Parsing Pipeline

The parsing pipeline is designed for reliability and repeatability. Every import is tracked as a session and can be audited or retried.

Pipeline stages

1. Upload

   • File stored in uploads/
   • SHA-256 hash computed for deduplication

2. Parser selection

   • ParserFactory chooses a bank-specific parser when available
   • Fallback parsers handle CSV/XLSX and generic AI/OCR paths

3. Import session

   • An import_session record is created
   • Metadata stores parser, source, and status

4. Transaction extraction

   • Parsed rows are normalized into transaction entities
   • Source mapping is stored for traceability

5. AI categorization

   • Optional AI pipeline (Gemini/OpenRouter)
   • Confidence thresholds and retry logic guard quality

6. Deduplication

   • Hash checks plus heuristics on date/amount/text
   • Conflicts flagged for manual review

Error handling

• Parsing errors are captured on the import session
• Failed stages can be retried without reuploading
• Logs include structured context for diagnostics

Related code

• backend/src/modules/parsing
• backend/src/modules/import
• backend/src/modules/classification

Next: Importing Statements