Our models leverages both product photos, a text prompt and more. That means they can spot a particular object (like a couch or a shirt) in a crowded photo and provide detailed information from the image.
.png)
Our model incorporates breakthrough advancements in open-domain object identification using text prompts. This dramatically improves upon traditional methods that depend on predetermined output labels.
Classifies 329 colors and roughly 50 different patterns.
Outputs the correct label name for a product, based on the image, description, and other associated text.
Uses multimodal understanding to classify materials such as brick, tile, ceramic, and wood.
Intelligently groups products according to whether they’re suited for infants, teens, or adults, among other demographic categories.
Uses multimodal understanding to classify products according to which room (such as kitchen, living room) they are typically used in.
Our model incorporates breakthrough advancements in open-domain object identification using text prompts. This dramatically improves upon traditional methods that depend on predetermined output labels.
Classifies 329 colors and roughly 50 different patterns.
Fast, efficient language generation model tailored for e-commerce based on input text and image.
.svg)
.png)
.png)
.png)
Vody's model combines visual and textual inputs to build a holistic understanding of each product. This allows the model to incorporate complex concepts like color perception to accurately classify the colors of objects.
.png)
Our model incorporates breakthrough advancements in open-domain object identification using text prompts. This dramatically improves upon traditional methods that depend on predetermined output labels. For example, the model can ignore the background color of an image and instead focus on the color of the relevant object, such as a table or chair.
We created synthetic training data to reflect various lighting conditions. This enhances the model's ability to recognize color changes caused by different environments.
.png)
Vody's model combines images, text, structured data, and user click patterns (“clickstream” data) to build a holistic understanding of each product (with our embedding model being the first ever to incorporate structured data and clickstream). Given that 60% of shoppers are frustrated by irrelevant results and 80% of shoppers say that accuracy and convenience are the most important aspects of online shopping, the improved model understanding of products will substantially improve the quality of users’s search experience and drive both customer acquisition and retention.
.png)
Our embedding models enable textual and visual search (both separately and in tandem with each other) and content-based recommendation out of the box. In addition, our embedding models can be fine-tuned on a variety of different e-commerce tasks, including product compositionality, cart filling, and product categorization, to fit whatever your company needs.
.png)
By training our model specifically on e-commerce related-data, we are able to dramatically improve our models’ understanding of e-commerce-related items across a variety of different product domains (furniture, pets, clothing, groceries, etc.).
.png)
Augmenting training data to reflect various lighting conditions, rotations, translations, etc. makes the model resistant to small perturbations in the input product image, meaning you can tailor our embeddings to your products rather than your products to the embeddings.
.png)
Vody's generative models combine images and text as part of a multimodal LLM, offering ~20% better performance than traditional LLMs at ~200x smaller size.
.png)
Similar to the emergent capabilities found in traditional LLMs, our e-commerce specific multimodal LLMs are able to do several e-commerce tasks out-of-the-box, including product description generation, product titling, product Q&A via chat, and much more. We greatly improve the zero-shot capabilities of unseen tasks by incorporating instruction tuning on our world-class proprietary e-commerce datasets.
.png)
By training our model specifically on e-commerce related-data, we are able to dramatically improve our models’ understanding of e-commerce-related items across a variety of different product domains (furniture, pets, clothing, groceries, etc.) as well as their ability to generate text more specific to e-commerce applications. For example, our models are far better at writing about products in a way that sells well to interested shoppers.
.png)
Unlike others in the LLM space, Vody provides a list of best-practice prompts in order to optimize performance across a variety of different generative tasks. Our research team works tirelessly to discover the best prompts so that you don’t have to.
1. Collect a labeled dataset for the task you want to test the model on. For our color model, this may look like a dataset of images and product titles mapped to the color of the requested item in the images.Use an automatic metric to compare the model’s predictions to the correct answers.
2. Accuracy will work well if your labels are relatively balanced (i.e., if you are checking for 10 patterns, each pattern should comprise close to 10% of the overall dataset). If they aren’t, we recommend using F1 instead.
.png)
Automatic. Can be further grouped into similarity-based evaluation metrics (L2, cosine) and top-k evaluation metrics (compare the k most similar predictions to a label of the top k predictions from a dataset and see how many match).
Human Evaluation. Take the approach from top-k evaluation metrics but have humans evaluate it in ways more relevant to your business. For example, if you’re using embeddings for search, you can return the top 10 results for a search and measure the number of clicks on purposes resulting from that list of 10.
Manual Evaluation. Have a human provide ratings (e.g., 1-5 stars) for each generation from the model.
Reward Model. Based on a labeled dataset of human ratings for a model’s generations, train a regression model to predict what a human rating for a generation would be. Use this model to label generative outputs at scale.
.svg)
.svg)
.svg)
.svg){kind=logo}
.png)
.png)