E-Facture Bank Reconciliation Pipeline

Problem

In multi-client accounting, invoices may arrive as receipt photos, PDFs, or semi-structured e-invoices, while bank data may be CAMT.053, OFX, CSV, or PDF statements. Manual reconciliation often misses small refunds, fees, and paid-but-unbilled edge cases. This skill turns the flow of invoices + bank statements → reconciliation results into a repeatable pipeline, focusing on structured ingestion, LLM extraction of unstructured evidence, and producing company.json and rapprochement.json.

How It Works and Boundaries

The entrypoint is scripts/main.py, with four steps:
- A. Invoice ingestion: prefer Factur-X, pure UBL/CII; receipt photos or XML-less PDFs go through LLM text/vision extraction with a confidence score.
- B. Bank ingestion: normalize CAMT.053, OFX, and CSV into signed transactions; PDF statements are also extracted by LLM.
- C. Reconciliation: delegated to the shared matching engine, using amount plus reference, counterparty, or date; misses go to unmatched_bank_lines, while salaries, taxes, and bank fees go to excluded_bank_lines.
- D. Output: write company.json and rapprochement.json, then validate rappr + unmatched + excluded == bank_transactions_count.

Execution is two-pass. Pass one runs the engine and stops with a worklist plus sidecar JSON paths for unstructured files. Pass two reads each file, extracts the normalized JSON, writes the sidecar, then reruns until the worklist is empty. Outputs default to a clients-test/ sandbox; use --real only after validation. It fits existing accounting backends that need better ingestion, extraction, and reconciliation output, but it does not replace backend work, email, front-end display, or manual editing of final JSON.