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

# Image search

> Let customers attach photos in the chat so the AI can identify products, answer questions about a scene, and read receipts.

Customers can attach photos to any chat message. The AI reads the image alongside the text, so a shopper can send a picture of a product and ask "do you have this?" instead of guessing at search keywords.

Image search is on by default for every widget. There is nothing to enable or configure.

## What customers can do with it

* **Find a specific product.** A photo of an item they own, or spotted elsewhere, matched against your catalog.
* **Ask about a whole scene.** A picture of a room or an outfit, answered with several products that work together.
* **Get recommendations.** "What goes with this?" against a photo, answered with compatible items from your catalog.
* **Support with proof.** A photo of a damaged item, or of a receipt or invoice, so the AI can see the problem instead of asking the customer to describe it.

The AI treats an image as part of the conversation, so follow-up questions like "do you have it in blue?" keep working without re-attaching the photo.

## Attaching images

In the chat widget, customers add images three ways:

* the attachment button in the message box
* drag and drop onto the chat
* paste from the clipboard

On phones, the attachment button opens the camera or photo library, so a shopper can photograph something and search for it without leaving the chat.

## Limits

| Limit              | Value                           |
| ------------------ | ------------------------------- |
| Images per message | 3                               |
| Maximum file size  | 10 MB per image                 |
| Accepted formats   | JPEG, PNG, GIF, WebP, SVG, HEIC |

Files that are empty, too large, or of an unsupported type are rejected in the widget before any upload starts.

<Note>HEIC is the default camera format on iPhones, so photos taken in the chat upload without conversion.</Note>

## How images are processed

1. The widget uploads the file through Molin's upload endpoint.
2. Molin stores it and records it against the conversation.
3. The AI receives a resized copy of the image, at most 2048×2048.
4. The widget and your [inbox](/home/inbox/inbox) display a smaller copy, at most 800×800.

Both copies are generated on demand from the original upload and cached at the edge, so a customer on a slow connection doesn’t download a full-resolution photo to see their own attachment.

Support agents can attach images too, when replying from the inbox. Those follow the same path and the same limits.

## Content Security Policy

If your shop sends a [Content Security Policy](/home/general/content-security-policy), image search needs no extra origins. Uploads and image display both go through `widget.molin.ai`, which the documented `connect-src` and `img-src` entries already allow.

## Data retention

Images belong to the conversation they were sent in and follow the same lifecycle:

* Attachments a customer selects but never sends are deleted when they start a new chat.
* Sent images stay with the conversation history so you can see what the customer sent when you review it in the inbox.
* When a conversation is erased, either through data retention or a GDPR erasure request, its images are deleted from storage and purged from the cache.

## Frequently asked questions

<AccordionGroup>
  <Accordion title="Does it work on Facebook Messenger and Instagram?">
    Image search works in the web chat widget and in the inbox. Photos sent through [Messenger](/home/channels/facebook-messenger) and [Instagram](/home/channels/instagram) aren’t read by the AI
    today. Please give our team a nudge and we will prioritize it for a future update.
  </Accordion>

  <Accordion title="What does the AI match photos against?">
    Your product catalog, as synced from your platform or [product feed](/home/general/product-feeds). It recommends only products you actually sell, the same as any other answer.
  </Accordion>

  <Accordion title="Can customers send several photos at once?">
    Yes, up to 3 per message. The AI considers them together, so a shopper can send a few references and get one answer covering all of them.
  </Accordion>

  <Accordion title="Can the AI read text in an image?">
    Yes. This is what makes receipts, invoices, and order confirmations useful in a support conversation. Accuracy depends on how legible the photo is.
  </Accordion>
</AccordionGroup>