Appearance
AskLilySearch
Home

TECHNOLOGY

Multimodal fashion intelligence: technical architecture.

How product inputs, expert-reviewed fashion knowledge and connected relationships support enrichment, retrieval, ranking and AI Outfitting.

On this page

Technology overview

Selectika’s proprietary multimodal fashion intelligence interprets product imagery together with titles, descriptions and catalogue context. “Multimodal” means that different kinds of input contribute to a unified understanding of the product. The resulting intelligence supports product enrichment, semantic retrieval, matching, ranking and AI Outfitting.

For a retailer, this links the information held in a catalogue with the decisions made in search, styling and recommendations. For technical teams, the published APIs provide routes to submit products, retrieve enriched records and request product relationships.

Architecture at a glance

CONCEPTUAL CAPABILITY MAP

Product inputs

Images · titles · descriptions · existing attributes · catalogue context

Proprietary multimodal fashion intelligence

Unified product understanding, informed by fashion expertise

Approximately 100 attributes
Category-dependent coverage
Rich synonym vocabularies
Related language, shared meaning

Fashion Decision Graph

Compatibility, alternatives, occasion and contextual relationships

Vector retrieval & ranking

Product embeddings, attributes, reviewed relationships and retailer controls

Retail experiences

Enriched catalogues · discovery · AI Outfitting · AI Stylist Chat · shopping agents

This map groups documented capabilities and relationships. It is not an infrastructure diagram or a prescribed processing sequence.

Inputs and the information they provide

Images convey visible product details. Titles, descriptions and existing attributes add catalogue context. Selectika’s model interprets these together rather than treating imagery and text as disconnected signals.

The Product Enrichment submission API requires an item ID, image URLs, gender and category. Optional fields include title, description, brand, colour, price and stock. Shopper intent and contextual signals apply to downstream experiences and are separate from those minimum enrichment fields.

Keep source-backed facts separate from visual or contextual interpretation. Product imagery can support shape, colour and visible-detail analysis; composition and performance claims need reliable catalogue or supplier information. Occasion and styling context should be evaluated against the retailer’s vocabulary and intended use.

Approximately 100 attributes, with a shared vocabulary

Attribute coverage depends on product category. Published examples include silhouette, fit, neckline, sleeves, waist, length, construction, fabric, texture, pattern, colour, pockets, buttons, distressing, technical performance, occasion, sub-occasion, style and trends.

Rich synonym vocabularies help relate different buyer phrases and retailer terms to the same underlying meaning. The company notes that individual tag values can include around 20 synonyms or related phrases; this is not a fixed requirement for every value. The fashion taxonomy provides a practical reference for the published attribute families.

Fashion knowledge reviewed by experts

Over several years, millions of fashion items have been tagged and reviewed by human stylists and fashion experts. This reviewed information captures product meaning as well as relationships between pieces. Selectika also reports processing hundreds of thousands of complete looks.

Within those looks, bounding boxes identify where individual products appear in the image. Catalogue matching connects each identified product to the exact retailer item. Fashion experts review or correct those matches and relationships. Selectika describes this relationship data as being used to train and refine its AI Outfitting intelligence.

This account explains the development of the company’s intelligence. It does not mean that every incoming customer product is manually reviewed. Review requirements for a particular catalogue should be agreed during implementation.

The Fashion Decision Graph

A catalogue record describes one item. The Fashion Decision Graph adds relationships: which products complement one another, which can serve as alternatives, and how style, occasion and context affect a choice.

It supports distinct questions: “What works with this?” calls for compatibility; “What could I choose instead?” calls for an alternative; “What is appropriate for this occasion?” adds intent and context. These relationships connect product understanding with the decisions required by a retailer’s shopping experience.

Vector retrieval and retailer-controlled ranking

Selectika indexes product embeddings, structured attributes and expert-reviewed relationships in a vector database. An embedding is a numerical representation used to help compare meaning. Semantic retrieval can find relevant products even when a shopper’s wording differs from a catalogue title.

The published architecture distinguishes retrieval from ranking. Retrieval narrows the catalogue to candidate items. Ranking then combines product meaning with shopper intent, context, compatibility, price, stock, retailer rules and performance signals to decide which candidates should appear first.

The company has not specified a public database vendor, embedding dimension or universal latency commitment. Technical evaluation should focus on the documented interfaces and the requirements of the intended deployment.

From intelligence to a retail workflow

Product Enrichment supplies structured data for catalogues, PIM systems and merchant feeds. Shop Similar presents alternatives. AI Outfitting uses reviewed relationships and complete looks to recommend complementary sets. AI Stylist Chat connects a conversation with catalogue understanding.

Published APIs and configured webhooks support activation. Catalogue ingestion, field mapping, available shopper signals, retailer controls and the presentation layer need to be agreed for the selected experience. Explore the developer documentation

MCP access and enriched product feeds

Selectika’s company-confirmed capabilities include an MCP layer that makes enriched catalogue information, similar products and complementary items accessible to connected AI applications. Selectika has also created a feed for Google Merchant Center; the same enriched feed serves as a product-information source for OpenAI, Claude and Grok, according to the company.

MCP access and feeds are separate delivery paths. Client setup, required fields, permissions, update arrangements and the receiving platform’s requirements need to be confirmed for the implementation. No endpoint, authentication method or refresh schedule is specified here. Read the catalogue enrichment and AI access explanation

INSIDE THE FASHION INTELLIGENCE

From product.
To possibility.

Illustrative walkthrough · A simplified explanation of documented concepts, not a live processing screen.

01 / INPUTS

Start with what the catalogue knows.

Selectika’s proprietary multimodal fashion intelligence interprets product imagery together with titles, descriptions and catalogue context. These inputs inform one connected representation of the product.

For a blazer, the image shows shape and construction while reliable catalogue information supplies details such as composition.

02 / ENRICH

Give details a consistent meaning.

Product Enrichment identifies and normalises fashion attributes against a controlled vocabulary. Coverage is category-specific, rather than the same fixed list for every SKU.

A supplier’s “off-white jacket” can be described through a product type, colour family, sleeve length and lapel detail, ready for agreed catalogue fields.

03 / RELATIONSHIPS

Learn how products work together.

Selectika describes a training-data workflow in which items in complete looks are identified, matched to catalogue products and reviewed by fashion experts. This evidence helps train and refine AI Outfitting.

Expert-reviewed looks inform compatibility and outfit logic. This does not mean every newly submitted product is manually reviewed.

04 / RETRIEVE

Find candidates. Then consider context.

A vector database indexes product representations, attributes and relationships. Retrieval finds relevant candidates; ranking then considers meaning, intent, context, compatibility, price, stock and retailer rules.

For “something tailored for dinner,” meaning can help retrieve relevant pieces even when the exact phrase is absent from their titles.

05 / ACTIVATE

Put that knowledge to work.

Structured product data and relationships support catalogue operations, discovery and recommendation experiences. APIs, webhooks and agreed field mappings connect those outputs with the retailer’s implementation.

Catalogue teams can use richer fields, search teams can use consistent product meaning, and merchandising teams can surface complementary products.

Read the technical architecture

Search Selectika

Explore a topic

Meet Lily

Your personal Selectika assistant.