The Problem

Without OCR:

• Images are just binary blobs in your knowledge base
• Handwritten notes can't be searched
• Scanned documents are dead weight
• Receipts and invoices are unusable
• Knowledge stays locked in images

The Solution

With Mistral OCR:

• AI extracts text from any image
• Handwriting becomes searchable
• Scanned docs fully indexed
• Receipt data automatically parsed
• Everything is findable

Mistral Vision OCR

• State-of-the-art accuracy: Powered by Mistral's multimodal vision models
• Handwriting recognition: Extracts text from handwritten notes and forms
• Multi-format support: Images (PNG, JPG, WEBP), PDFs, scanned documents
• Automatic extraction: No manual data entry required

OCR Models

Three Mistral OCR models available:

• mistral-ocr-latest: Latest OCR model (recommended)
• mistral-ocr-2505: May 2025 version
• mistral-ocr-2503: March 2025 version

Mistral OCR Parser

• Seamless integration with llm_knowledge processing pipeline
• Automatic text extraction from image attachments
• Preserves original images while extracting text content
• Works with existing chunking and embedding systems

Step 1: View OCR Models

The module comes pre-configured with Mistral's OCR models. View them under LLM → Configuration → Models, filtered by "ocr".

Available models:

• mistral-ocr-latest - Latest OCR model (recommended)
• mistral-ocr-2505 - May 2025 version
• mistral-ocr-2503 - March 2025 version

Step 2: Configure Parser

When creating or editing a knowledge resource:

1. Select "Mistral OCR Parser" from the Parser dropdown
2. Choose your preferred OCR model (mistral-ocr-latest recommended)
3. Upload images as attachments
4. Click "Process Resources"

Parser settings:

• Parser: Mistral OCR Parser
• Provider: Mistral AI (auto-selected)
• OCR Model: mistral-ocr-latest or specific version
• Supported formats: PNG, JPG, WEBP, PDF

Step 3: Process and Search

Click "Process Resources" to extract text from your images. The extracted text becomes searchable through your AI assistant.

Handwritten Grocery List

Input: Photo of handwritten grocery list

Output: Extracted text

- potatoes
- peas & carrots
- pastina
- garbage bags
- dog treats
- aluminum foil
- almond milk
- creamer - vanilla
- eggs (2)
- crushed tomatoes
- hot sauce
- paper towels?

Result: Fully searchable in knowledge base. Ask "What items are on the grocery list?" and AI finds and lists all items.

Expense Management

Goal: Track business expenses from receipt photos

Setup:

• Upload receipt photos to knowledge collection
• Use Mistral OCR Parser
• Process resources

Result: Extract vendor, amount, date, and items from receipts. Search "Find all Starbucks receipts from last month" → AI finds all matching receipts and totals.

Meeting Notes Archive

Goal: Make handwritten meeting notes searchable

Setup:

• Scan or photograph handwritten meeting notes
• Upload to knowledge base
• Process with Mistral OCR

Result: Every decision, action item, and idea becomes searchable. Ask "What did we decide about the Q4 budget?" → AI cites exact meeting notes.

Product Label Extraction

Goal: Index product information from label photos

Setup:

• Photograph product labels
• Add to product knowledge collection
• Process with Mistral OCR

Result: Extract ingredients, nutritional info, warnings, and instructions. AI can answer product questions using label data.

Processing Pipeline

When you process an image resource with Mistral OCR:

1. Upload: Attach image to llm.resource
2. Parse: Mistral OCR Parser sends image to Mistral AI vision model
3. Extract: Vision model analyzes image and extracts all text
4. Save: Extracted text saved to resource's content field
5. Chunk: Text chunked using collection's chunker settings
6. Embed: Chunks embedded and stored in vector database
7. Search: AI can now search and cite this content in responses

Supported Image Types

• Handwritten text: Notes, forms, letters
• Printed text: Documents, books, manuals
• Receipts: Business expenses, invoices
• Screenshots: Error messages, UI text
• Product labels: Ingredients, instructions
• Whiteboards: Brainstorming sessions, diagrams
• Forms: Filled-out applications, surveys
• Scanned documents: PDFs, legacy files

Mistral OCR Models

The available OCR models are fetched from Mistral AI when you configure the provider:

Note: You must set up the Mistral provider via llm_mistral module and click "Fetch Models" to download the available OCR models from Mistral AI.

Parser Registration

The Mistral OCR Parser is registered in models/mistral_resource_parser.py on the llm.resource model:

@api.model
def _get_available_parsers(self):
    parsers = super()._get_available_parsers()
    parsers.extend([
        ("mistral_ocr", "Mistral OCR Parser"),
    ])
    return parsers

The parser method parse_mistral_ocr() processes images:

def parse_mistral_ocr(self, record, field):
    mimetype = field["mimetype"]
    if not self.llm_model_id or not self.llm_provider_id:
        raise ValueError("Please select a model and provider.")
    value = field["rawcontent"]
    ocr_response = self.llm_provider_id.process_ocr(
        self.llm_model_id.name, value, mimetype
    )
    final_content = self._format_mistral_ocr_text(ocr_response, record.id)
    self.content = final_content
    return True

Fields Added to llm.resource

This module extends llm.resource with:

• llm_model_id: Many2one to OCR model (domain: model_use = 'ocr')
• llm_provider_id: Many2one to Mistral provider (domain: service = 'mistral')

OCR not extracting text

1. Verify image quality is sufficient (not too blurry)
2. Check Mistral API credentials are configured
3. Review system logs for API errors
4. Try different OCR model version

Handwriting not recognized

1. Ensure handwriting is legible
2. Use high-resolution images
3. Try mistral-ocr-latest (best for handwriting)
4. Avoid low-light or skewed photos

Wrong text extracted

1. Check image orientation (rotate if needed)
2. Verify image is not corrupted
3. Ensure sufficient contrast between text and background
4. Try cropping to focus on text area