Merged
Conversation
- Add GeminiValidator for MedGemma output validation with retry logic - Add AnnotationSerializer with smart JSON truncation (4000 char limit) - Add AnnotationPipeline orchestrator (MedGemma → Gemini → Pydantic → DB) - Refactor GeminiAnnotationAgent to use bulletproof pipeline - Update /datasets/analyze endpoint for validated DB insertion - Add comprehensive tests and documentation Pipeline flow: MedGemma → Gemini Validation → Pydantic → DB Cost: ~$0.0005/image (validation only) Budget: $300 = 600k images 🤖 Generated with Claude Code Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
- Keep bulletproof pipeline implementation - Remove conflicting old code - Update .gitignore for __pycache__
- Enhanced DB schema with annotation_request staging table
- Added AgenticAnnotationRepo for two-tier operations
- Created ClinicalSummary Pydantic model with validation
- Implemented GeminiSummaryGenerator for polished summaries
- Built AgenticAnnotationPipeline orchestrator
- All tests passing (imports, models, repository)
Architecture:
MedGemma → Gemini Validator → Pydantic → annotation_request (raw)
↓
Gemini Summary
↓
annotation (clean)
Benefits:
- Full traceability (raw → summary)
- Pydantic validation at every step
- Clean summaries for frontend
- Pipeline statistics & debugging
- Cost: ~$0.0008/image
🤖 Generated with Claude Code
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
- annotation table now has 6 columns (added request_id) - Updated save_annotations to explicitly set request_id=NULL - Maintains backward compatibility with old API code
- Automated backup of existing database - Creates new database with enhanced schema - Verifies table and index creation - Shows helpful next steps Usage: ./migrate_database.sh
- Updated POST /datasets/load to create annotation_request placeholders - Updated POST /datasets/analyze to use AgenticAnnotationPipeline - Flow: load → annotation_request (pending) → analyze → update request → process to annotation - annotation table now only receives processed, refined data with Gemini summaries Breaking change: /datasets/analyze now requires /datasets/load first
Changes: - Added force_reanalyze parameter to PromptRequest schema - Updated POST /datasets/analyze to support re-analyzing processed images - Improved error messages (distinguish "no images" vs "already processed") - Added analyze_dataset() support for force_reanalyze in API client - Added "Analyze Dataset" button to UI with force re-analyze checkbox - Fixed 404/400 error when trying to analyze already-processed images Features: - Users can now re-analyze images with different prompts - Clear error messages with pipeline stats - UI form with customizable prompt and force re-analyze option
Critical fixes for concurrent database access: 1. Enabled WAL (Write-Ahead Logging) mode in both repositories - Allows multiple readers + one writer simultaneously - Prevents "database is locked" errors 2. Added connection timeout and autocommit mode - timeout=30.0: Wait up to 30 seconds for lock acquisition - isolation_level=None: Enable autocommit mode 3. Reused global agentic_repo instance across endpoints - Prevents creating multiple concurrent connections - Fixed: POST /datasets/load creating new AgenticAnnotationRepo - Fixed: POST /datasets/analyze creating new AgenticAnnotationRepo - Fixed: POST /chat creating new AgenticAnnotationRepo Changes: - DB/agentic_repository.py: WAL mode + timeout + autocommit - DB/repository.py: WAL mode + timeout + autocommit - src/api/main.py: Global agentic_repo reused across endpoints Root cause: Multiple database connections trying to write simultaneously in async FastAPI context with SQLite's limited concurrency support.
Critical performance fix: 1. Created global agentic_pipeline instance - Initialized once at startup - Reused across all analyze requests - Prevents MedGemma model from reloading every time 2. Fixed lazy loading logic in MedGemmaTool - Only set _model_loaded flag AFTER successful load - Prevents flag from being set on failed loads - Better error handling and logging Root cause: - AgenticAnnotationPipeline was created fresh on every POST /datasets/analyze - Each new pipeline created a new MedGemmaTool() instance - Each new MedGemmaTool() reset _model_loaded = False - Model was reloaded from disk on EVERY request (~10-20 seconds each time) After fix: - First request: Loads model (~10-20 seconds) - Subsequent requests: Reuses loaded model (~instant) Changes: - src/api/main.py: Global agentic_pipeline initialized at startup - src/tools/medgemma_tool.py: Fixed lazy loading flag logic
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.