Visual search

How do I find the product when the buyer only sends a photo?

Eva compares the image against your catalog images, returns candidates with confidence levels and verifies the record before stating price or stock.

Plenty of buyers do not know what the thing they want is called. They send a photo of the part, a screenshot of a post or a collage of three models and write «do you have this one?». Whoever is on the chat searches from memory, and if they are new, the search ends in «let me check» and dies there.

Eva compares that image against your catalog images and returns candidates with their confidence level and their evidence. The verified product record stays the source of truth for price, stock and variants. Here is how it works and what it needs to work.

Whatever the customer sends, Eva understands

Voice note · 0:23understood
Photo of a productidentified
Shared locationnearest branch
Incoming callanswered
Video · 0:41understood
Order_march.pdfread
Voice note · 0:23understood
Photo of a productidentified
Shared locationnearest branch
Incoming callanswered
Video · 0:41understood
Order_march.pdfread

Before and after

Today

The photo arrives and somebody either recognizes it or does not. If the product has ten similar variants, the answer depends on who is on shift, and the buyer gets the wrong model or a promise to look into it later.

With Eva

The image is matched against the catalog and candidates come back ranked with their confidence level. Before naming a price or stating availability, Eva pulls the verified product record, and when verification is incomplete it says so instead of asserting more than it knows.

How it works

  1. 1. The buyer sends the image

    A photo of their own, a screenshot or a collage with several options all work. Eva processes JPEG, PNG, GIF and WebP images on the channels where it receives files, and can use the most recent ones in the conversation to get ahead of the search.

  2. 2. Eva matches it against your catalog

    The comparison runs against your own product images, not a general image bank. What comes back are candidates with a confidence level and evidence for the resemblance, not an assertion that this is the product.

  3. 3. It verifies before asserting

    Before giving price, currency, availability or variant, Eva pulls the verified product record and uses only the fields you marked visible. If that verification is incomplete, the answer is capped rather than completed with a guess.

Live

See it working

Whatever the customer sends, Eva understands

Voice note · 0:23understood
Photo of a productidentified
Shared locationnearest branch
Incoming callanswered
Video · 0:41understood
Order_march.pdfread
Voice note · 0:23understood
Photo of a productidentified
Shared locationnearest branch
Incoming callanswered
Video · 0:41understood
Order_march.pdfread

Eva understands text, audio, photos and documents.

The answer is built from your data

Catalog

$30,375 · 14 in stock

Knowledge

Free delivery in León

Calendar

Thursday 10:30 open

We do: 1.5-ton inverter mini-split, $30,375 each with 14 in stock. Delivery in León is included and I can book the install Thursday at 10:30.

Verified against your data

The answer is built from your data, never invented.

Collects and invoices on its own

Payment confirmed

Invoice_A241.pdf

CFDI 4.0 invoice issued and sent

Eva collects payment and invoices inside the conversation.

What Eva needs

  • A catalog loaded with product images and its visual index ready; without an image, a product cannot come back as a candidate.
  • The product search and product detail tools enabled for that agent, with the fields you want shown marked as visible.
  • A connected channel: WhatsApp Business and, if you want them, Instagram, Messenger, calls, email, your website widget or the API.
  • Written rules for when the resemblance is not enough: ask for a dimension, ask for another photo, or hand the conversation to a person.

Good to know

  • Results are candidates, not final facts. Price, currency, availability and variants are always confirmed against the verified product record.
  • Incomplete verification caps the result at medium confidence, and Eva answers within that cap instead of asserting more.
  • Failures in detection, embeddings, reranking or reading text inside the image degrade results, and what Eva may assert stays bounded by the confidence levels.
  • Proactive preloading uses at most the three most recent customer images, and requires the search and detail tools plus a ready visual index.
  • The comparison runs against your catalog images. Eva does not identify objects in the world or other brands' products that are not loaded in it.
  • Supported images are JPEG, PNG, GIF and WebP with a documented ten-megabyte maximum, and not every model supports every file type.
  • Search returns up to twenty-five candidates and the detail lookup only returns fields marked visible: what is not in the catalog, Eva does not answer.
  • API-connected sources sync every twenty-four hours and file sources refresh when you upload them again; minute-by-minute stock requires a live lookup exposed from your own system.
+4×

online sales in six months

Serviacero Comercial, a steel distributor, multiplied its online sales more than fourfold in six months with Eva answering and quoting immediately.

Read the full story

Put Eva to work.

Send us a photo a customer sent you this week and we will show you which candidates Eva would have returned from your catalog.

Message us on WhatsApp

Updated

Message us on WhatsApp