For those who aren't familiar with Lookr, we are building an AI sourcing agent that turns a brand owner's clothing idea into a product. Today, a request from our Lookr waitlist prompted us to think deeply. The customer's clothing idea is to 'Represent versatile style! From today’s modern wear to 80’s streetwear...' If this customer were to search for something similar on Amazon, Alibaba, or Google, the results would be empty.
This underscores a long-existing problem: traditional search engines rely on keywords, but what we need in fashion is a deep, semantic understanding.
To solve this problem, we use FashionCLIP
to create image and text embeddings, a model from Hugging Face that's been fine-tuned with 800,000 fashion-related data points. It’s excellent for standard product images and is designed to interpret fashion contexts with high accuracy. We still need more validation on the model of text embedding because other advanced models like SFR-Embedding-Mistral
may also do a good job since it is featured top in the MTEB database.
We’re storing these image and text embeddings in Qdrant
- a vector database.
Choosing the right tool for indexing and querying is crucial. While Langchain
provides a flexible architecture that could lower development overhead, Llamaindex
offers superior speed and performance. In our world, where user satisfaction depends on speed and accuracy, Llamaindex is the preferred choice.
Our AI doesn't just search; it communicates. GPT-4 has been instrumental in this aspect, articulating product recommendations with the confidence and finesse of an experienced salesperson. It understands customer inquiries and responds in a way that not only answers questions but also persuades and engages.
Looking forward, we aim to innovate further by integrating AI-generated tech packs into our search framework. Tech packs are detailed documents that describe how to manufacture a garment, including materials, measurements, and assembly instructions. By translating all product data into tech pack formats before embedding them in our database, we can achieve high precision in matching products to customer queries.
Lastly, I want to express my gratitude to my ex-cofounder & CRO, Cheng Chi
, whose contributions have been crucial in shaping our technological direction. His work has left a lasting impact on our strategies and our capability to innovate.
Stay tuned as we continue to develop and refine our technology. At Lookr, we’re not just following trends—we’re setting them. Join us as we transform how the fashion industry understands and responds to customer needs.