How to build image search for a jewellery catalogue: build it yourself or buy it

What does jewellery image search have to do?

Jewellery image search has one job: a buyer sends a photo, and the closest designs from the seller's own catalogue come back, ranked, with the details a salesperson needs to reply. To do that, it has to:

• Match the photo to your own SKUs. Results come from your catalogue images, not the public web, so every match is one of your own designs.
• Rank the matches. The closest designs come at the top, with a match score, so the salesperson knows what to send first.
• Show the item details. Each match carries its SKU, weight and purity from your item master.
• Cope with real buyer photos. Buyers send WhatsApp screenshots, showroom pictures with reflections and crops from other companies' catalogues, rarely a clean catalogue shot.
• Work where enquiries arrive. Most enquiries arrive on WhatsApp, often as a forwarded photo.

That is a different job from a web search such as Google Lens, which finds what is published online, and from visual search on an online store, which serves shoppers rather than a sales desk.

What does it take to build jewellery image search yourself?

Building it yourself means putting eight parts together and keeping them running. Each part needs someone who owns it.

1. Clean catalogue photos, tied to the item master. Each design needs a clear photo, and each photo has to carry the right SKU, so a match can be turned into a weight, a purity and a stock line. Catalogue photos often sit across ERP tables, Google Drive folders and salespeople's phones, so collecting them, removing duplicates and fixing wrong codes is work in itself.
2. An image-matching model. The model turns each photo into a vector, a numeric fingerprint of its shape, motif and finish, so similar designs sit close together. You can host an open-source vision model yourself, which your team may need to adapt to jewellery, or call a cloud vision API, which runs the model for you. Check how it handles fine detail, because many chains look alike up close. If you choose a cloud service, check that it is still actively developed.
3. A vector search index. The fingerprints go into a vector database or search index that returns the nearest designs to a new photo and filters them on fields such as category, weight band and purity. It has to be updated whenever designs are added, changed or discontinued.
4. A search screen. The sales desk needs somewhere to drop in a photo and see the ranked matches with SKU, weight and purity: a web page, a screen inside the ERP or a dealer portal. Someone has to build it, host it and decide who can log in.
5. A WhatsApp Business connection, if buyers send photos there. Taking photos in and sending matches back on WhatsApp means connecting a business number to the WhatsApp Business Platform and following WhatsApp's business messaging rules.
6. A sync with the ERP item master. New designs, weight changes and discontinued designs have to reach the index, or search offers designs you no longer have. Read the item master over whatever the ERP exposes (an API, database views or scheduled files), and give search read access, never write access, so it can't change a record. How to add image search without replacing your ERP covers that connection.
7. Testing on real buyer photos. Catalogue shots are the easy case. Test on what buyers actually send: WhatsApp screenshots, steep side angles, tight close-ups of a clasp or a few links of a chain, glare and coloured shop lighting, and crops with price tags or other pieces in the frame. Keep a set of real enquiries, each with the design the buyer meant, and check every change against it. What makes a buyer's photo easy or hard to match lists the usual problems.
8. Ongoing upkeep. After launch, someone has to own it: adding new collections to the index, checking wrong matches the sales desk reports, updating the model and the servers, and keeping the WhatsApp connection running.

This guide gives no figures for cost, build time or match quality, because each depends on your catalogue, your photos and your team. A test on a sample of your own designs tells you more than a general figure.

When does building it yourself make sense?

Building fits when:

• You have an in-house engineering team that can build, host and maintain a search system. You control every part, and your team runs it and keeps the index in sync as designs change.
• Your needs are unusual, such as matching rules specific to your designs, a system no vendor connects to, or a requirement that no product meets once you have tested them.
• Image search is part of a product you sell, and owning the model and the index matters more to you than buying them.

When does buying make sense?

Buying fits when:

• Your team sells jewellery, not software. Nobody in-house is set up to build, host and maintain a search system.
• Enquiries arrive on WhatsApp, and you want search where buyers already send photos without building and running the WhatsApp connection yourself.
• You want to see it work before you commit. A vendor can show the search on a sample of your own designs before you decide.
• You would rather not run the system. In a hosted service, the model, the search index and the servers are the vendor's to run.
• You build software for jewellers, such as an ERP or POS. The QuickLens partner programme lets you offer its search under your own brand.

Whichever you choose, test it on a sample of your own designs and on real buyer photos before you decide.

Read: how to choose a jewellery catalogue search tool

How does QuickLens do it?

QuickLens is the buy option: AI visual search for B2B catalogues, from Mumbai, India, built first for jewellery manufacturers, wholesalers and retail chains.

• Your own catalogue. A buyer's photo returns the closest designs from your own catalogue, ranked with a match score. Nothing comes from the public web.
• Nobody has to tag the catalogue first. A re-shoot is usually not needed. Search is built to run on catalogue photos as they are, and where photo quality genuinely limits matching, we tell you which parts of the catalogue are affected.
• WhatsApp, web app and REST API. The salesperson forwards the buyer's photo on WhatsApp, and the closest designs come back with their SKU, weight and purity. The same search runs in the QuickLens web app, and in your own website, app, ERP screen or dealer portal through the REST API.
• Alongside your ERP. QuickLens reads your item master over whatever your ERP exposes: an API, database views or scheduled files. Search never writes to the ERP. The records written back are orders and quotations a person has approved.
• Used for one thing. Your designs and photos are used to search your own catalogue. They are never used to train QuickLens's AI models.
• Hosted on AWS in Mumbai, or on your own servers. QuickLens can run fully on-premise, in your own environment, instead of the hosted setup on AWS in the Mumbai region.
• Tested on your designs first. The free 15-minute AI audit includes a live search on a sample of your designs, and we expand once your team trusts the results.

QuickLens plans start at $675 a year plus taxes, with no charge per search. There are three packages, Core, Teams and Commerce, and the yearly fee depends on the package and the size of your catalogue. A one-time fee of around $560 applies if you need ERP integration or custom requirements, and not otherwise. We give an exact quote after a free 15-minute AI audit.

Read: what decides your price and what every plan includes

Building jewellery image search: common questions

Get started

Try it on a sample of your own designs

Book a 15-minute AI audit and we will map your catalogue, show a live search on a sample of your designs, and give you the honest version of what will and will not work.