- Simplify STT_SERVICE config to local/MODEL_SIZE format
- Remove separate STT routes, integrate with document upload
- Add local STT support to audio file processing pipeline
- Remove React component, use existing upload interface
- Support both local Faster-Whisper and external STT services
- Tested with real speech: 99% accuracy, 2.87s processing
- RBAC soon??
- Updated various services and routes to handle search space-specific LLM preferences.
- Modified frontend components to pass search space ID for LLM configuration management.
- Removed onboarding page and settings page as part of the refactor.
- Deleted the document_processing module and its associated docling_service.
- Updated imports in documents_routes.py and background_tasks.py to reflect the new service structure.
- Ensured compatibility with the task logging system by adjusting type hints for log entries.
- Integrated Docling ETL service with new task logging system
- Maintained consistent logging pattern across all ETL services
- Added progress and success/failure logging for Docling processing
- Added DOCLING as third ETL_SERVICE option (alongside UNSTRUCTURED/LLAMACLOUD)
- Implemented add_received_file_document_using_docling function
- Added Docling processing logic in documents_routes.py
- Enhanced chunking with configurable overlap support
- Added comprehensive document processing service
- Supports both CPU and GPU processing with user selection
Addresses #161 - Add Docling Support as an ETL_SERVICE
Follows same pattern as LlamaCloud integration (PR #123)
\
- Earlier for each chunk, whole message (with all annotations included)
were streamed. Leading to extremely large data length.
- Fixed to only stream new chunk.
- Updated ANSWER part to be streamed as message content (following
Vercel's Stream Protocol)\
- Fixed yield typo
