feat(fusion_accounting_ocr): pluggable OCR for vendor bills
Replaces Enterprise's account_invoice_extract with a Fusion-native pipeline:
Stage 1 (text extraction): Tesseract OCRs the bill attachment via
pytesseract + pdf2image. Pluggable OCRProvider adapter pattern allows
future Mindee / Google Document AI / Ollama-vision backends.
Stage 2 (field parsing): The fusion_accounting_ai LLMProvider reads the
raw OCR text and returns structured invoice fields (vendor, invoice
number, dates, amounts, line items) as JSON.
Draft invoice fields are auto-populated for empty-only fields (never
overwriting user-entered data). Vendor matching by name against
res.partner with supplier_rank > 0.
Adds:
- account.move.ocr_state (selection: not_requested/pending/processing/
done/failed/manual)
- account.move.ocr_raw_text, ocr_extracted_data (Json), ocr_backend,
ocr_confidence
- fusion.ocr.log (audit trail per OCR run)
- res.company.fusion_ocr_enabled / fusion_ocr_default_backend / auto_run
- /fusion/ocr/request_for_invoice JSON-RPC endpoint
Backend availability detected at runtime via OCRProvider.is_available()
classmethods. Tesseract 5.3.4 + pytesseract 0.3.13 + pdf2image 1.17.0
are installed in the container.
Tests: 13 (TesseractAdapter availability + image OCR; flow tests for
draft autofill, no-attachment guard, customer-invoice guard, ref-not-
overwritten; field parser empty/clean-json/markdown-fence/bad-JSON/
provider-exception). All pass on westin-v19 OrbStack VM.
Made-with: Cursor
125f48377a