embed images

def/02-hybrid-opencv-ocr-llm.md

@@ -0,0 +1,111 @@
+# 02 - Hybrid OpenCV + OCR + LLM Approach
+
+## Date
+2025-10-28
+
+## Context
+Vision models (llava) were hallucinating text content badly - showing HTML code when there was none, inventing text that didn't exist. Pure OCR was fast and accurate but lost code formatting and structure.
+
+## Problem
+- **Vision models**: Hallucinate text content, can't be trusted for accurate extraction
+- **Pure OCR**: Accurate text but messy output, lost indentation/formatting
+- **Need**: Accurate text extraction + preserved code structure
+
+## Solution: Three-Stage Hybrid Approach
+
+### Stage 1: OpenCV Text Detection
+Use morphological operations to find text regions:
+- Adaptive thresholding (handles varying lighting)
+- Dilation with horizontal kernel to connect text lines
+- Contour detection to find bounding boxes
+- Filter by area and aspect ratio
+- Merge overlapping regions
+
+### Stage 2: Region-Based OCR
+- Sort regions by reading order (top-to-bottom, left-to-right)
+- Crop each region from original image
+- Run OCR on cropped regions (more accurate than full frame)
+- Tesseract with PSM 6 mode to preserve layout
+- Preserve indentation in cleaning step
+
+### Stage 3: Optional LLM Cleanup
+- Take accurate OCR output (no hallucination)
+- Use lightweight LLM (llama3.2:3b for speed) to:
+  - Fix obvious OCR errors (l→1, O→0)
+  - Restore code indentation and structure
+  - Preserve exact text content
+  - No added explanations or hallucinated content
+
+## Benefits
+✓ **Accurate**: OCR reads actual pixels, no hallucination
+✓ **Fast**: OpenCV detection is instant, focused OCR is quick
+✓ **Structured**: Regions separated with headers showing position
+✓ **Formatted**: Optional LLM cleanup preserves/restores code structure
+✓ **Deterministic**: Same input = same output (unlike vision models)
+
+## Implementation
+
+**New file:** `meetus/hybrid_processor.py`
+- `HybridProcessor` class with OpenCV detection + OCR + optional LLM
+- Region sorting for proper reading order
+- Visual separators between regions
+
+**CLI flags:**
+```bash
+--use-hybrid                 # Enable hybrid mode
+--hybrid-llm-cleanup        # Add LLM post-processing (optional)
+--hybrid-llm-model MODEL    # LLM model (default: llama3.2:3b)
+```
+
+**OCR improvements:**
+- Tesseract PSM 6 mode for better layout preservation
+- Modified text cleaning to keep indentation
+- `preserve_layout` parameter
+
+## Usage
+
+```bash
+# Basic hybrid (OpenCV + OCR)
+python process_meeting.py samples/video.mkv --use-hybrid --scene-detection
+
+# With LLM cleanup for best code formatting
+python process_meeting.py samples/video.mkv --use-hybrid --hybrid-llm-cleanup --scene-detection -v
+
+# Iterate on threshold
+python process_meeting.py samples/video.mkv --use-hybrid --scene-detection --scene-threshold 5 --skip-cache-frames --skip-cache-analysis
+```
+
+## Output Format
+
+```
+[Region 1 at y=120]
+function calculateTotal(items) {
+  return items.reduce((sum, item) => sum + item.price, 0);
+}
+
+============================================================
+
+[Region 2 at y=450]
+const result = calculateTotal(cartItems);
+console.log('Total:', result);
+```
+
+## Performance
+- **Without LLM cleanup**: Very fast (~2-3s per frame)
+- **With LLM cleanup**: Slower but still faster than vision models (~5-8s per frame)
+- **Accuracy**: Much better than vision model hallucinations
+
+## When to Use What
+
+| Method | Best For | Pros | Cons |
+|--------|----------|------|------|
+| **Hybrid** | Code/terminal text extraction | Accurate, fast, no hallucination | Formatting may be messy |
+| **Hybrid + LLM** | Code with preserved structure | Accurate + formatted | Slower, needs Ollama |
+| **Vision** | Understanding layout/context | Semantic understanding | Hallucinates text |
+| **Pure OCR** | Simple text, no structure needed | Fast, simple | Full-frame, no region detection |
+
+## Files Modified
+- `meetus/hybrid_processor.py` - New hybrid processor
+- `meetus/ocr_processor.py` - Layout preservation
+- `meetus/workflow.py` - Hybrid mode integration
+- `process_meeting.py` - CLI flags and examples

def/03-embed-images-for-llm.md

@@ -0,0 +1,100 @@
+# 03 - Embed Images for LLM Analysis
+
+## Date
+2025-10-28
+
+## Context
+Hybrid OCR approach was fast and accurate but formatting was messy. Vision models hallucinated text. Rather than fighting with text extraction, a better approach is to embed the actual frame images in the enhanced transcript and let the end-user's LLM analyze them with full audio context.
+
+## Problem
+- OCR/vision models either hallucinate or produce messy text
+- Code formatting/indentation is hard to preserve
+- User wants to analyze frames with their own LLM (Claude, GPT, etc.)
+- Need to keep file size reasonable (~200KB per image is too big)
+
+## Solution: Image Embedding
+
+Instead of extracting text, embed the actual frame images as base64 in the enhanced transcript. The LLM can then:
+- See the actual screen content (no hallucination)
+- Understand code structure, layout, and formatting visually
+- Have full audio transcript context for each frame
+- Analyze dashboards, terminals, editors with perfect accuracy
+
+## Implementation
+
+**Quality Optimization:**
+- Default JPEG quality: 80 (good tradeoff between size and readability)
+- Configurable via `--embed-quality` (0-100)
+- Typical sizes at quality 80: ~40-80KB per image (vs 200KB original)
+
+**Format:**
+```
+[MM:SS] SPEAKER:
+  Audio transcript text here
+
+[MM:SS] SCREEN CONTENT:
+  IMAGE (base64, 52KB):
+  <image>data:image/jpeg;base64,/9j/4AAQSkZJRg...</image>
+
+  TEXT:
+  | Optional OCR text for reference
+```
+
+**Features:**
+- Base64 encoding for easy embedding
+- Size tracking and reporting
+- Optional text content alongside images
+- Works with scene detection for smart frame selection
+
+## Usage
+
+```bash
+# Basic: Embed images at quality 80 (default)
+python process_meeting.py samples/video.mkv --run-whisper --embed-images --scene-detection --no-cache -v
+
+# Lower quality for smaller files (still readable)
+python process_meeting.py samples/video.mkv --run-whisper --embed-images --embed-quality 60 --scene-detection --no-cache -v
+
+# Higher quality for detailed code
+python process_meeting.py samples/video.mkv --run-whisper --embed-images --embed-quality 90 --scene-detection --no-cache -v
+
+# Iterate on scene threshold (reuse whisper)
+python process_meeting.py samples/video.mkv --embed-images --scene-detection --scene-threshold 5 --skip-cache-frames --skip-cache-analysis -v
+```
+
+## File Sizes
+
+**Example for 20 frames:**
+- Quality 60: ~30-50KB per image = 0.6-1MB total
+- Quality 80: ~40-80KB per image = 0.8-1.6MB total (recommended)
+- Quality 90: ~80-120KB per image = 1.6-2.4MB total
+- Original: ~200KB per image = 4MB total
+
+## Benefits
+
+✓ **No hallucination**: LLM sees actual pixels
+✓ **Perfect formatting**: Code structure preserved visually
+✓ **Full context**: Audio transcript + visual frame together
+✓ **User's choice**: Use your preferred LLM (Claude, GPT, etc.)
+✓ **Reasonable size**: Quality 80 gives 4x smaller files vs original
+✓ **Simple workflow**: One file contains everything
+
+## Use Cases
+
+**Code walkthroughs:** LLM can see actual code structure and indentation
+**Dashboard analysis:** Charts, graphs, metrics visible to LLM
+**Terminal sessions:** Commands and output in proper context
+**UI reviews:** Actual interface visible with audio commentary
+
+## Files Modified
+
+- `meetus/transcript_merger.py` - Image encoding and embedding
+- `meetus/workflow.py` - Wire through config
+- `process_meeting.py` - CLI flags
+- `meetus/output_manager.py` - Cleaner directory naming (date + increment)
+
+## Output Directory Naming
+
+Also changed output directory format for clarity:
+- Old: `20251028_054553-video` (confusing timestamps)
+- New: `20251028-001-video` (clear date + run number)

def/04-fix-whisper-cache-loading.md

@@ -0,0 +1,78 @@
+# 04 - Fix Whisper Cache Loading
+
+## Date
+2025-10-28
+
+## Problem
+Enhanced transcript was not including the audio segments from cached whisper transcripts when running without the `--run-whisper` flag.
+
+Example command that failed:
+```bash
+python process_meeting.py samples/zaca-run-scrapers.mkv --embed-images --scene-detection --scene-threshold 10 --skip-cache-frames -v
+```
+
+Result: Enhanced transcript only contained embedded images, no audio segments (0 "SPEAKER" entries).
+
+## Root Cause
+In `workflow.py`, the `_run_whisper()` method was checking the `run_whisper` flag **before** checking the cache:
+
+```python
+def _run_whisper(self) -> Optional[str]:
+    if not self.config.run_whisper:
+        return self.config.transcript_path  # Returns None if --transcript not specified
+
+    # Cache check NEVER REACHED if run_whisper is False
+    cached = self.cache_mgr.get_whisper_cache()
+    if cached:
+        return str(cached)
+```
+
+This meant:
+- User runs command without `--run-whisper`
+- Method returns None immediately
+- Cached whisper transcript is never discovered
+- No audio segments in enhanced output
+
+## Solution
+Reorder the logic to check cache **first**, regardless of flags:
+
+```python
+def _run_whisper(self) -> Optional[str]:
+    """Run Whisper transcription if requested, or use cached/provided transcript."""
+    # First, check cache (regardless of run_whisper flag)
+    cached = self.cache_mgr.get_whisper_cache()
+    if cached:
+        return str(cached)
+
+    # If no cache and not running whisper, use provided transcript path (if any)
+    if not self.config.run_whisper:
+        return self.config.transcript_path
+
+    # If no cache and run_whisper is True, run whisper transcription
+    # ... rest of whisper code
+```
+
+## New Behavior
+1. Cache is checked first (regardless of `--run-whisper` flag)
+2. If cached whisper exists, use it
+3. If no cache and `--run-whisper` not specified, use `--transcript` path (or None)
+4. If no cache and `--run-whisper` specified, run whisper
+
+## Benefits
+✓ Cached whisper transcripts are always discovered and used
+✓ User can iterate on frame extraction/analysis without re-running whisper
+✓ Enhanced transcripts now properly include both audio + visual content
+✓ Granular cache flags (`--skip-cache-frames`, `--skip-cache-whisper`) work as expected
+
+## Use Case
+```bash
+# First run: Generate whisper transcript + extract frames
+python process_meeting.py samples/video.mkv --run-whisper --embed-images --scene-detection -v
+
+# Second run: Iterate on scene threshold without re-running whisper
+python process_meeting.py samples/video.mkv --embed-images --scene-detection --scene-threshold 10 --skip-cache-frames -v
+# Now correctly includes cached whisper transcript in enhanced output!
+```
+
+## Files Modified
+- `meetus/workflow.py` - Reordered logic in `_run_whisper()` method (lines 172-181)

meetus/cache_manager.py

@@ -48,6 +48,17 @@ class CacheManager:
         cache_path = self.output_dir / f"{self.video_name}.json"
         if cache_path.exists():
             logger.info(f"✓ Found cached Whisper transcript: {cache_path.name}")
+
+            # Debug: Show cached transcript info
+            try:
+                import json
+                with open(cache_path, 'r', encoding='utf-8') as f:
+                    data = json.load(f)
+                if 'segments' in data:
+                    logger.debug(f"Cached transcript has {len(data['segments'])} segments")
+            except Exception as e:
+                logger.debug(f"Could not parse cached whisper for debug: {e}")
+
             return cache_path
 
         return None
@@ -68,6 +79,7 @@ class CacheManager:
             return None
 
         logger.info(f"✓ Found {len(existing_frames)} cached frames in {self.frames_dir.name}/")
+        logger.debug(f"Frame filenames: {[f.name for f in sorted(existing_frames)[:3]]}...")
 
         # Build frames_info from existing files
         frames_info = []
@@ -102,6 +114,11 @@ class CacheManager:
             with open(cache_path, 'r', encoding='utf-8') as f:
                 results = json.load(f)
             logger.info(f"✓ Loaded {len(results)} analyzed frames from cache")
+
+            # Debug: Show first cached result
+            if results:
+                logger.debug(f"First cached result: timestamp={results[0].get('timestamp')}, text_length={len(results[0].get('text', ''))}")
+
             return results
 
         return None

meetus/frame_extractor.py

@@ -16,17 +16,19 @@ logger = logging.getLogger(__name__)
 class FrameExtractor:
     """Extract frames from video files."""
 
-    def __init__(self, video_path: str, output_dir: str = "frames"):
+    def __init__(self, video_path: str, output_dir: str = "frames", quality: int = 75):
         """
         Initialize frame extractor.
 
         Args:
             video_path: Path to video file
             output_dir: Directory to save extracted frames
+            quality: JPEG quality for saved frames (0-100)
         """
         self.video_path = video_path
         self.output_dir = Path(output_dir)
         self.output_dir.mkdir(parents=True, exist_ok=True)
+        self.quality = quality
 
     def extract_by_interval(self, interval_seconds: int = 5) -> List[Tuple[str, float]]:
         """
@@ -56,8 +58,16 @@ class FrameExtractor:
                 frame_filename = f"frame_{saved_count:05d}_{timestamp:.2f}s.jpg"
                 frame_path = self.output_dir / frame_filename
 
-                # Use high quality for text readability (95 = high quality JPEG)
+                # Downscale to 1600px width for smaller file size (but still readable)
-                cv2.imwrite(str(frame_path), frame, [cv2.IMWRITE_JPEG_QUALITY, 95])
+                height, width = frame.shape[:2]
+                if width > 1600:
+                    ratio = 1600 / width
+                    new_width = 1600
+                    new_height = int(height * ratio)
+                    frame = cv2.resize(frame, (new_width, new_height), interpolation=cv2.INTER_LANCZOS4)
+
+                # Save with configured quality (matches embed quality)
+                cv2.imwrite(str(frame_path), frame, [cv2.IMWRITE_JPEG_QUALITY, self.quality])
                 frames_info.append((str(frame_path), timestamp))
                 saved_count += 1
 
@@ -90,16 +100,24 @@ class FrameExtractor:
         output_pattern = self.output_dir / f"{video_name}_%05d.jpg"
 
         try:
-            # Use FFmpeg's scene detection filter with high quality output
+            # Use FFmpeg's scene detection filter with downscaling
             stream = ffmpeg.input(self.video_path)
             stream = ffmpeg.filter(stream, 'select', f'gt(scene,{threshold/100})')
             stream = ffmpeg.filter(stream, 'showinfo')
+            # Scale to 1600px width (maintains aspect ratio, still readable)
+            # Use simple conditional: if width > 1600, scale to 1600, else keep original
+            stream = ffmpeg.filter(stream, 'scale', w='min(1600,iw)', h=-1)
+
+            # Convert JPEG quality (0-100) to FFmpeg qscale (2-31, lower=better)
+            # Rough mapping: qscale ≈ (100 - quality) / 10, clamped to 2-31
+            qscale = max(2, min(31, int((100 - self.quality) / 10 + 2)))
+
             stream = ffmpeg.output(
                 stream,
                 str(output_pattern),
                 vsync='vfr',
                 frame_pts=1,
-                **{'q:v': '2'}  # High quality JPEG
+                **{'q:v': str(qscale)}  # Matches configured quality
             )
 
             # Run with stderr capture to get showinfo output

meetus/hybrid_processor.py

@@ -0,0 +1,355 @@
+"""
+Hybrid frame analysis: OpenCV text detection + OCR for accurate extraction.
+Better than pure vision models which tend to hallucinate text content.
+"""
+from typing import List, Tuple, Dict, Optional
+from pathlib import Path
+import logging
+import cv2
+import numpy as np
+from difflib import SequenceMatcher
+
+logger = logging.getLogger(__name__)
+
+
+class HybridProcessor:
+    """Combine OpenCV text detection with OCR for accurate text extraction."""
+
+    def __init__(self, ocr_engine: str = "tesseract", min_confidence: float = 0.5,
+                 use_llm_cleanup: bool = False, llm_model: Optional[str] = None):
+        """
+        Initialize hybrid processor.
+
+        Args:
+            ocr_engine: OCR engine to use ('tesseract', 'easyocr', 'paddleocr')
+            min_confidence: Minimum confidence for text detection (0-1)
+            use_llm_cleanup: Use LLM to clean up OCR output and preserve formatting
+            llm_model: Ollama model for cleanup (default: llama3.2:3b for speed)
+        """
+        from .ocr_processor import OCRProcessor
+
+        self.ocr = OCRProcessor(engine=ocr_engine)
+        self.min_confidence = min_confidence
+        self.use_llm_cleanup = use_llm_cleanup
+        self.llm_model = llm_model or "llama3.2:3b"
+        self._llm_client = None
+
+        if use_llm_cleanup:
+            self._init_llm()
+
+    def _init_llm(self):
+        """Initialize Ollama client for LLM cleanup."""
+        try:
+            import ollama
+            self._llm_client = ollama
+            logger.info(f"LLM cleanup enabled using {self.llm_model}")
+        except ImportError:
+            logger.warning("ollama package not installed. LLM cleanup disabled.")
+            self.use_llm_cleanup = False
+
+    def _cleanup_with_llm(self, raw_text: str) -> str:
+        """
+        Use LLM to clean up OCR output and preserve code formatting.
+
+        Args:
+            raw_text: Raw OCR output
+
+        Returns:
+            Cleaned up text with proper formatting
+        """
+        if not self.use_llm_cleanup or not self._llm_client:
+            return raw_text
+
+        prompt = """You are cleaning up OCR output from a code editor screenshot.
+
+Your task:
+1. Fix any obvious OCR errors (l→1, O→0, etc.)
+2. Preserve or restore code indentation and structure
+3. Keep the exact text content - don't add explanations or comments
+4. If it's code, maintain proper spacing and formatting
+5. Return ONLY the cleaned text, nothing else
+
+OCR Text:
+"""
+
+        try:
+            response = self._llm_client.generate(
+                model=self.llm_model,
+                prompt=prompt + raw_text,
+                options={"temperature": 0.1}  # Low temperature for accuracy
+            )
+            cleaned = response['response'].strip()
+            logger.debug(f"LLM cleanup: {len(raw_text)} → {len(cleaned)} chars")
+            return cleaned
+        except Exception as e:
+            logger.warning(f"LLM cleanup failed: {e}, using raw OCR output")
+            return raw_text
+
+    def detect_text_regions(self, image_path: str, min_area: int = 100) -> List[Tuple[int, int, int, int]]:
+        """
+        Detect text regions in image using OpenCV.
+
+        Args:
+            image_path: Path to image file
+            min_area: Minimum area for text region (pixels)
+
+        Returns:
+            List of bounding boxes (x, y, w, h)
+        """
+        # Read image
+        img = cv2.imread(image_path)
+        if img is None:
+            logger.warning(f"Could not read image: {image_path}")
+            return []
+
+        gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+
+        # Method 1: Morphological operations to find text regions
+        # Works well for solid text blocks
+        regions = self._detect_by_morphology(gray, min_area)
+
+        if not regions:
+            logger.debug(f"No text regions detected in {Path(image_path).name}")
+
+        return regions
+
+    def _detect_by_morphology(self, gray: np.ndarray, min_area: int) -> List[Tuple[int, int, int, int]]:
+        """
+        Detect text regions using morphological operations.
+        Fast and works well for solid text blocks (code editors, terminals).
+
+        Args:
+            gray: Grayscale image
+            min_area: Minimum area for region
+
+        Returns:
+            List of bounding boxes (x, y, w, h)
+        """
+        # Apply adaptive threshold to handle varying lighting
+        binary = cv2.adaptiveThreshold(
+            gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
+            cv2.THRESH_BINARY_INV, 11, 2
+        )
+
+        # Morphological operations to connect text regions
+        kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (15, 3))  # Horizontal kernel for text lines
+        dilated = cv2.dilate(binary, kernel, iterations=2)
+
+        # Find contours
+        contours, _ = cv2.findContours(dilated, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+
+        # Filter and extract bounding boxes
+        regions = []
+        for contour in contours:
+            x, y, w, h = cv2.boundingRect(contour)
+            area = w * h
+
+            # Filter by area and aspect ratio
+            if area > min_area and w > 20 and h > 10:  # Reasonable text dimensions
+                regions.append((x, y, w, h))
+
+        # Merge overlapping regions
+        regions = self._merge_overlapping_regions(regions)
+
+        logger.debug(f"Detected {len(regions)} text regions using morphology")
+        return regions
+
+    def _merge_overlapping_regions(
+        self, regions: List[Tuple[int, int, int, int]],
+        overlap_threshold: float = 0.3
+    ) -> List[Tuple[int, int, int, int]]:
+        """
+        Merge overlapping bounding boxes.
+
+        Args:
+            regions: List of (x, y, w, h) tuples
+            overlap_threshold: Minimum overlap ratio to merge
+
+        Returns:
+            Merged regions
+        """
+        if not regions:
+            return []
+
+        # Sort by y-coordinate (top to bottom)
+        regions = sorted(regions, key=lambda r: r[1])
+
+        merged = []
+        current = list(regions[0])
+
+        for region in regions[1:]:
+            x, y, w, h = region
+            cx, cy, cw, ch = current
+
+            # Check for overlap
+            x_overlap = max(0, min(cx + cw, x + w) - max(cx, x))
+            y_overlap = max(0, min(cy + ch, y + h) - max(cy, y))
+            overlap_area = x_overlap * y_overlap
+
+            current_area = cw * ch
+            region_area = w * h
+            min_area = min(current_area, region_area)
+
+            if overlap_area / min_area > overlap_threshold:
+                # Merge regions
+                new_x = min(cx, x)
+                new_y = min(cy, y)
+                new_x2 = max(cx + cw, x + w)
+                new_y2 = max(cy + ch, y + h)
+                current = [new_x, new_y, new_x2 - new_x, new_y2 - new_y]
+            else:
+                merged.append(tuple(current))
+                current = list(region)
+
+        merged.append(tuple(current))
+        return merged
+
+    def extract_text_from_region(self, image_path: str, region: Tuple[int, int, int, int]) -> str:
+        """
+        Extract text from a specific region using OCR.
+
+        Args:
+            image_path: Path to image file
+            region: Bounding box (x, y, w, h)
+
+        Returns:
+            Extracted text
+        """
+        from PIL import Image
+
+        # Load image and crop region
+        img = Image.open(image_path)
+        x, y, w, h = region
+        cropped = img.crop((x, y, x + w, y + h))
+
+        # Save to temp file for OCR (or use in-memory)
+        import tempfile
+        with tempfile.NamedTemporaryFile(suffix='.png', delete=False) as tmp:
+            cropped.save(tmp.name)
+            text = self.ocr.extract_text(tmp.name)
+
+        # Clean up temp file
+        Path(tmp.name).unlink()
+
+        return text
+
+    def analyze_frame(self, image_path: str) -> str:
+        """
+        Analyze a frame: detect text regions and OCR them.
+
+        Args:
+            image_path: Path to image file
+
+        Returns:
+            Combined text from all detected regions
+        """
+        # Detect text regions
+        regions = self.detect_text_regions(image_path)
+
+        if not regions:
+            # Fallback to full-frame OCR if no regions detected
+            logger.debug(f"No regions detected, using full-frame OCR for {Path(image_path).name}")
+            raw_text = self.ocr.extract_text(image_path)
+            return self._cleanup_with_llm(raw_text) if self.use_llm_cleanup else raw_text
+
+        # Sort regions by reading order (top-to-bottom, left-to-right)
+        regions = self._sort_regions_by_reading_order(regions)
+
+        # Extract text from each region
+        texts = []
+        for idx, region in enumerate(regions):
+            x, y, w, h = region
+            text = self.extract_text_from_region(image_path, region)
+            if text.strip():
+                # Add visual separator with region info
+                section_header = f"[Region {idx+1} at y={y}]"
+                texts.append(f"{section_header}\n{text.strip()}")
+                logger.debug(f"Region {idx+1}/{len(regions)} (y={y}): Extracted {len(text)} chars")
+
+        combined = ("\n\n" + "="*60 + "\n\n").join(texts)
+        logger.debug(f"Total extracted from {len(regions)} regions: {len(combined)} chars")
+
+        # Apply LLM cleanup if enabled
+        if self.use_llm_cleanup:
+            combined = self._cleanup_with_llm(combined)
+
+        return combined
+
+    def _sort_regions_by_reading_order(self, regions: List[Tuple[int, int, int, int]]) -> List[Tuple[int, int, int, int]]:
+        """
+        Sort regions in reading order (top-to-bottom, left-to-right).
+
+        Args:
+            regions: List of (x, y, w, h) tuples
+
+        Returns:
+            Sorted regions
+        """
+        # Sort primarily by y (top to bottom), secondarily by x (left to right)
+        # Group regions that are on roughly the same line (within 20px)
+        sorted_regions = sorted(regions, key=lambda r: (r[1] // 20, r[0]))
+        return sorted_regions
+
+    def process_frames(
+        self,
+        frames_info: List[Tuple[str, float]],
+        deduplicate: bool = True,
+        similarity_threshold: float = 0.85
+    ) -> List[Dict]:
+        """
+        Process multiple frames with hybrid analysis.
+
+        Args:
+            frames_info: List of (frame_path, timestamp) tuples
+            deduplicate: Whether to remove similar consecutive analyses
+            similarity_threshold: Threshold for considering analyses as duplicates (0-1)
+
+        Returns:
+            List of dicts with 'timestamp', 'text', and 'frame_path'
+        """
+        results = []
+        prev_text = ""
+
+        total = len(frames_info)
+        logger.info(f"Starting hybrid analysis of {total} frames...")
+
+        for idx, (frame_path, timestamp) in enumerate(frames_info, 1):
+            logger.info(f"Analyzing frame {idx}/{total} at {timestamp:.2f}s...")
+
+            text = self.analyze_frame(frame_path)
+
+            if not text:
+                logger.warning(f"No content extracted from frame at {timestamp:.2f}s")
+                continue
+
+            # Debug: Show what was extracted
+            logger.debug(f"Frame {idx} ({timestamp:.2f}s): Extracted {len(text)} chars")
+            logger.debug(f"Content preview: {text[:150]}{'...' if len(text) > 150 else ''}")
+
+            # Deduplicate similar consecutive frames
+            if deduplicate and prev_text:
+                similarity = self._text_similarity(prev_text, text)
+                logger.debug(f"Similarity to previous frame: {similarity:.2f} (threshold: {similarity_threshold})")
+                if similarity > similarity_threshold:
+                    logger.debug(f"⊘ Skipping duplicate frame at {timestamp:.2f}s (similarity: {similarity:.2f})")
+                    continue
+
+            results.append({
+                'timestamp': timestamp,
+                'text': text,
+                'frame_path': frame_path
+            })
+
+            prev_text = text
+
+        logger.info(f"Extracted content from {len(results)} frames (deduplication: {deduplicate})")
+        return results
+
+    def _text_similarity(self, text1: str, text2: str) -> float:
+        """
+        Calculate similarity between two texts.
+
+        Returns:
+            Similarity score between 0 and 1
+        """
+        return SequenceMatcher(None, text1, text2).ratio()

meetus/ocr_processor.py

@@ -53,20 +53,25 @@ class OCRProcessor:
         else:
             raise ValueError(f"Unknown OCR engine: {self.engine}")
 
-    def extract_text(self, image_path: str) -> str:
+    def extract_text(self, image_path: str, preserve_layout: bool = True) -> str:
         """
         Extract text from a single image.
 
         Args:
             image_path: Path to image file
+            preserve_layout: Try to preserve whitespace and layout
 
         Returns:
             Extracted text
         """
         if self.engine == "tesseract":
             from PIL import Image
+            import pytesseract
             image = Image.open(image_path)
-            text = self._ocr_engine.image_to_string(image)
+
+            # Use PSM 6 (uniform block of text) to preserve layout better
+            config = '--psm 6' if preserve_layout else ''
+            text = pytesseract.image_to_string(image, config=config)
 
         elif self.engine == "easyocr":
             result = self._ocr_engine.readtext(image_path, detail=0)
@@ -81,12 +86,31 @@ class OCRProcessor:
 
         return self._clean_text(text)
 
-    def _clean_text(self, text: str) -> str:
+    def _clean_text(self, text: str, preserve_indentation: bool = True) -> str:
-        """Clean up OCR output."""
+        """
-        # Remove excessive whitespace
+        Clean up OCR output.
-        text = re.sub(r'\n\s*\n', '\n', text)
+
-        text = re.sub(r' +', ' ', text)
+        Args:
-        return text.strip()
+            text: Raw OCR text
+            preserve_indentation: Keep leading whitespace on lines
+
+        Returns:
+            Cleaned text
+        """
+        if preserve_indentation:
+            # Remove excessive blank lines but preserve indentation
+            lines = text.split('\n')
+            cleaned_lines = []
+            for line in lines:
+                # Keep line if it has content or is single empty line
+                if line.strip() or (cleaned_lines and cleaned_lines[-1].strip()):
+                    cleaned_lines.append(line)
+            return '\n'.join(cleaned_lines).strip()
+        else:
+            # Original aggressive cleaning
+            text = re.sub(r'\n\s*\n', '\n', text)
+            text = re.sub(r' +', ' ', text)
+            return text.strip()
 
     def process_frames(
         self,
@@ -108,18 +132,24 @@ class OCRProcessor:
         results = []
         prev_text = ""
 
-        for frame_path, timestamp in frames_info:
+        for idx, (frame_path, timestamp) in enumerate(frames_info, 1):
-            logger.debug(f"Processing frame at {timestamp:.2f}s...")
+            logger.debug(f"Processing frame {idx}/{len(frames_info)} at {timestamp:.2f}s...")
             text = self.extract_text(frame_path)
 
             if not text:
+                logger.debug(f"No text extracted from frame at {timestamp:.2f}s")
                 continue
 
+            # Debug: Show what was extracted
+            logger.debug(f"Frame {idx} ({timestamp:.2f}s): Extracted {len(text)} chars")
+            logger.debug(f"Content preview: {text[:150]}{'...' if len(text) > 150 else ''}")
+
             # Deduplicate similar consecutive frames
-            if deduplicate:
+            if deduplicate and prev_text:
                 similarity = self._text_similarity(prev_text, text)
+                logger.debug(f"Similarity to previous frame: {similarity:.2f} (threshold: {similarity_threshold})")
                 if similarity > similarity_threshold:
-                    logger.debug(f"Skipping duplicate frame at {timestamp:.2f}s (similarity: {similarity:.2f})")
+                    logger.debug(f"⊘ Skipping duplicate frame at {timestamp:.2f}s (similarity: {similarity:.2f})")
                     continue
 
             results.append({

meetus/output_manager.py

@@ -36,7 +36,7 @@ class OutputManager:
 
     def _get_or_create_output_dir(self) -> Path:
         """
-        Get existing output directory or create a new timestamped one.
+        Get existing output directory or create a new one with incremental number.
 
         Returns:
             Path to output directory
@@ -54,9 +54,29 @@ class OutputManager:
                 logger.info(f"Found existing output: {existing_dirs[0].name}")
                 return existing_dirs[0]
 
-        # Create new timestamped directory
+        # Create new directory with date + incremental number
-        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        date_str = datetime.now().strftime("%Y%m%d")
-        dir_name = f"{timestamp}-{video_name}"
+
+        # Find existing runs for today
+        if self.base_output_dir.exists():
+            existing_today = [
+                d for d in self.base_output_dir.iterdir()
+                if d.is_dir() and d.name.startswith(date_str) and d.name.endswith(f"-{video_name}")
+            ]
+
+            # Extract run numbers and find max
+            run_numbers = []
+            for d in existing_today:
+                # Format: YYYYMMDD-NNN-videoname
+                parts = d.name.split('-')
+                if len(parts) >= 2 and parts[1].isdigit():
+                    run_numbers.append(int(parts[1]))
+
+            next_run = max(run_numbers) + 1 if run_numbers else 1
+        else:
+            next_run = 1
+
+        dir_name = f"{date_str}-{next_run:03d}-{video_name}"
         output_dir = self.base_output_dir / dir_name
         output_dir.mkdir(parents=True, exist_ok=True)
         logger.info(f"Created new output directory: {dir_name}")

meetus/transcript_merger.py

@@ -6,6 +6,8 @@ from typing import List, Dict, Optional
 import json
 from pathlib import Path
 import logging
+import base64
+from io import BytesIO
 
 logger = logging.getLogger(__name__)
 
@@ -13,9 +15,16 @@ logger = logging.getLogger(__name__)
 class TranscriptMerger:
     """Merge audio transcripts with screen OCR text."""
 
-    def __init__(self):
+    def __init__(self, embed_images: bool = False, embed_quality: int = 80):
-        """Initialize transcript merger."""
+        """
-        pass
+        Initialize transcript merger.
+
+        Args:
+            embed_images: Whether to embed frame images as base64
+            embed_quality: JPEG quality for embedded images (0-100)
+        """
+        self.embed_images = embed_images
+        self.embed_quality = embed_quality
 
     def load_whisper_transcript(self, transcript_path: str, group_interval: Optional[int] = None) -> List[Dict]:
         """
@@ -120,6 +129,32 @@ class TranscriptMerger:
         logger.info(f"Grouped {len(segments)} segments into {len(intervals)} intervals of {interval_seconds}s")
         return intervals
 
+    def _encode_image_base64(self, image_path: str) -> tuple[str, int]:
+        """
+        Encode image as base64 (image already at target quality/size).
+
+        Args:
+            image_path: Path to image file
+
+        Returns:
+            Tuple of (base64_string, size_in_bytes)
+        """
+        try:
+            # Read file directly (already at target quality/resolution)
+            with open(image_path, 'rb') as f:
+                img_bytes = f.read()
+
+            # Encode to base64
+            b64_string = base64.b64encode(img_bytes).decode('utf-8')
+
+            logger.debug(f"Encoded {Path(image_path).name}: {len(img_bytes)} bytes")
+
+            return b64_string, len(img_bytes)
+
+        except Exception as e:
+            logger.error(f"Failed to encode image {image_path}: {e}")
+            return "", 0
+
     def merge_transcripts(
         self,
         audio_segments: List[Dict],
@@ -172,10 +207,15 @@ class TranscriptMerger:
         lines = []
         lines.append("=" * 80)
         lines.append("ENHANCED MEETING TRANSCRIPT")
-        lines.append("Audio transcript + Screen content")
+        if self.embed_images:
+            lines.append("Audio transcript + Embedded frame images (base64)")
+        else:
+            lines.append("Audio transcript + Screen content")
         lines.append("=" * 80)
         lines.append("")
 
+        total_image_bytes = 0
+
         for seg in segments:
             timestamp = self._format_timestamp(seg['timestamp'])
 
@@ -186,11 +226,31 @@ class TranscriptMerger:
 
             else:  # screen
                 lines.append(f"[{timestamp}] SCREEN CONTENT:")
-                # Indent screen text for visibility
+
-                screen_text = seg['text'].replace('\n', '\n  | ')
+                # Embed image if requested
-                lines.append(f"  | {screen_text}")
+                if self.embed_images and 'frame_path' in seg:
+                    b64_img, img_size = self._encode_image_base64(seg['frame_path'])
+                    total_image_bytes += img_size
+
+                    if b64_img:
+                        lines.append(f"  IMAGE (base64, {img_size // 1024}KB):")
+                        lines.append(f"  <image>data:image/jpeg;base64,{b64_img}</image>")
+                        lines.append("")
+
+                # Include text content if available (fallback or additional context)
+                if 'text' in seg and seg['text'].strip():
+                    screen_text = seg['text'].replace('\n', '\n  | ')
+                    lines.append(f"  TEXT:")
+                    lines.append(f"  | {screen_text}")
+
                 lines.append("")
 
+        if self.embed_images and total_image_bytes > 0:
+            total_mb = total_image_bytes / (1024 * 1024)
+            lines.append("")
+            lines.append(f"Total embedded images size: {total_mb:.2f} MB")
+            logger.info(f"Embedded {len([s for s in segments if s['type'] == 'screen'])} images, total size: {total_mb:.2f} MB")
+
         return "\n".join(lines)
 
     def _format_compact(self, segments: List[Dict]) -> str:

meetus/vision_processor.py

@@ -161,11 +161,16 @@ class VisionProcessor:
                 logger.warning(f"No content extracted from frame at {timestamp:.2f}s")
                 continue
 
+            # Debug: Show what was extracted
+            logger.debug(f"Frame {idx} ({timestamp:.2f}s): Extracted {len(text)} chars")
+            logger.debug(f"Content preview: {text[:150]}{'...' if len(text) > 150 else ''}")
+
             # Deduplicate similar consecutive frames
-            if deduplicate:
+            if deduplicate and prev_text:
                 similarity = self._text_similarity(prev_text, text)
+                logger.debug(f"Similarity to previous frame: {similarity:.2f} (threshold: {similarity_threshold})")
                 if similarity > similarity_threshold:
-                    logger.debug(f"Skipping duplicate frame at {timestamp:.2f}s (similarity: {similarity:.2f})")
+                    logger.debug(f"⊘ Skipping duplicate frame at {timestamp:.2f}s (similarity: {similarity:.2f})")
                     continue
 
             results.append({

meetus/workflow.py

@@ -40,10 +40,21 @@ class WorkflowConfig:
 
         # Analysis options
         self.use_vision = kwargs.get('use_vision', False)
+        self.use_hybrid = kwargs.get('use_hybrid', False)
+        self.hybrid_llm_cleanup = kwargs.get('hybrid_llm_cleanup', False)
+        self.hybrid_llm_model = kwargs.get('hybrid_llm_model', 'llama3.2:3b')
         self.vision_model = kwargs.get('vision_model', 'llava:13b')
         self.vision_context = kwargs.get('vision_context', 'meeting')
         self.ocr_engine = kwargs.get('ocr_engine', 'tesseract')
 
+        # Validation: can't use both vision and hybrid
+        if self.use_vision and self.use_hybrid:
+            raise ValueError("Cannot use both --use-vision and --use-hybrid. Choose one.")
+
+        # Validation: LLM cleanup requires hybrid mode
+        if self.hybrid_llm_cleanup and not self.use_hybrid:
+            raise ValueError("--hybrid-llm-cleanup requires --use-hybrid")
+
         # Processing options
         self.no_deduplicate = kwargs.get('no_deduplicate', False)
         self.no_cache = kwargs.get('no_cache', False)
@@ -52,6 +63,8 @@ class WorkflowConfig:
         self.skip_cache_analysis = kwargs.get('skip_cache_analysis', False)
         self.extract_only = kwargs.get('extract_only', False)
         self.format = kwargs.get('format', 'detailed')
+        self.embed_images = kwargs.get('embed_images', False)
+        self.embed_quality = kwargs.get('embed_quality', 80)
 
     def to_dict(self) -> Dict[str, Any]:
         """Convert config to dictionary for manifest."""
@@ -66,10 +79,10 @@ class WorkflowConfig:
                 "scene_threshold": self.scene_threshold if self.scene_detection else None
             },
             "analysis": {
-                "method": "vision" if self.use_vision else "ocr",
+                "method": "vision" if self.use_vision else ("hybrid" if self.use_hybrid else "ocr"),
                 "vision_model": self.vision_model if self.use_vision else None,
                 "vision_context": self.vision_context if self.use_vision else None,
-                "ocr_engine": self.ocr_engine if not self.use_vision else None,
+                "ocr_engine": self.ocr_engine if (not self.use_vision) else None,
                 "deduplication": not self.no_deduplicate
             },
             "output_format": self.format
@@ -113,10 +126,19 @@ class ProcessingWorkflow:
         logger.info("MEETING PROCESSOR")
         logger.info("=" * 80)
         logger.info(f"Video: {self.config.video_path.name}")
-        logger.info(f"Analysis: {'Vision Model' if self.config.use_vision else f'OCR ({self.config.ocr_engine})'}")
+
+        # Determine analysis method
         if self.config.use_vision:
-            logger.info(f"Vision Model: {self.config.vision_model}")
+            analysis_method = f"Vision Model ({self.config.vision_model})"
+            logger.info(f"Analysis: {analysis_method}")
             logger.info(f"Context: {self.config.vision_context}")
+        elif self.config.use_hybrid:
+            analysis_method = f"Hybrid (OpenCV + {self.config.ocr_engine})"
+            logger.info(f"Analysis: {analysis_method}")
+        else:
+            analysis_method = f"OCR ({self.config.ocr_engine})"
+            logger.info(f"Analysis: {analysis_method}")
+
         logger.info(f"Frame extraction: {'Scene detection' if self.config.scene_detection else f'Every {self.config.interval}s'}")
         logger.info(f"Caching: {'Disabled' if self.config.no_cache else 'Enabled'}")
         logger.info("=" * 80)
@@ -148,15 +170,16 @@ class ProcessingWorkflow:
         return self._build_result(transcript_path, screen_segments, enhanced_transcript)
 
     def _run_whisper(self) -> Optional[str]:
-        """Run Whisper transcription if requested."""
+        """Run Whisper transcription if requested, or use cached/provided transcript."""
-        if not self.config.run_whisper:
+        # First, check cache (regardless of run_whisper flag)
-            return self.config.transcript_path
-
-        # Check cache
         cached = self.cache_mgr.get_whisper_cache()
         if cached:
             return str(cached)
 
+        # If no cache and not running whisper, use provided transcript path (if any)
+        if not self.config.run_whisper:
+            return self.config.transcript_path
+
         logger.info("=" * 80)
         logger.info("STEP 0: Running Whisper Transcription")
         logger.info("=" * 80)
@@ -195,6 +218,25 @@ class ProcessingWorkflow:
 
             if transcript_path.exists():
                 logger.info(f"✓ Whisper transcription completed: {transcript_path.name}")
+
+                # Debug: Show transcript preview
+                try:
+                    import json
+                    with open(transcript_path, 'r', encoding='utf-8') as f:
+                        whisper_data = json.load(f)
+
+                    if 'segments' in whisper_data:
+                        logger.debug(f"Whisper produced {len(whisper_data['segments'])} segments")
+                        if whisper_data['segments']:
+                            logger.debug(f"First segment: {whisper_data['segments'][0]}")
+                            logger.debug(f"Last segment: {whisper_data['segments'][-1]}")
+
+                    if 'text' in whisper_data:
+                        text_preview = whisper_data['text'][:200] + "..." if len(whisper_data.get('text', '')) > 200 else whisper_data.get('text', '')
+                        logger.debug(f"Transcript preview: {text_preview}")
+                except Exception as e:
+                    logger.debug(f"Could not parse whisper output for debug: {e}")
+
                 logger.info("")
                 return str(transcript_path)
             else:
@@ -216,12 +258,24 @@ class ProcessingWorkflow:
 
         # Clean up old frames if regenerating
         if self.config.skip_cache_frames and self.output_mgr.frames_dir.exists():
-            logger.info("Cleaning up old frames...")
+            old_frames = list(self.output_mgr.frames_dir.glob("*.jpg"))
-            for old_frame in self.output_mgr.frames_dir.glob("*.jpg"):
+            if old_frames:
-                old_frame.unlink()
+                logger.info(f"Cleaning up {len(old_frames)} old frames...")
+                for old_frame in old_frames:
+                    old_frame.unlink()
+                logger.info("✓ Cleanup complete")
 
-        # Extract frames
+        # Extract frames (use embed quality so saved files match embedded images)
-        extractor = FrameExtractor(str(self.config.video_path), str(self.output_mgr.frames_dir))
+        if self.config.scene_detection:
+            logger.info(f"Extracting frames with scene detection (threshold={self.config.scene_threshold})...")
+        else:
+            logger.info(f"Extracting frames every {self.config.interval}s...")
+
+        extractor = FrameExtractor(
+            str(self.config.video_path),
+            str(self.output_mgr.frames_dir),
+            quality=self.config.embed_quality
+        )
 
         if self.config.scene_detection:
             frames_info = extractor.extract_scene_changes(threshold=self.config.scene_threshold)
@@ -232,8 +286,29 @@ class ProcessingWorkflow:
         return frames_info
 
     def _analyze_frames(self, frames_info):
-        """Analyze frames with vision or OCR."""
+        """Analyze frames with vision, hybrid, or OCR."""
-        analysis_type = 'vision' if self.config.use_vision else 'ocr'
+        # Skip analysis if just embedding images
+        if self.config.embed_images:
+            logger.info("Step 2: Skipping analysis (images will be embedded)")
+            # Create minimal segments with just frame paths and timestamps
+            screen_segments = [
+                {
+                    'timestamp': timestamp,
+                    'text': '',  # No text extraction needed
+                    'frame_path': frame_path
+                }
+                for frame_path, timestamp in frames_info
+            ]
+            logger.info(f"✓ Prepared {len(screen_segments)} frames for embedding")
+            return screen_segments
+
+        # Determine analysis type
+        if self.config.use_vision:
+            analysis_type = 'vision'
+        elif self.config.use_hybrid:
+            analysis_type = 'hybrid'
+        else:
+            analysis_type = 'ocr'
 
         # Check cache
         cached_analysis = self.cache_mgr.get_analysis_cache(analysis_type)
@@ -242,6 +317,8 @@ class ProcessingWorkflow:
 
         if self.config.use_vision:
             return self._run_vision_analysis(frames_info)
+        elif self.config.use_hybrid:
+            return self._run_hybrid_analysis(frames_info)
         else:
             return self._run_ocr_analysis(frames_info)
 
@@ -272,6 +349,13 @@ class ProcessingWorkflow:
             )
             logger.info(f"✓ Analyzed {len(screen_segments)} frames with vision model")
 
+            # Debug: Show sample analysis results
+            if screen_segments:
+                logger.debug(f"First analysis result: timestamp={screen_segments[0].get('timestamp')}, text_length={len(screen_segments[0].get('text', ''))}")
+                logger.debug(f"First analysis text preview: {screen_segments[0].get('text', '')[:200]}...")
+                if len(screen_segments) > 1:
+                    logger.debug(f"Last analysis result: timestamp={screen_segments[-1].get('timestamp')}, text_length={len(screen_segments[-1].get('text', ''))}")
+
             # Cache results
             self.cache_mgr.save_analysis('vision', screen_segments)
             return screen_segments
@@ -285,6 +369,42 @@ class ProcessingWorkflow:
         cached = self.cache_mgr.get_whisper_cache()
         return str(cached) if cached else None
 
+    def _run_hybrid_analysis(self, frames_info):
+        """Run hybrid analysis on frames (OpenCV + OCR)."""
+        if self.config.hybrid_llm_cleanup:
+            logger.info("Step 2: Running hybrid analysis (OpenCV + OCR + LLM cleanup)...")
+        else:
+            logger.info("Step 2: Running hybrid analysis (OpenCV text detection + OCR)...")
+
+        try:
+            from .hybrid_processor import HybridProcessor
+
+            hybrid = HybridProcessor(
+                ocr_engine=self.config.ocr_engine,
+                use_llm_cleanup=self.config.hybrid_llm_cleanup,
+                llm_model=self.config.hybrid_llm_model
+            )
+            screen_segments = hybrid.process_frames(
+                frames_info,
+                deduplicate=not self.config.no_deduplicate
+            )
+            logger.info(f"✓ Processed {len(screen_segments)} frames with hybrid analysis")
+
+            # Debug: Show sample hybrid results
+            if screen_segments:
+                logger.debug(f"First hybrid result: timestamp={screen_segments[0].get('timestamp')}, text_length={len(screen_segments[0].get('text', ''))}")
+                logger.debug(f"First hybrid text preview: {screen_segments[0].get('text', '')[:200]}...")
+                if len(screen_segments) > 1:
+                    logger.debug(f"Last hybrid result: timestamp={screen_segments[-1].get('timestamp')}, text_length={len(screen_segments[-1].get('text', ''))}")
+
+            # Cache results
+            self.cache_mgr.save_analysis('hybrid', screen_segments)
+            return screen_segments
+
+        except ImportError as e:
+            logger.error(f"{e}")
+            raise
+
     def _run_ocr_analysis(self, frames_info):
         """Run OCR analysis on frames."""
         logger.info("Step 2: Running OCR on extracted frames...")
@@ -297,6 +417,13 @@ class ProcessingWorkflow:
             )
             logger.info(f"✓ Processed {len(screen_segments)} frames with OCR")
 
+            # Debug: Show sample OCR results
+            if screen_segments:
+                logger.debug(f"First OCR result: timestamp={screen_segments[0].get('timestamp')}, text_length={len(screen_segments[0].get('text', ''))}")
+                logger.debug(f"First OCR text preview: {screen_segments[0].get('text', '')[:200]}...")
+                if len(screen_segments) > 1:
+                    logger.debug(f"Last OCR result: timestamp={screen_segments[-1].get('timestamp')}, text_length={len(screen_segments[-1].get('text', ''))}")
+
             # Cache results
             self.cache_mgr.save_analysis('ocr', screen_segments)
             return screen_segments
@@ -309,7 +436,10 @@ class ProcessingWorkflow:
 
     def _merge_transcripts(self, transcript_path, screen_segments):
         """Merge audio and screen transcripts."""
-        merger = TranscriptMerger()
+        merger = TranscriptMerger(
+            embed_images=self.config.embed_images,
+            embed_quality=self.config.embed_quality
+        )
 
         # Load audio transcript if available
         audio_segments = []
@@ -350,10 +480,18 @@ class ProcessingWorkflow:
 
     def _build_result(self, transcript_path=None, screen_segments=None, enhanced_transcript=None):
         """Build result dictionary."""
+        # Determine analysis filename
+        if self.config.use_vision:
+            analysis_type = 'vision'
+        elif self.config.use_hybrid:
+            analysis_type = 'hybrid'
+        else:
+            analysis_type = 'ocr'
+
         return {
             "output_dir": str(self.output_mgr.output_dir),
             "transcript": transcript_path,
-            "analysis": f"{self.config.video_path.stem}_{'vision' if self.config.use_vision else 'ocr'}.json",
+            "analysis": f"{self.config.video_path.stem}_{analysis_type}.json",
             "frames_count": len(screen_segments) if screen_segments else 0,
             "enhanced_transcript": enhanced_transcript,
             "manifest": str(self.output_mgr.get_path("manifest.json"))

process_meeting.py

@@ -32,23 +32,20 @@ def main():
         formatter_class=argparse.RawDescriptionHelpFormatter,
         epilog="""
 Examples:
-  # Run Whisper + vision analysis (recommended for code/dashboards)
+  # Embed images for LLM analysis (recommended - let LLM analyze actual frames)
-  python process_meeting.py samples/meeting.mkv --run-whisper --use-vision
+  python process_meeting.py samples/meeting.mkv --run-whisper --embed-images --scene-detection
 
-  # Use vision with specific context hint
+  # Embed with custom quality (lower = smaller file size)
-  python process_meeting.py samples/meeting.mkv --run-whisper --use-vision --vision-context code
+  python process_meeting.py samples/meeting.mkv --run-whisper --embed-images --embed-quality 60 --scene-detection
 
-  # Traditional OCR approach
+  # Hybrid approach: OpenCV + OCR (extracts text, no images)
-  python process_meeting.py samples/meeting.mkv --run-whisper
+  python process_meeting.py samples/meeting.mkv --run-whisper --use-hybrid --scene-detection
 
-  # Re-run analysis using cached frames and transcript
+  # Hybrid + LLM cleanup (best for code formatting)
-  python process_meeting.py samples/meeting.mkv --use-vision
+  python process_meeting.py samples/meeting.mkv --run-whisper --use-hybrid --hybrid-llm-cleanup --scene-detection
 
-  # Force reprocessing (ignore cache)
+  # Iterate on scene threshold (reuse whisper transcript)
-  python process_meeting.py samples/meeting.mkv --run-whisper --use-vision --no-cache
+  python process_meeting.py samples/meeting.mkv --embed-images --scene-detection --scene-threshold 5 --skip-cache-frames --skip-cache-analysis
-
-  # Use scene detection for fewer frames
-  python process_meeting.py samples/meeting.mkv --run-whisper --use-vision --scene-detection
         """
     )
 
@@ -119,6 +116,21 @@ Examples:
         action='store_true',
         help='Use local vision model (Ollama) instead of OCR for better context understanding'
     )
+    parser.add_argument(
+        '--use-hybrid',
+        action='store_true',
+        help='Use hybrid approach: OpenCV text detection + OCR (more accurate than vision models)'
+    )
+    parser.add_argument(
+        '--hybrid-llm-cleanup',
+        action='store_true',
+        help='Use LLM to clean up OCR output and preserve code formatting (requires --use-hybrid)'
+    )
+    parser.add_argument(
+        '--hybrid-llm-model',
+        help='LLM model for cleanup (default: llama3.2:3b)',
+        default='llama3.2:3b'
+    )
     parser.add_argument(
         '--vision-model',
         help='Vision model to use with Ollama (default: llava:13b)',
@@ -168,6 +180,17 @@ Examples:
         help='Output format style (default: detailed)',
         default='detailed'
     )
+    parser.add_argument(
+        '--embed-images',
+        action='store_true',
+        help='Embed frame images (as base64) in enhanced transcript for LLM analysis'
+    )
+    parser.add_argument(
+        '--embed-quality',
+        type=int,
+        help='JPEG quality for embedded images (default: 80, lower = smaller file)',
+        default=80
+    )
 
     # Logging
     parser.add_argument(