> ## Documentation Index
> Fetch the complete documentation index at: https://docs.civic.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Mistral OCR

> Extract text and images from PDFs and images using Mistral's Pixtral OCR API. Supports URL-based and base64-encoded documents.

## Overview

The Mistral OCR server uses Mistral's Pixtral model to extract text from images and PDF documents. It supports both URL-based documents and base64-encoded content, making it ideal for processing invoices, receipts, contracts, and other scanned documents.

## How to Add Mistral OCR

<Steps>
  <Step title="Get Mistral API Key">
    1. Go to [https://console.mistral.ai/](https://console.mistral.ai/)
    2. Sign up or log in to your Mistral account
    3. Navigate to **API Keys** in your account settings
    4. Create a new API key or copy an existing one
  </Step>

  <Step title="Connect to Civic">
    1. Add the Mistral OCR server to your Civic environment through the server directory
    2. Enter your Mistral API key when prompted
  </Step>

  <Step title="Test Connection">
    Start with a simple command like "Extract text from this PDF: \[URL]" to verify the connection works properly.
  </Step>
</Steps>

## What You Can Do

<CardGroup cols={2}>
  <Card title="PDF Text Extraction" icon="file-pdf">
    Extract text content from PDF documents via URL or base64 encoding
  </Card>

  <Card title="Image OCR" icon="image">
    Extract text from images including PNG, JPEG, WebP, and AVIF formats
  </Card>

  <Card title="Invoice Processing" icon="file-invoice">
    Process invoices, receipts, and bills to extract amounts, dates, and vendor info
  </Card>

  <Card title="Document Analysis" icon="magnifying-glass">
    Extract and analyze text from contracts, forms, and scanned documents
  </Card>
</CardGroup>

## Available Tools (2)

### OCR from URL

<AccordionGroup>
  <Accordion title="ocr_url">
    Extract text from an image or PDF document via URL

    * **Input**:
      * `url` (required) - Public URL of the image or PDF document
      * `type` (required) - Document type: `image` or `pdf`
      * `includeImages` (optional) - Include base64-encoded images in response (default: false)
    * **Use Cases**: Process publicly accessible documents, extract text from hosted files

    <Warning>
      If the URL does not have a clear file extension (.pdf, .png, .jpg, etc.), you must specify the document type explicitly. Do not guess - ask the user.
    </Warning>
  </Accordion>
</AccordionGroup>

### OCR from Base64

<AccordionGroup>
  <Accordion title="ocr_base64">
    Extract text from a base64-encoded image or PDF document

    * **Input**:
      * `data` (required) - Base64-encoded image or PDF data (without data URI prefix)
      * `mimeType` (required) - MIME type of the document
      * `includeImages` (optional) - Include base64-encoded images in response (default: false)
    * **Supported MIME Types**:
      * `image/png`
      * `image/jpeg`
      * `image/webp`
      * `image/avif`
      * `application/pdf`
    * **Use Cases**: Process documents from file uploads, extract text from embedded content
  </Accordion>
</AccordionGroup>

## Use Cases

<AccordionGroup>
  <Accordion title="Invoice & Receipt Processing" icon="file-invoice">
    * **Extract Totals**: "What is the total amount on this invoice?"
    * **Get Line Items**: "Extract all the line items from this receipt"
    * **Vendor Info**: "Pull out the vendor name, date, and amount from this bill"
    * **Batch Processing**: Process multiple invoices to extract payment details
  </Accordion>

  <Accordion title="Document Digitization" icon="file-lines">
    * **PDF Conversion**: "Extract the text from this scanned PDF"
    * **Form Processing**: "What are the values filled in on this form?"
    * **Contract Analysis**: "Extract the key dates and parties from this contract"
    * **Archive Search**: Enable searching through scanned document archives
  </Accordion>

  <Accordion title="Data Extraction" icon="table">
    * **Contact Info**: "Find all phone numbers and email addresses in this document"
    * **Dates & Deadlines**: "What are the dates mentioned in this document?"
    * **Financial Data**: "Extract all monetary amounts from this statement"
    * **Structured Output**: Convert unstructured documents into usable data
  </Accordion>

  <Accordion title="Image Text Extraction" icon="image">
    * **Screenshots**: "What does this screenshot say?"
    * **Photos of Documents**: "Read the text from this photo of the receipt"
    * **Signage & Labels**: "What text is visible in this image?"
    * **Handwritten Notes**: Extract text from photos of handwritten content
  </Accordion>
</AccordionGroup>

## Sample Prompts

### Basic OCR

* "Extract the text from this PDF: [https://example.com/invoice.pdf](https://example.com/invoice.pdf)"
* "What does this image say? [https://example.com/receipt.png](https://example.com/receipt.png)"
* "OCR this document and give me the contents"

### Invoice Processing

* "Extract all the line items and totals from this invoice"
* "What is the invoice number, date, and total amount?"
* "Pull out the vendor details and payment terms"

### Document Analysis

* "Extract the text and summarize the key terms in this contract"
* "What are all the dates mentioned in this document?"
* "Find all contact information (emails, phone numbers) in this PDF"

### With Embedded Images

* "Extract the text and include any images from this PDF"
* "OCR this document and return the embedded charts as images"

## Security Guardrails

The Mistral OCR server has built-in security constraints:

### PII Redaction

<Note>
  Personally identifiable information (SSN, credit cards, phone numbers, emails) detected in OCR output is automatically redacted to prevent data leakage.
</Note>

### Input Validation

| Constraint          | Description                                                                                              |
| ------------------- | -------------------------------------------------------------------------------------------------------- |
| Base64 Size Limit   | Blocks payloads over \~10MB to prevent memory exhaustion                                                 |
| Block Internal URLs | Prevents SSRF attacks by blocking localhost and private IP ranges (10.x.x.x, 192.168.x.x, 172.16-31.x.x) |

## Known Limitations

### URL Requirements

* URLs must be publicly accessible
* Internal/private network URLs are blocked for security
* URLs without clear file extensions require explicit `type` parameter

### File Size

* Base64-encoded documents are limited to \~10MB
* For larger documents, use the `ocr_url` tool with a public URL

### Document Types

* Only supports images (PNG, JPEG, WebP, AVIF) and PDFs
* Scanned documents with poor quality may have reduced accuracy
* Handwritten text recognition may be limited

### PII Handling

* PII is automatically redacted from responses
* If you need unredacted data, contact your administrator to adjust guardrails

## Documentation

* [Mistral AI Documentation](https://docs.mistral.ai/)
* [Mistral API Reference](https://docs.mistral.ai/api/)
* [Pixtral Vision Model](https://docs.mistral.ai/models/model-cards/pixtral-large-24-11)

***

## Guardrails

In addition to the [14 universal guardrails](/civic/concepts/guardrails), this server has **4 server-specific guardrails** across 2 operations.

<Note>
  Covers both security (SSRF protection, payload size limits) and privacy (PII redaction from extracted text).
</Note>

| Guardrail                                   | Operation    | Timing   | Description                                                     |
| ------------------------------------------- | ------------ | -------- | --------------------------------------------------------------- |
| OCR Base64 - Maximum Payload Size           | `ocr_base64` | Request  | Blocks base64 payloads over \~10MB to prevent memory exhaustion |
| OCR Base64 - Redact PII from Extracted Text | `ocr_base64` | Response | Redacts SSN, credit cards, phones, emails from OCR output       |
| OCR URL - Block Internal Network URLs       | `ocr_url`    | Request  | Blocks SSRF attacks — prevents localhost/private IP URLs        |
| OCR URL - Redact PII from Extracted Text    | `ocr_url`    | Response | Redacts PII from URL-based OCR output                           |

<Tip>
  Configure guardrails via the Civic UI or ask the Configurator Agent: "Add guardrails to my Mistral OCR server."
</Tip>