eBay

Systems and methods are directed providing a personalized search experience. The system receives a broad term query and identifies a plurality of search results, each having a corresponding embedding. The plurality of search results are clustered into embedding clusters. The system then presents a set of prompts whereby each prompt comprises a display of two or more search results, each chosen from a different embedding cluster. With each selection from a prompt, the system narrows down on one or more embedding clusters. Individual search results from the narrowed-down embedding cluster(s) are then displayed and a user can provide an indication of like or dislike. With each indication, the system dynamically updates a user preference embedding that represents an item that the user desires. The updated user preference embedding is then used to identify updated search results having closest corresponding embeddings to the user preference embedding.

Methods, systems, and computer media are disclosed for generating enhanced query suggestions by integrating collaborative filtering, embedding-based similarity, and large language model generation techniques. The system accesses a query keyword and produces three types of suggestions: one derived from historical user behavior, one based on semantic similarity using embeddings, and one generated by an LLM. These suggestions are stored in association with the query keyword and may be retrieved for future queries. For new or previously unseen queries, the system generates a semantic hash and identifies the most similar stored keyword using hash comparison or embedding similarity. The technology improves upon traditional collaborative filtering by enabling relevant and diverse suggestions even for low-frequency or non-head queries. This approach enhances user engagement and search effectiveness on e-commerce platforms by aligning suggestions more closely with user intent, regardless of historical data availability. The system supports scalable, real-time performance across large query datasets.

Some aspects relate to technologies for a search engine that performs item retrieval on a listing platform using chain-of-thought (CoT) query categorization. The system receives a search query (e.g., a natural language query) and performs a tree search on a hierarchical taxonomy of categories using a language model that semantically scores category nodes. Less relevant branches are pruned, and traversal continues along high-scoring paths using different traversal approaches, such as a breadth-first or depth-first approach. This process identifies candidate leaf categories aligned with the search query. The language model then scores these candidate categories. Based on the relevance scores, one or more relevant leaf categories are selected. The system then queries an item listings data store constrained by the selected categories and returns the resulting search results.

Methods, systems, and computer storage media for providing a cross-domain secure authentication engine in an item listing system are described. The cross-domain secure authentication engine is an advanced, modular system designed to handle user authentication seamlessly across different service domains, while ensuring security features and an adaptive user experience. In an item listing system, the cross-domain secure authentication engine operates based on a security-focused framework that streamlines user authentication across multiple domains without sacrificing user experience. It integrates a modal containing an iframe-based sign-in interface associated with a separate authentication domain and employs a postMessage API for cross-origin communication. Key features include dynamic modal resizing-adjusting the sign-in window to different authentication steps (like email, password, or OTP)-and parameterized iframes that remove unnecessary UI elements for a cleaner user experience. The cross-domain secure authentication engine also supports security measures and supports incremental, multi-step authentication flows to reduce user friction.

The technology described herein relates to systems, methods, and computer storage media, among other things, for generating recommendations corresponding to whether requests (e.g., transmitted by computing devices) are malicious network traffic. For example, the recommendations can be generated using various historical and current network traffic trends (e.g., associated with a particular application programming interface). As another example, the recommendations can be generated based on particular models. Based on the generated recommendation, dynamic rate limiting rules can be applied for determining whether a request is malicious network traffic. Based on determining the request is malicious network traffic, the request (and additional requests associated with that particular request) can be blocked. The applied dynamic rate limiting rule can be adjusted based on a change to the current network traffic associated with the application programming interface.

In an example embodiment, a transaction log index is used to index log records. The transaction log index comprises an in-memory transaction log index and an on-disk transaction log index. The log records are temporarily indexed in the in-memory transaction log index until the storage server can ensure that the corresponding log records are stored without any missing log records in a sequence, as well as possibly based on other correctness and performance considerations.

Some aspects relate to technologies for performing item retrieval on a listing platform using clusters of interchangeable parts formed using fitment match. In accordance with some aspects, item data is accessed for part item listings on a listing platform, where the item data for each part item listing includes fitment data for each of a number of different fitments. For each part item listing, a fitment hash is generated using fitment data for each fitment of the part item listing. The part item listings are clustered based on overlap of fitment hashes for the part item listings. Cluster data is stored for the part item listing clusters. The cluster data for each part item listing cluster associates a cluster identifier and an item listing identifier for each part item listing in the part item listing cluster. The cluster data can be leveraged to perform item retrieval for the listing platform.

Methods, systems, and computer storage media for providing an advertising campaign management engine in an artificial intelligence system are described. Advertising campaign management refers to the process of planning, executing, monitoring, and optimizing advertising campaigns. Operationally, first, high-conversion design modeling and mining analyzes performance data and user preferences to identify design elements-such as image style and text layout-that drive engagement. Second, design requirement generation converts ad optimization profiles into structured prompts that guide the refinement of ad designs. Lastly, design fine-tuning leverages advanced generative models to adjust ad elements, such as images and layouts, based on the generated requirements. The advertising campaign management engine supports an iterative process that enables automated updates and optimizations, allowing ads to be tailored to specific demographics. For example, it could produce quirky designs for young adults or simplified layouts for seniors. Refined ads are then deployed and continuously monitored for performance improvements.

Some aspects relate to technologies for generating and validating content providing summary information for parts on a listing platform, such as compatibility information and item characteristic information for the parts. In some aspects, input data comprising attribute-value pairs based on one or more items is accessed for content generation. The input data and one or more pre-defined generation prompts are provided to a first generative model, which generates content comprising natural language text as output. In some aspects, the content is validated by providing the content and one or more pre-defined validation prompts to a second generative model. If approved, the content can be provided for presentation on a user device. If rejected, the content is not provided for presentation. In further aspects, the rejected content is used by a third generative model for tuning the pre-defined generation prompts.

A search query is received. A first plurality of items is identified. A reference item is identified based on the search query and the first plurality of items. A plurality of contextual distances is determined using a distance calculation metric. A machine learning model is used to rank a second plurality of items based on the plurality of contextual distances. The second plurality of items is caused to be displayed. The second plurality of items is arranged in an order in accordance with the ranking.