Catalog-Aware Fashion Chatbots: Using Knowledge Graphs as Context Memory

Session track: Data Intelligence

Session time: 07:30 am - 08:00 am UTC November 12

Session description:

Fashion chatbots are everywhere — but most of them are fashion-illiterate. Ask "show me something like what I picked earlier but in linen" and the bot hallucinates, forgets, or returns something entirely unrelated. The problem isn't the LLM. It's that the LLM has no grounded memory of the catalog, the user's session, or the relationships between the things being discussed. This talk makes the case that a product knowledge graph is the missing layer between a conversational LLM and a fashion catalog. The graph does three things that vectors and flat metadata cannot: it encodes relationships (this style pairs with that; this attribute implies that category), it stores conversational state as a traversable structure (what the user expressed, what was shown, what was liked), and it provides grounded retrieval context that gets injected back into the LLM at every turn — making every response catalog-aware without hallucination. I'll walk through a concrete architecture: how to model a fashion catalog as a knowledge graph (products, attributes, styles, outfit concepts, user intent nodes), how to update the graph as a conversation unfolds, and how to use graph traversal at inference time to construct the context window for the LLM. We'll look at Cypher query patterns for extracting conversational context, the design of the graph schema for incremental updates mid-conversation, and evaluation techniques for measuring whether the chatbot is actually staying grounded. The result is a chatbot that doesn't just respond — it shops with you, carries what you've said forward, and never loses the thread of what you're looking for.