Adaptive Transformer Search (ATS) is a proprietary technology developed by Particular Audience designed to address the significant issues in ecommerce search. Traditional keyword search systems, which rely on exact token matching and manual rules, are considered ineffective for a large majority of consumers and result in substantial financial losses for retailers due to lost sales and customer abandonment. The problems with legacy search include poor performance for long-tail queries, difficulty understanding synonyms and context, reliance on inconsistent data, and the significant manual effort required for configuration, particularly for the vast majority of search terms which fall into the long tail.
ATS tackles these challenges by utilizing Large Language Models (LLMs) and vector embedding technology to comprehend the meaning and intent behind user queries and product information. Instead of matching keywords literally, ATS transforms both queries and product data into dense vectors. These vectors are numerical representations in a high-dimensional space, and the proximity or angle between them in this space indicates their semantic similarity. This approach allows for a much richer understanding and more relevant data retrieval compared to simple token matching.
Particular Audience creates proprietary Vertical Tuned Models (VTMs) by fine-tuning open-source transformers on specific retail vertical data sets. This fine-tuning process enhances the accuracy of vector embeddings specifically for retail contexts. VTMs enable the system to understand shared meaning based on context (like 'lightweight', 'portable', 'small', etc., when referring to a tripod) and significantly improve relevancy scores. VTMs also allow ATS to suggest substitute products when a searched item is not available (e.g., showing Samsung phones if a customer searches for a Huawei phone and the retailer doesn't stock them). This capability can dramatically reduce zero search results.
ATS also incorporates Synthetic Data, generated by AI, to train query-click pairs. This helps to fill gaps in existing data and improve model training. Furthermore, the system employs Adaptive Reinforcement Learning (ARL), allowing the VTMs to learn continuously from live user behavior on a retailer's site. This adaptive process automatically adjusts embeddings over time, improving precision and accuracy in the specific context of that retailer without requiring manual intervention.
Beyond just relevancy, ATS offers Contextual Ranking, which can personalize search results by interpreting implicit user signals (such as clickstream data and items in the shopping basket). This helps to predict demand and potentially increase click-through rates. ATS also enhances Relevant Search Ads by moving beyond keyword matching to semantic understanding, improving the coverage and performance of sponsored products, especially for long-tail queries.
Privacy is a core design principle for ATS; it collects no personally identifiable information and does not use third-party tracking. The technology's ability to interpret natural language text also prepares it to handle future conversational interfaces and complex question-and-answer queries.
In summary, ATS combines the power of vector search and traditional keyword search with vertical tuning and localized reinforcement learning. This results in intuitive semantic search experiences that aim to boost conversion, increase sales, build customer loyalty, and reduce the manual effort required for search configuration.
