Product Description
To connect it, you need to contact your manager, or contact support at support@retailrocket.net
Key Features
- Semantic search engine, based on modern language models (LLM): processes natural language queries, including alternative formulations and common errors;
- User behavior. Helps to find first and foremost products and their categories that are most important in the context of the search phrase. Different products or their groups may equally well match the search phrase in meaning, but users may expect only some of them first. For example, for the phrase “dishes” one can and should find antique and rare dishes, but first of all one should show ordinary sets and plates. At the same time, it is impossible to show even very popular plates everywhere;
- Dynamic SERP personalization (search results page): based on user behavior and the current state of the catalog. Connected upon partner request;
- UI in the spirit of native search: results are integrated into the design of the website or mobile application, maintaining the brand style of partners;
- Understanding slang: Semantic search out of the box understands common human language, but in some areas specific vocabulary is used, which we can add to the algorithm;
- Multilingual support: understanding queries in different languages, including alternative or mixed formulations;
Main Components
Semantic search using modern LLMs
It is important that this mechanism works even if the search phrase describes the product not by words from its name or description, but by arbitrary synonyms, by its purpose or other closely related means.
- Finds products that are suitable in meaning to the search phrase;
- Uses semantic vectors;
User behavior
Tracking the behavior of all users of the store allows identifying products and categories that are truly important for a specific search query, because even if different groups of products formally match the query, users more often expect to see some of them first.
For example, for the query "dishes", the system should find both antique items and rare collections, but with higher priority — familiar sets of plates. At the same time, it is important to maintain balance: popular products should be shown where they are appropriate, and not in all cases.
Current user behavior - takes into account the individual actions of the user and allows showing them products that best match their interests first.
You can find out how to integrate and set up user behavior tracking by this link.
Classic search mechanisms
- Using keywords. Although semantic search reduces dependence on keywords, their use in some cases helps to increase the accuracy of results.
- Attribute prioritization. Semantic search takes into account many semantic factors and is able to select products based on matching or similar attributes. However, sometimes it is required that the system first show products with precisely those attribute values that are explicitly specified in the query.
- Accounting for slang. Semantic search processes natural language, but in certain subjects, specific terminology is used. Such a limited set of expressions can be additionally trained in the system.
- Correcting layout errors. If a query is entered in an incorrect layout, but at the same time forms a correct word in Russian or English, the system automatically corrects it.
- Correction of typos. If the query contains a non-existent word similar to one known in the catalog, it is replaced with the closest semantic variant.
Terminology
- Search results: The set of products shown to the user;
- Search phrase: User's text query;
- Search phrase input field: The input field for the search phrase, usually visible on every page of the store's website in its upper part;
- Search preview: Part of the Rocket Search product with a small search results and additional tools, which appears under the search phrase input field at the moment of its use. Implemented as a separate web component;
- Search results: Part of the Rocket Search product with full search results and additional tools. Located on a separate page, implemented as a separate web component;
- Facet filters: Part of the search results page that implements faceted filtering of search results;