using traditional OCR or manual verification")
+else:
+    print("✅ Good confidence - result likely reliable")
+```
+
+#### Confidence Scoring Features
+
+- **Hallucination Detection**: Penalizes historical/cultural references that often indicate AI misinterpretation
+- **Pattern Recognition**: Rewards structured text (numbers, punctuation, proper formatting)
+- **Uncertainty Detection**: Flags vague language ("appears to be", "seems like", etc.)
+- **Quality Assessment**: Considers content length, repetition, and coherence
+
+#### Common Hallucination Patterns (Automatically Detected)
+- **Historical Content**: "ancient", "medieval", "Egyptian papyrus", "hieroglyphs"
+- **Artistic Interpretations**: "painting", "artwork", "masterpiece", "Renaissance"  
+- **Fantasy Content**: "mystical", "magical", "dragon", "wizard"
+- **Scientific Misinterpretation**: "fossil", "geological formation", "crystal structure"
+- **Vague Language**: "unclear", "difficult to read", "appears to be"
+
 ### Command-Line OCR Example
 
 A complete example script for command-line usage is available:
@@ -220,6 +259,11 @@ with open("ocr_example.py", "wb") as f:
 # python ocr_example.py document.png
 # python ocr_example.py chart.jpg --mode=describe --lang=italian --style=detailed
 # python ocr_example.py diagram.png --mode=describe --length=long --temp=0.3
+
+# NEW: Confidence scoring examples
+# python ocr_example.py suspicious.png --show-confidence
+# python ocr_example.py medical.png --confidence=0.9 --show-confidence
+# python ocr_example.py diagram.png --confidence=0.5 --mode=describe --show-confidence
 ```
 
 The script supports:
@@ -229,6 +273,8 @@ The script supports:
 - **Style control**: descriptive, technical, simple, detailed
 - **Length control**: short, medium, long captions
 - **Temperature**: Adjust LLM creativity (0.0-1.0)
+- **Confidence scoring**: Set threshold and display confidence scores
+- **Quality filtering**: Automatically reject low-confidence results
 
 ### Show Available Formats
 ```python
@@ -272,6 +318,7 @@ text = xtxt(attachment_bytes)
 
 ## ⚠️ Known Limitations
 
+### General Limitations
 - **Legacy file detection**: When using raw streams without filenames, legacy files (.doc, .xls, .ppt) may not be correctly detected due to identical file signatures in libmagic
 - **Filename hints recommended**: When available, providing original filenames improves detection accuracy
 - **MSWrite .doc files**: Require `antiword` installation:
@@ -279,6 +326,41 @@ text = xtxt(attachment_bytes)
   sudo apt-get update && sudo apt-get install antiword
   ```
 
+### 🤖 AI-Powered Features - Important Warnings
+
+**⚠️ EXPERIMENTAL TECHNOLOGY**: AI-powered features (OCR-Ollama, audio transcription) are based on machine learning models and may produce:
+
+#### Potential Issues:
+- **Hallucinations**: AI may "see" or "hear" content that isn't actually present
+- **Misinterpretations**: Complex images may be incorrectly identified (e.g., X-ray images mistaken for historical artifacts)
+- **Language Errors**: Transcription accuracy depends on audio quality, accents, and background noise
+- **Context Confusion**: AI may apply inappropriate cultural/historical context to technical content
+- **Model Dependence**: Results vary significantly between different AI models (gemma3, llava, whisper versions)
+- **Bias and Inconsistency**: Models may exhibit cultural, linguistic, or domain-specific biases
+
+#### Critical Applications Warning:
+**🚨 DO NOT USE for critical applications** such as:
+- Medical diagnosis or medical image interpretation
+- Legal document analysis requiring perfect accuracy  
+- Financial data extraction where errors have monetary impact
+- Security/safety systems where false positives/negatives are dangerous
+- Academic research requiring citation-quality accuracy
+
+#### Best Practices:
+- **Always validate AI results** against source material when accuracy matters
+- **Use confidence scoring** to identify potentially unreliable results
+- **Cross-reference** with traditional OCR/transcription tools for important content
+- **Human review** recommended for any production use case
+- **Test thoroughly** with your specific content types and use cases
+- **Fallback options**: Keep traditional OCR (EasyOCR) available as backup
+
+#### Recommended Use Cases:
+✅ Content discovery and initial text extraction  
+✅ Batch processing of low-stakes content  
+✅ Development and prototyping workflows  
+✅ Personal document organization  
+✅ Educational and learning projects
+
 ## 📖 Full Examples
 
 ### Accessing Examples After Installation
@@ -323,7 +405,13 @@ Pull requests, issues, and feedback are warmly welcome! 🚀
 - ✅ **NEW**: Caption length control (short/medium/long)
 - ✅ **NEW**: Temperature and token limit configuration
 - ✅ **NEW**: Command-line OCR example script with full parameter support
+- ✅ **NEW**: **Confidence Scoring System** - Advanced AI hallucination detection
+- ✅ **NEW**: `xtxt_image_with_confidence()` - Returns text + confidence score
+- ✅ **NEW**: Automatic rejection of low-confidence results with configurable thresholds
+- ✅ **NEW**: 60+ hallucination patterns detected (historical, artistic, fantasy, scientific)
+- ✅ **NEW**: `--show-confidence` and `--confidence` CLI parameters
 - ✅ **ENHANCED**: OCR-Ollama mode with both text extraction and image description
+- ✅ **ENHANCED**: Comprehensive AI safety warnings and best practices documentation
 - ✅ Support for gemma3:4b, gemma3:12b, gemma3:27b, llava:7b, llava:13b models
 
 ### v0.2.4 

{pyxtxt-0.3.2 → pyxtxt-0.3.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "pyxtxt"
-version = "0.3.2"
+version = "0.3.4"
 description = "A Python library for extracting text from different types of files (PDF, DOCX, PPTX, XLSX, ODT, ecc.)."
 classifiers = [
   "Development Status :: 4 - Beta",

{pyxtxt-0.3.2 → pyxtxt-0.3.4}/src/pyxtxt/__init__.py

@@ -4,12 +4,14 @@ from .core import xtxt, extxt_available_formats, xtxt_from_url
 try:
     from .estrattori.ocr_ollama import (
         set_ollama_model, get_ollama_model, xtxt_image_describe,
-        set_ollama_config, get_ollama_config, reset_ollama_config
+        set_ollama_config, get_ollama_config, reset_ollama_config,
+        xtxt_image_with_confidence
     )
     __all__ = [
         "xtxt", "extxt_available_formats", "xtxt_from_url", 
         "set_ollama_model", "get_ollama_model", "xtxt_image_describe",
-        "set_ollama_config", "get_ollama_config", "reset_ollama_config"
+        "set_ollama_config", "get_ollama_config", "reset_ollama_config",
+        "xtxt_image_with_confidence"
     ]
 except ImportError:
     __all__ = ["xtxt", "extxt_available_formats", "xtxt_from_url"]

pyxtxt-0.3.4/src/pyxtxt/estrattori/ocr_ollama.py

@@ -0,0 +1,520 @@
+# pyxtxt/extractors/image_ocr_ollama.py
+from . import register_extractor
+from io import BytesIO
+import base64
+
+try:
+    import ollama
+    from PIL import Image
+except ImportError:
+    ollama = None
+    Image = None
+
+# Global configuration for Ollama model and parameters
+OLLAMA_MODEL = "gemma3:4b"  # Default multimodal model
+OLLAMA_CONFIG = {
+    'language': 'auto',  # Language hint for caption generation
+    'caption_length': 'medium',  # short, medium, long
+    'style': 'descriptive',  # descriptive, technical, simple, detailed
+    'temperature': 0.1,  # Response creativity (0.0-1.0)
+    'max_tokens': 1500,  # Maximum response length
+    'confidence_threshold': 0.7,  # Minimum confidence for text extraction
+    'context': 'general'  # Context hint: general, cookbook, document, diagram, etc.
+}
+
+def set_ollama_model(model_name: str):
+    """
+    Set the Ollama model to use for OCR.
+    
+    Recommended multimodal models:
+    - gemma3:4b (default, balanced speed/quality)
+    - gemma3:12b (higher quality, slower)
+    - gemma3:27b (best quality, very slow)
+    - llava:7b (alternative vision model)
+    - llava:13b (higher quality LLAVA)
+    """
+    global OLLAMA_MODEL
+    OLLAMA_MODEL = model_name
+    print(f"✅ Ollama OCR model set to: {model_name}")
+
+def get_ollama_model():
+    """Get current Ollama model name"""
+    return OLLAMA_MODEL
+
+def set_ollama_config(**kwargs):
+    """
+    Configure Ollama LLM parameters for better caption generation.
+    
+    Parameters:
+    - language: Language hint ('auto', 'italian', 'english', 'spanish', etc.)
+    - caption_length: Caption length ('short', 'medium', 'long')
+    - style: Caption style ('descriptive', 'technical', 'simple', 'detailed')
+    - context: Content context ('general', 'cookbook', 'document', 'handwriting', 'technical')
+    - temperature: Response creativity 0.0-1.0 (default: 0.1)
+    - max_tokens: Maximum response length (default: 1500)
+    - confidence_threshold: Text extraction confidence 0.0-1.0 (default: 0.7)
+    
+    Examples:
+        set_ollama_config(language='italian', style='detailed')
+        set_ollama_config(context='document', caption_length='long')
+        set_ollama_config(context='handwriting', temperature=0.2)
+    """
+    global OLLAMA_CONFIG
+    for key, value in kwargs.items():
+        if key in OLLAMA_CONFIG:
+            OLLAMA_CONFIG[key] = value
+            print(f"✅ Ollama config updated: {key} = {value}")
+        else:
+            print(f"⚠️ Unknown config parameter: {key}")
+
+def get_ollama_config():
+    """Get current Ollama configuration"""
+    return OLLAMA_CONFIG.copy()
+
+def reset_ollama_config():
+    """Reset Ollama configuration to defaults"""
+    global OLLAMA_CONFIG
+    OLLAMA_CONFIG = {
+        'language': 'auto',
+        'caption_length': 'medium', 
+        'style': 'descriptive',
+        'temperature': 0.1,
+        'max_tokens': 1500,
+        'confidence_threshold': 0.7,
+        'context': 'general'
+    }
+    print("✅ Ollama configuration reset to defaults")
+
+def _calculate_confidence_score(content: str, mode: str) -> float:
+    """
+    Calculate confidence score (0.0-1.0) for OCR/caption quality.
+    
+    Evaluates response based on:
+    - Content length and structure
+    - Presence of meaningful text patterns
+    - Absence of hallucination indicators
+    - Appropriate response format
+    """
+    if not content or len(content.strip()) < 3:
+        return 0.0
+    
+    content_lower = content.lower()
+    score = 0.5  # Base score
+    
+    # Positive indicators (add to confidence)
+    positive_patterns = [
+        # Structured text indicators
+        (r'\b\d+\b', 0.1, "Contains numbers/measurements"),
+        (r'[.!?]', 0.1, "Has sentence punctuation"), 
+        (r'\b(the|and|of|in|to|for|with|on|at|by)\b', 0.1, "Common words present"),
+        (r'\b[A-Z][a-z]+\b', 0.1, "Proper capitalization"),
+        (r'[,;:]', 0.05, "Has proper punctuation"),
+        
+        # Content-specific indicators
+        (r'\b(text|document|page|title|heading)\b', 0.1, "Document references"),
+        (r'\b(image|photo|picture|shows|contains)\b', 0.1, "Visual descriptions"),
+        (r'\b\d+[%$€£¥]\b|\b[%$€£¥]\d+\b', 0.1, "Currency/percentages"),
+        (r'\b\d{1,2}[:/]\d{1,2}\b|\b\d{4}\b', 0.05, "Times/years"),
+    ]
+    
+    # Negative indicators (reduce confidence) 
+    negative_patterns = [
+        # Historical/Archaeological hallucinations (common with X-rays, scans)
+        ("ancient", 0.2, "May be hallucinating historical content"),
+        ("egypt", 0.15, "Egyptian hallucination"),
+        ("hieroglyph", 0.25, "Hieroglyphic hallucination"),
+        ("papyrus", 0.2, "Papyrus hallucination"),
+        ("scroll", 0.1, "Ancient scroll hallucination"),
+        ("medieval", 0.15, "Medieval hallucination"),
+        ("manuscript", 0.1, "Manuscript hallucination"),
+        ("archaeological", 0.15, "Archaeological hallucination"),
+        ("artifact", 0.12, "Artifact hallucination"),
+        ("roman", 0.1, "Roman historical hallucination"),
+        ("greek", 0.1, "Greek historical hallucination"),
+        ("biblical", 0.15, "Religious historical hallucination"),
+        ("stone tablet", 0.2, "Stone tablet hallucination"),
+        ("carved", 0.1, "Carving hallucination"),
+        
+        # Art/Cultural hallucinations (common with medical images, diagrams)
+        ("painting", 0.1, "Artistic interpretation hallucination"),
+        ("artwork", 0.12, "Artwork hallucination"),  
+        ("masterpiece", 0.15, "Art masterpiece hallucination"),
+        ("renaissance", 0.15, "Renaissance art hallucination"),
+        ("portrait", 0.08, "Portrait hallucination"),
+        ("landscape", 0.08, "Landscape hallucination"),
+        ("abstract art", 0.12, "Abstract art hallucination"),
+        
+        # Fantasy/Fictional content (can occur with any unclear image)
+        ("mystical", 0.15, "Fantasy content hallucination"),
+        ("magical", 0.15, "Magical content hallucination"),
+        ("mythological", 0.15, "Mythology hallucination"),
+        ("fairy tale", 0.15, "Fairy tale hallucination"),
+        ("legend", 0.1, "Legend/folklore hallucination"),
+        ("dragon", 0.2, "Dragon hallucination"),
+        ("wizard", 0.15, "Fantasy character hallucination"),
+        
+        # Scientific misinterpretations (X-rays as geological, etc.)
+        ("fossil", 0.15, "Fossil hallucination"),
+        ("geological", 0.1, "Geological misinterpretation"),
+        ("mineral", 0.1, "Mineral hallucination"),
+        ("crystal", 0.1, "Crystal hallucination"),
+        ("rock formation", 0.12, "Rock formation hallucination"),
+        ("sediment", 0.1, "Sediment hallucination"),
+        
+        # Vague/uncertain language
+        ("unclear", 0.1, "Uncertainty indicator"),
+        ("difficult to read", 0.1, "Reading difficulty"),
+        ("appears to be", 0.05, "Tentative language"),
+        ("seems to", 0.05, "Uncertain language"),
+        ("might be", 0.1, "Possibility language"),
+        ("possibly", 0.1, "Uncertain language"),
+        ("probably", 0.08, "Probability language"),
+        ("looks like", 0.08, "Appearance-based guess"),
+        ("reminds me of", 0.1, "Subjective association"),
+        ("similar to", 0.05, "Similarity guess"),
+        
+        # Quality/visibility issues (legitimate but indicate uncertainty)
+        ("blurry", 0.05, "Image quality issue"),
+        ("faded", 0.05, "Faded content"),
+        ("damaged", 0.05, "Damaged content"),
+        ("corrupted", 0.1, "Corrupted content"),
+        ("low resolution", 0.08, "Resolution issue"),
+        ("hard to make out", 0.1, "Visibility issue"),
+        
+        # Generic/evasive responses
+        ("sorry, i cannot", 0.3, "Refusal response"),
+        ("cannot determine", 0.2, "Unable to process"),
+        ("unable to identify", 0.15, "Identification failure"),
+        ("not clear enough", 0.1, "Clarity issue"),
+        ("too dark", 0.08, "Darkness issue"),
+        ("too bright", 0.08, "Brightness issue"),
+        ("no visible text", 0.0, "No text found - appropriate response"),
+        
+        # Repetitive/nonsensical content (AI breakdown indicators)
+        ("lorem ipsum", 0.2, "Placeholder text hallucination"),
+        ("test test test", 0.25, "Repetitive test pattern"),
+        ("abc abc abc", 0.2, "Repetitive pattern"),
+        ("error error", 0.15, "Error message hallucination"),
+    ]
+    
+    import re
+    
+    # Apply positive pattern scoring
+    for pattern, bonus, description in positive_patterns:
+        if re.search(pattern, content_lower):
+            score += bonus
+    
+    # Apply negative pattern scoring  
+    for pattern, penalty, description in negative_patterns:
+        if pattern in content_lower:
+            score -= penalty
+            print(f"⚠️ Confidence penalty ({penalty}): {description}")
+    
+    # Length-based adjustments
+    if len(content) > 200:  # Substantial content
+        score += 0.1
+    elif len(content) < 20:  # Very short responses are suspicious
+        score -= 0.2
+    
+    # Mode-specific adjustments
+    if mode == "describe":
+        # Check for proper formatting in describe mode
+        if "TEXT:" in content and "DESCRIPTION:" in content:
+            score += 0.1  # Proper format bonus
+        elif len(content) > 50:  # Has substantial descriptive content
+            score += 0.05
+    
+    # Context coherence check - very basic
+    words = content_lower.split()
+    if len(words) > 5:
+        # Check for repetitive patterns (hallucination indicator)
+        unique_words = set(words)
+        repetition_ratio = len(unique_words) / len(words)
+        if repetition_ratio < 0.3:  # Too repetitive
+            score -= 0.2
+    
+    # Clamp to valid range
+    return max(0.0, min(1.0, score))
+
+if ollama and Image:
+    def xtxt_image_ocr_ollama(file_buffer, mode="ocr", model=None):
+        """
+        Extract text from images using Ollama with multimodal models.
+        
+        Args:
+            file_buffer: Image file buffer
+            mode: "ocr" (text only) or "describe" (text + description)  
+            model: Override default model (optional)
+        """
+        try:
+            # Use specified model or global default
+            current_model = model or OLLAMA_MODEL
+            
+            # Convert buffer to PIL Image
+            # Reset buffer position if it has read method
+            if hasattr(file_buffer, 'seek'):
+                file_buffer.seek(0)
+            image_data = file_buffer.read()
+            image = Image.open(BytesIO(image_data))
+            
+            # Convert to RGB if needed
+            if image.mode != 'RGB':
+                image = image.convert('RGB')
+            
+            # Convert image to base64
+            buffered = BytesIO()
+            image.save(buffered, format="PNG")
+            img_base64 = base64.b64encode(buffered.getvalue()).decode()
+            
+            # Get current configuration
+            config = OLLAMA_CONFIG
+            
+            # Build language hint
+            lang_hint = ""
+            if config['language'] != 'auto':
+                lang_hint = f"Text language: {config['language']}. "
+            
+            # Different prompts based on mode
+            if mode == "ocr":
+                prompt = f"""Look carefully at this image and extract ALL text that you can see, including:
+- Titles, headings, and main text content  
+- Small print, captions, labels, and annotations
+- Numbers, measurements, quantities, and symbols
+- Menu items, ingredient lists, cooking instructions
+- Any text in boxes, speech bubbles, or decorative elements
+
+IMPORTANT: 
+- Read carefully and include even small or partially visible text
+- Preserve the original formatting and line breaks where possible
+- {lang_hint}Process the text from left to right, top to bottom
+- If you cannot find any readable text at all, respond with 'NO_TEXT_FOUND'
+
+Extracted text:"""
+            
+            else:  # mode == "describe"
+                # Build style-specific prompts
+                style_prompts = {
+                    'descriptive': "Provide a clear, descriptive explanation",
+                    'technical': "Use technical terminology and precise descriptions", 
+                    'simple': "Use simple, easy-to-understand language",
+                    'detailed': "Provide comprehensive details about all visual elements"
+                }
+                
+                length_hints = {
+                    'short': "Keep descriptions brief (1-2 sentences)",
+                    'medium': "Provide moderate detail (2-4 sentences)",
+                    'long': "Give comprehensive descriptions (4-8 sentences)"
+                }
+                
+                style_instruction = style_prompts.get(config['style'], style_prompts['descriptive'])
+                length_instruction = length_hints.get(config['caption_length'], length_hints['medium'])
+                
+                # Context-specific hints (only when explicitly set)
+                context_hint = ""
+                context = config.get('context', 'general').lower()
+                if context == 'cookbook' or context == 'recipe':
+                    context_hint = """
+   - If this appears to be a recipe/cookbook page, include: ingredients, cooking steps, quantities, cooking times
+   - Mention any photos of prepared dishes or cooking techniques shown
+   - Note any special formatting like ingredient lists, step numbers, or cooking tips"""
+                elif context == 'document':
+                    context_hint = """
+   - Focus on document structure: headers, paragraphs, sections, page numbers
+   - Note any official formatting, letterheads, signatures, or stamps"""
+                elif context == 'handwriting' or context == 'notes':
+                    context_hint = """
+   - Pay special attention to handwritten text which may be harder to read
+   - Note any sketches, diagrams, or informal formatting typical of personal notes"""
+                elif context == 'technical' or context == 'diagram':
+                    context_hint = """
+   - Focus on technical elements: labels, measurements, specifications, diagrams
+   - Include any mathematical formulas, technical symbols, or engineering notations"""
+
+                prompt = f"""Analyze this image and provide:
+1. All visible text exactly as written (preserve formatting, line breaks, bullet points)
+2. Image description following these guidelines:
+   - {style_instruction}
+   - {length_instruction}
+   - {lang_hint}Focus on key visual elements, layout, and context{context_hint}
+
+Format:
+TEXT: [all visible text here, or NO_TEXT_FOUND if none]
+DESCRIPTION: [image description following the guidelines above]"""
+            
+            # Send request to Ollama with configured parameters
+            response = ollama.generate(
+                model=current_model,
+                prompt=prompt,
+                images=[img_base64],
+                options={
+                    'temperature': config['temperature'],
+                    'top_p': 0.9,
+                    'num_predict': config['max_tokens']
+                }
+            )
+            
+            # Extract and clean response
+            extracted_content = response.get('response', '').strip()
+            
+            # Calculate confidence score based on response quality indicators
+            confidence_score = _calculate_confidence_score(extracted_content, mode)
+            
+            # Check confidence threshold
+            if confidence_score < config['confidence_threshold']:
+                print(f"⚠️ Low confidence ({confidence_score:.2f} < {config['confidence_threshold']}): {extracted_content[:100]}...")
+                if mode == "ocr":
+                    return ""  # Return empty for OCR if below threshold
+                else:
+                    # For describe mode, add warning prefix
+                    extracted_content = f"[LOW_CONFIDENCE_{confidence_score:.2f}] {extracted_content}"
+            else:
+                print(f"✅ Good confidence ({confidence_score:.2f}): Processing successful")
+            
+            # Handle no-text case for OCR mode
+            if mode == "ocr" and ('NO_TEXT_FOUND' in extracted_content or len(extracted_content) < 3):
+                return ""
+            
+            return extracted_content
+            
+        except Exception as e:
+            print(f"⚠️ Error extracting from image with Ollama {current_model}: {e}")
+            return ""
+    
+    def xtxt_image_ocr_ollama_with_confidence(file_buffer, mode="ocr", model=None):
+        """
+        Version of OCR-Ollama that returns both text and confidence score.
+        
+        Returns:
+            tuple: (extracted_text, confidence_score)
+        """
+        try:
+            # Use specified model or global default
+            current_model = model or OLLAMA_MODEL
+            
+            # Convert buffer to PIL Image
+            if hasattr(file_buffer, 'seek'):
+                file_buffer.seek(0)
+            image_data = file_buffer.read()
+            image = Image.open(BytesIO(image_data))
+            
+            # Convert to RGB if needed
+            if image.mode != 'RGB':
+                image = image.convert('RGB')
+            
+            # Convert image to base64
+            buffered = BytesIO()
+            image.save(buffered, format="PNG")
+            img_base64 = base64.b64encode(buffered.getvalue()).decode()
+            
+            # Get current configuration
+            config = OLLAMA_CONFIG
+            
+            # Build prompts (same logic as main function)
+            lang_hint = ""
+            if config['language'] != 'auto':
+                lang_hint = f"Text language: {config['language']}. "
+            
+            if mode == "ocr":
+                prompt = f"""Look carefully at this image and extract ALL text that you can see, including:
+- Titles, headings, and main text content  
+- Small print, captions, labels, and annotations
+- Numbers, measurements, quantities, and symbols
+- Menu items, ingredient lists, cooking instructions
+- Any text in boxes, speech bubbles, or decorative elements
+
+IMPORTANT: 
+- Read carefully and include even small or partially visible text
+- Preserve the original formatting and line breaks where possible
+- {lang_hint}Process the text from left to right, top to bottom
+- If you cannot find any readable text at all, respond with 'NO_TEXT_FOUND'
+
+Extracted text:"""
+            else:  # describe mode
+                style_prompts = {
+                    'descriptive': "Provide a clear, descriptive explanation",
+                    'technical': "Use technical terminology and precise descriptions", 
+                    'simple': "Use simple, easy-to-understand language",
+                    'detailed': "Provide comprehensive details about all visual elements"
+                }
+                
+                length_hints = {
+                    'short': "Keep descriptions brief (1-2 sentences)",
+                    'medium': "Provide moderate detail (2-4 sentences)",
+                    'long': "Give comprehensive descriptions (4-8 sentences)"
+                }
+                
+                style_instruction = style_prompts.get(config['style'], style_prompts['descriptive'])
+                length_instruction = length_hints.get(config['caption_length'], length_hints['medium'])
+                
+                prompt = f"""Analyze this image and provide:
+1. All visible text exactly as written (preserve formatting, line breaks, bullet points)
+2. Image description following these guidelines:
+   - {style_instruction}
+   - {length_instruction}
+   - {lang_hint}Focus on key visual elements, layout, and context
+
+Format:
+TEXT: [all visible text here, or NO_TEXT_FOUND if none]
+DESCRIPTION: [image description following the guidelines above]"""
+            
+            # Send request to Ollama
+            response = ollama.generate(
+                model=current_model,
+                prompt=prompt,
+                images=[img_base64],
+                options={
+                    'temperature': config['temperature'],
+                    'top_p': 0.9,
+                    'num_predict': config['max_tokens']
+                }
+            )
+            
+            # Extract response
+            extracted_content = response.get('response', '').strip()
+            
+            # Calculate confidence without threshold filtering
+            confidence_score = _calculate_confidence_score(extracted_content, mode)
+            
+            # Return both text and confidence
+            return extracted_content, confidence_score
+            
+        except Exception as e:
+            print(f"⚠️ Error extracting from image with Ollama {current_model}: {e}")
+            return "", 0.0
+
+    # Wrapper functions for each mode
+    def xtxt_image_ocr_only(file_input):
+        """Traditional OCR: extract only visible text using Ollama"""
+        # Handle both file paths and buffers
+        if isinstance(file_input, str):
+            with open(file_input, 'rb') as f:
+                return xtxt_image_ocr_ollama(f, mode="ocr")
+        else:
+            return xtxt_image_ocr_ollama(file_input, mode="ocr")
+    
+    def xtxt_image_describe(file_input):
+        """OCR + Description: text + image context using Ollama"""
+        # Handle both file paths and buffers  
+        if isinstance(file_input, str):
+            with open(file_input, 'rb') as f:
+                return xtxt_image_ocr_ollama(f, mode="describe")
+        else:
+            return xtxt_image_ocr_ollama(file_input, mode="describe")
+
+    def xtxt_image_with_confidence(file_input, mode="ocr"):
+        """Get both text and confidence score from image OCR"""
+        if isinstance(file_input, str):
+            with open(file_input, 'rb') as f:
+                return xtxt_image_ocr_ollama_with_confidence(f, mode=mode)
+        else:
+            return xtxt_image_ocr_ollama_with_confidence(file_input, mode=mode)
+    
+    # Register OCR-only version as default
+    # Note: Will override traditional EasyOCR if both modules are loaded
+    image_formats = [
+        "image/jpeg", "image/jpg", "image/png", 
+        "image/bmp", "image/tiff", "image/webp"
+    ]
+    
+    for format_type in image_formats:
+        register_extractor(format_type, xtxt_image_ocr_only, name="OCR-Ollama")

{pyxtxt-0.3.2 → pyxtxt-0.3.4/src/pyxtxt.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyxtxt
-Version: 0.3.2
+Version: 0.3.4
 Summary: A Python library for extracting text from different types of files (PDF, DOCX, PPTX, XLSX, ODT, ecc.).
 Author-email: Giuseppe Levi <giuseppe.levi@gmail.com>
 License: MIT License
@@ -308,6 +308,45 @@ image_response = requests.get("https://example.com/document.png")
 text = xtxt(image_response.content)
 ```
 
+### AI OCR Confidence Scoring (NEW)
+
+⚠️ **IMPORTANT**: AI-powered OCR is experimental technology that may produce errors, hallucinations, or misinterpretations. Always validate results for critical applications.
+
+The OCR-Ollama system includes confidence scoring to help identify unreliable results:
+
+```python
+from pyxtxt import xtxt_image_with_confidence, set_ollama_config
+
+# Configure confidence threshold (0.0-1.0, default: 0.7)
+set_ollama_config(confidence_threshold=0.8)  # More restrictive
+
+# Get text with confidence score
+text, confidence = xtxt_image_with_confidence("document.png", mode="ocr")
+print(f"Confidence: {confidence:.2f} ({confidence*100:.1f}%)")
+print(f"Text: {text}")
+
+# Check if reliable
+if confidence < 0.7:
+    print("⚠️ Low confidence - result may be unreliable")
+    print("Consider using traditional OCR or manual verification")
+else:
+    print("✅ Good confidence - result likely reliable")
+```
+
+#### Confidence Scoring Features
+
+- **Hallucination Detection**: Penalizes historical/cultural references that often indicate AI misinterpretation
+- **Pattern Recognition**: Rewards structured text (numbers, punctuation, proper formatting)
+- **Uncertainty Detection**: Flags vague language ("appears to be", "seems like", etc.)
+- **Quality Assessment**: Considers content length, repetition, and coherence
+
+#### Common Hallucination Patterns (Automatically Detected)
+- **Historical Content**: "ancient", "medieval", "Egyptian papyrus", "hieroglyphs"
+- **Artistic Interpretations**: "painting", "artwork", "masterpiece", "Renaissance"  
+- **Fantasy Content**: "mystical", "magical", "dragon", "wizard"
+- **Scientific Misinterpretation**: "fossil", "geological formation", "crystal structure"
+- **Vague Language**: "unclear", "difficult to read", "appears to be"
+
 ### Command-Line OCR Example
 
 A complete example script for command-line usage is available:
@@ -324,6 +363,11 @@ with open("ocr_example.py", "wb") as f:
 # python ocr_example.py document.png
 # python ocr_example.py chart.jpg --mode=describe --lang=italian --style=detailed
 # python ocr_example.py diagram.png --mode=describe --length=long --temp=0.3
+
+# NEW: Confidence scoring examples
+# python ocr_example.py suspicious.png --show-confidence
+# python ocr_example.py medical.png --confidence=0.9 --show-confidence
+# python ocr_example.py diagram.png --confidence=0.5 --mode=describe --show-confidence
 ```
 
 The script supports:
@@ -333,6 +377,8 @@ The script supports:
 - **Style control**: descriptive, technical, simple, detailed
 - **Length control**: short, medium, long captions
 - **Temperature**: Adjust LLM creativity (0.0-1.0)
+- **Confidence scoring**: Set threshold and display confidence scores
+- **Quality filtering**: Automatically reject low-confidence results
 
 ### Show Available Formats
 ```python
@@ -376,6 +422,7 @@ text = xtxt(attachment_bytes)
 
 ## ⚠️ Known Limitations
 
+### General Limitations
 - **Legacy file detection**: When using raw streams without filenames, legacy files (.doc, .xls, .ppt) may not be correctly detected due to identical file signatures in libmagic
 - **Filename hints recommended**: When available, providing original filenames improves detection accuracy
 - **MSWrite .doc files**: Require `antiword` installation:
@@ -383,6 +430,41 @@ text = xtxt(attachment_bytes)
   sudo apt-get update && sudo apt-get install antiword
   ```
 
+### 🤖 AI-Powered Features - Important Warnings
+
+**⚠️ EXPERIMENTAL TECHNOLOGY**: AI-powered features (OCR-Ollama, audio transcription) are based on machine learning models and may produce:
+
+#### Potential Issues:
+- **Hallucinations**: AI may "see" or "hear" content that isn't actually present
+- **Misinterpretations**: Complex images may be incorrectly identified (e.g., X-ray images mistaken for historical artifacts)
+- **Language Errors**: Transcription accuracy depends on audio quality, accents, and background noise
+- **Context Confusion**: AI may apply inappropriate cultural/historical context to technical content
+- **Model Dependence**: Results vary significantly between different AI models (gemma3, llava, whisper versions)
+- **Bias and Inconsistency**: Models may exhibit cultural, linguistic, or domain-specific biases
+
+#### Critical Applications Warning:
+**🚨 DO NOT USE for critical applications** such as:
+- Medical diagnosis or medical image interpretation
+- Legal document analysis requiring perfect accuracy  
+- Financial data extraction where errors have monetary impact
+- Security/safety systems where false positives/negatives are dangerous
+- Academic research requiring citation-quality accuracy
+
+#### Best Practices:
+- **Always validate AI results** against source material when accuracy matters
+- **Use confidence scoring** to identify potentially unreliable results
+- **Cross-reference** with traditional OCR/transcription tools for important content
+- **Human review** recommended for any production use case
+- **Test thoroughly** with your specific content types and use cases
+- **Fallback options**: Keep traditional OCR (EasyOCR) available as backup
+
+#### Recommended Use Cases:
+✅ Content discovery and initial text extraction  
+✅ Batch processing of low-stakes content  
+✅ Development and prototyping workflows  
+✅ Personal document organization  
+✅ Educational and learning projects
+
 ## 📖 Full Examples
 
 ### Accessing Examples After Installation
@@ -427,7 +509,13 @@ Pull requests, issues, and feedback are warmly welcome! 🚀
 - ✅ **NEW**: Caption length control (short/medium/long)
 - ✅ **NEW**: Temperature and token limit configuration
 - ✅ **NEW**: Command-line OCR example script with full parameter support
+- ✅ **NEW**: **Confidence Scoring System** - Advanced AI hallucination detection
+- ✅ **NEW**: `xtxt_image_with_confidence()` - Returns text + confidence score
+- ✅ **NEW**: Automatic rejection of low-confidence results with configurable thresholds
+- ✅ **NEW**: 60+ hallucination patterns detected (historical, artistic, fantasy, scientific)
+- ✅ **NEW**: `--show-confidence` and `--confidence` CLI parameters
 - ✅ **ENHANCED**: OCR-Ollama mode with both text extraction and image description
+- ✅ **ENHANCED**: Comprehensive AI safety warnings and best practices documentation
 - ✅ Support for gemma3:4b, gemma3:12b, gemma3:27b, llava:7b, llava:13b models
 
 ### v0.2.4 

pyxtxt-0.3.2/src/pyxtxt/estrattori/ocr_ollama.py

@@ -1,202 +0,0 @@
-# pyxtxt/extractors/image_ocr_ollama.py
-from . import register_extractor
-from io import BytesIO
-import base64
-
-try:
-    import ollama
-    from PIL import Image
-except ImportError:
-    ollama = None
-    Image = None
-
-# Global configuration for Ollama model and parameters
-OLLAMA_MODEL = "gemma3:4b"  # Default multimodal model
-OLLAMA_CONFIG = {
-    'language': 'auto',  # Language hint for caption generation
-    'caption_length': 'medium',  # short, medium, long
-    'style': 'descriptive',  # descriptive, technical, simple, detailed
-    'temperature': 0.1,  # Response creativity (0.0-1.0)
-    'max_tokens': 1500,  # Maximum response length
-    'confidence_threshold': 0.7  # Minimum confidence for text extraction
-}
-
-def set_ollama_model(model_name: str):
-    """
-    Set the Ollama model to use for OCR.
-    
-    Recommended multimodal models:
-    - gemma3:4b (default, balanced speed/quality)
-    - gemma3:12b (higher quality, slower)
-    - gemma3:27b (best quality, very slow)
-    - llava:7b (alternative vision model)
-    - llava:13b (higher quality LLAVA)
-    """
-    global OLLAMA_MODEL
-    OLLAMA_MODEL = model_name
-    print(f"✅ Ollama OCR model set to: {model_name}")
-
-def get_ollama_model():
-    """Get current Ollama model name"""
-    return OLLAMA_MODEL
-
-def set_ollama_config(**kwargs):
-    """
-    Configure Ollama LLM parameters for better caption generation.
-    
-    Parameters:
-    - language: Language hint ('auto', 'italian', 'english', 'spanish', etc.)
-    - caption_length: Caption length ('short', 'medium', 'long')
-    - style: Caption style ('descriptive', 'technical', 'simple', 'detailed')
-    - temperature: Response creativity 0.0-1.0 (default: 0.1)
-    - max_tokens: Maximum response length (default: 1500)
-    - confidence_threshold: Text extraction confidence 0.0-1.0 (default: 0.7)
-    
-    Examples:
-        set_ollama_config(language='italian', style='detailed')
-        set_ollama_config(caption_length='long', temperature=0.3)
-    """
-    global OLLAMA_CONFIG
-    for key, value in kwargs.items():
-        if key in OLLAMA_CONFIG:
-            OLLAMA_CONFIG[key] = value
-            print(f"✅ Ollama config updated: {key} = {value}")
-        else:
-            print(f"⚠️ Unknown config parameter: {key}")
-
-def get_ollama_config():
-    """Get current Ollama configuration"""
-    return OLLAMA_CONFIG.copy()
-
-def reset_ollama_config():
-    """Reset Ollama configuration to defaults"""
-    global OLLAMA_CONFIG
-    OLLAMA_CONFIG = {
-        'language': 'auto',
-        'caption_length': 'medium', 
-        'style': 'descriptive',
-        'temperature': 0.1,
-        'max_tokens': 1500,
-        'confidence_threshold': 0.7
-    }
-    print("✅ Ollama configuration reset to defaults")
-
-if ollama and Image:
-    def xtxt_image_ocr_ollama(file_buffer, mode="ocr", model=None):
-        """
-        Extract text from images using Ollama with multimodal models.
-        
-        Args:
-            file_buffer: Image file buffer
-            mode: "ocr" (text only) or "describe" (text + description)  
-            model: Override default model (optional)
-        """
-        try:
-            # Use specified model or global default
-            current_model = model or OLLAMA_MODEL
-            
-            # Convert buffer to PIL Image
-            image = Image.open(BytesIO(file_buffer.read()))
-            
-            # Convert to RGB if needed
-            if image.mode != 'RGB':
-                image = image.convert('RGB')
-            
-            # Convert image to base64
-            buffered = BytesIO()
-            image.save(buffered, format="PNG")
-            img_base64 = base64.b64encode(buffered.getvalue()).decode()
-            
-            # Get current configuration
-            config = OLLAMA_CONFIG
-            
-            # Build language hint
-            lang_hint = ""
-            if config['language'] != 'auto':
-                lang_hint = f"Text language: {config['language']}. "
-            
-            # Different prompts based on mode
-            if mode == "ocr":
-                prompt = f"""Extract ALL visible text from this image exactly as it appears. 
-Rules:
-- Only return text that is actually written/printed in the image
-- Preserve reading order (left to right, top to bottom)  
-- Maintain line breaks and formatting
-- Include numbers, symbols, special characters
-- Do NOT add descriptions, interpretations, or context
-- {lang_hint}If no text is visible, return 'NO_TEXT_FOUND'
-
-Extracted text:"""
-            
-            else:  # mode == "describe"
-                # Build style-specific prompts
-                style_prompts = {
-                    'descriptive': "Provide a clear, descriptive explanation",
-                    'technical': "Use technical terminology and precise descriptions", 
-                    'simple': "Use simple, easy-to-understand language",
-                    'detailed': "Provide comprehensive details about all visual elements"
-                }
-                
-                length_hints = {
-                    'short': "Keep descriptions brief (1-2 sentences)",
-                    'medium': "Provide moderate detail (2-4 sentences)",
-                    'long': "Give comprehensive descriptions (4-8 sentences)"
-                }
-                
-                style_instruction = style_prompts.get(config['style'], style_prompts['descriptive'])
-                length_instruction = length_hints.get(config['caption_length'], length_hints['medium'])
-                
-                prompt = f"""Analyze this image and provide:
-1. All visible text exactly as written
-2. Image description following these guidelines:
-   - {style_instruction}
-   - {length_instruction}
-   - {lang_hint}Focus on key visual elements, layout, and context
-
-Format:
-TEXT: [all visible text here, or NO_TEXT_FOUND if none]
-DESCRIPTION: [image description following the guidelines above]"""
-            
-            # Send request to Ollama with configured parameters
-            response = ollama.generate(
-                model=current_model,
-                prompt=prompt,
-                images=[img_base64],
-                options={
-                    'temperature': config['temperature'],
-                    'top_p': 0.9,
-                    'num_predict': config['max_tokens']
-                }
-            )
-            
-            # Extract and clean response
-            extracted_content = response.get('response', '').strip()
-            
-            # Handle no-text case for OCR mode
-            if mode == "ocr" and ('NO_TEXT_FOUND' in extracted_content or len(extracted_content) < 3):
-                return ""
-            
-            return extracted_content
-            
-        except Exception as e:
-            print(f"⚠️ Error extracting from image with Ollama {current_model}: {e}")
-            return ""
-    
-    # Wrapper functions for each mode
-    def xtxt_image_ocr_only(file_buffer):
-        """Traditional OCR: extract only visible text using Ollama"""
-        return xtxt_image_ocr_ollama(file_buffer, mode="ocr")
-    
-    def xtxt_image_describe(file_buffer):
-        """OCR + Description: text + image context using Ollama"""
-        return xtxt_image_ocr_ollama(file_buffer, mode="describe")
-    
-    # Register OCR-only version as default
-    # Note: Will override traditional EasyOCR if both modules are loaded
-    image_formats = [
-        "image/jpeg", "image/jpg", "image/png", 
-        "image/bmp", "image/tiff", "image/webp"
-    ]
-    
-    for format_type in image_formats:
-        register_extractor(format_type, xtxt_image_ocr_only, name="OCR-Ollama")