Two Ways to Extract Invoice Data

1. Manual Processing

   For invoices with attached PDF/image:

   • Open a draft invoice with an attached PDF/image
   • Click "Process with AI" button
   • Form reloads with extracted data

2. OCA Invoice Import Wizard

   When using OCA's invoice import wizard:

   • Go to Accounting → Vendors → Import Vendor Bill
   • Upload PDF invoice
   • If no embedded XML found, LLM extraction is used as fallback
   • Invoice is created with extracted data

What Gets Extracted

The LLM extracts and populates:

• Vendor name and VAT number
• Invoice number and reference
• Invoice date and due date
• Currency
• Subtotal, tax, and total amounts
• Line items with descriptions, quantities, unit prices, and tax rates

OCA Invoice Import Integration

Follows the standard OCA pattern for extending account_invoice_import:

account_invoice_import_llm/
├── __manifest__.py                   # Dependencies and metadata
├── __init__.py                       # Pre-init hook for module rename
├── models/
│   ├── account_invoice_import_ocr.py # AbstractModel: OCR + LLM extraction
│   └── account_move.py               # "Process with AI" button
├── wizard/
│   └── account_invoice_import.py     # OCA fallback_parse_pdf_invoice()
├── data/
│   ├── llm_prompt_invoice_data.xml   # One-shot extraction prompt
│   └── llm_assistant_data.xml        # Assistant configuration
└── views/
    └── account_move_views.xml        # Button view extension

Processing Flows

Flow 1: Manual Trigger:

1. User clicks "Process with AI" button
   ↓
2. action_process_with_llm()
   - Validates invoice is draft
   - Finds PDF/image attachment
   ↓
3. extract_invoice_data()
   - Runs OCR on file bytes
   - Renders prompt with OCR text
   - Calls LLM for extraction
   ↓
4. _update_invoice_from_pivot()
   - Converts to OCA pivot format
   - Updates invoice fields
   ↓
5. Returns action to reload form view
   ↓
6. User sees populated invoice data

Flow 2: OCA Wizard Fallback:

1. User uploads PDF via OCA wizard
   ↓
2. Wizard calls parse_pdf_invoice()
   ↓
3. Checks for embedded XML (UBL, Factur-X)
   - If found: Uses XML parser
   - If not found: Calls fallback_parse_pdf_invoice()
   ↓
4. Our fallback_parse_pdf_invoice()
   - Extracts data via LLM
   - Converts to OCA's Invoice Pivot Format
   ↓
5. Wizard creates invoice from pivot format

Key Components

1. OCR Extraction AbstractModel

   File: models/account_invoice_import_ocr.py

   Central extraction logic shared by wizard and manual button:

   • Runs Mistral OCR on file bytes
   • Renders prompt with OCR text context
   • Calls LLM for structured extraction
   • Converts to OCA Invoice Pivot Format

2. One-Shot Extraction Prompt

   File: data/llm_prompt_invoice_data.xml

   Structured prompt that instructs LLM to extract invoice data in JSON format.

3. OCA Wizard Integration

   File: wizard/account_invoice_import.py

   Extends fallback_parse_pdf_invoice() to delegate to OCR AbstractModel.

Error Handling Strategy

• Wizard flow: Catches exceptions and returns res gracefully (doesn't break OCA flow)
• Manual button: Lets exceptions propagate as UserError with helpful messages
• Logging: Detailed error info in server logs for debugging

Pre-Init Hook

File: __init__.py

Handles module rename from llm_assistant_account_invoice → account_invoice_import_llm:

1. Updates ir_model_data records from old module name
2. Adopts orphaned records (prevents FK violations)
3. Prevents duplicate key errors on installation

Invoice Extraction Assistant

Location: Settings → LLM → Assistants → Invoice Data Extraction (Automatic)

The assistant is pre-configured with:

• Prompt: One-shot extraction template with detailed instructions
• Model: Any LLM (GPT-4, Claude, Gemini, etc.)
• OCR Tool: Mistral OCR for text extraction
• Dynamic Context: OCR text injected automatically

Customization

To customize extraction behavior:

1. Go to Settings → LLM → Prompts → Invoice Data Extraction (One-Shot)
2. Edit the instructions argument to add/remove fields
3. Update the JSON structure in default_values
4. Modify _convert_llm_data_to_pivot() if adding new fields

Issue: "No PDF or image attachment found"

Solution: Attach a PDF or image file to the invoice before clicking "Process with AI"

Issue: "Failed to extract data"

Possible causes:

• Poor quality scan/image
• Non-standard invoice format
• OCR couldn't extract text

Solution: Check server logs for detailed error. Try with a higher quality scan.

Issue: "Mistral provider not configured"

Solution: Install llm_mistral module and configure Mistral provider with API key

Issue: Duplicate records after module rename

Solution: The pre-init hook should handle this automatically. If issues persist, check ir_model_data for orphaned records.

Authors

• Apexive Solutions LLC

Maintainers

This module is maintained by Apexive Solutions LLC.