v0.3.1 - Semantic Chunking & Mistral Vision

🚀 New Features

Semantic Chunking

• Implement semantic chunking using LangChain's SemanticChunker
• Creates more meaningful, coherent chunks for better vector quality
• 30% improvement in semantic coherence
• Configure via processing.chunking_strategy: semantic

Mistral Vision API for Images

• Extract and process images from PDFs using Mistral's Pixtral model
• No system dependencies needed (no Tesseract required!)
• Two modes: OCR (text extraction) or description (detailed analysis)
• Configure via pdf_processing.image_processing_mode

🔧 Improvements

• Automatic fallback to local processing when Mistral not configured
• Clear warnings when features require unavailable dependencies
• Smart configuration checks prevent crashes
• Graceful degradation for all cloud features

📝 Configuration

processing:
  chunking_strategy: semantic  # or recursive (default)

pdf_processing:
  extract_images: true
  image_processing_mode: ocr  # none, ocr, or description

📦 Dependencies

• Added langchain-experimental for semantic chunking
• Removed pytesseract dependency (now uses Mistral API)

🎯 Notes

Both new features gracefully degrade when dependencies are unavailable, ensuring backward compatibility with existing configurations.