In the realm of e-Commerce, a fundamental problem is accurate interpretation of users' core intent. The intent is often subtly expressed implicitly or stated explicitly with the usage of verbose tokens or key phrases in a user query. In this work, we focus on the later class of problems where we identify a subset of query tokens which are the primary intent bearing phrases that convey explicit intents. We did not solve this as an intent detection problem but rather an immutable component detection problem because we believe that discovering the immutable phrases in a query entails that those are the intent bearing phrases. Furthermore, identifying a certain set of query tokens as immutable ensures better downstream processing in terms of unprecedented token handling, query category detection or query rewrites. We have developed a BERT based supervised learned model which can identify core-intent tokens, thereby improving F1 score over the baseline by over 35%. Furthermore, we integrated our proposed approach for a query recovery strategy which produces approximately 11.9% improvement in offline relevance scores compared to the production model.

In search, autocomplete (AC) is an essential tool that provides suggestions for each keystroke, functioning well with token-based queries. However, it is challenging to handle at scale when input queries are conversational and semantically rich. Identifying relevant queries for sub-tokens requires efficient lookup strategies, real-time ranking, and relevance in the results. This work integrates Retrieval-Augmented Generation (RAG), AI safety, and relevance ranking to produce autocomplete suggestions for conversational queries in a production system. RAG-based responses ensure a high hit ratio for popular AC inputs and maintain a very low risk category by not triggering any critical AI safety concerns.

In the realm of e-commerce, the process of search stands as the primary point of interaction for users, wielding a profound influence on the platform's revenue generation. Notably, spelling correction assumes a pivotal role in shaping the user's search experience by rectifying erroneous query inputs, thus facilitating more accurate retrieval outcomes. Within the scope of this research paper, our aim is to enhance the existing state-of-the-art discriminative model performance with generative modelling strategies while concurrently addressing the engineering concerns associated with real-time online latency, inherent to models of this category. We endeavor to refine LSTM-based classification models for spelling correction through a generative fine-tuning approach hinged upon pre-trained language models. Our comprehensive offline assessments have yielded compelling results, showcasing that transformer-based architectures, such as BART (developed by Facebook) and T5 (a product of Google), have achieved a 4% enhancement in F1 score compared to baseline models for the English language sites. Furthermore, to mitigate the challenges posed by latency, we have incorporated model pruning techniques like no-teacher distillation. We have undertaken the deployment of our model (English only) as an A/B test candidate for real-time e-commerce traffic, encompassing customers from the US and the UK. The model attest to a 100% successful request service rate within real-time scenarios, with median, 90th percentile, and 99th percentile (p90/p99) latencies comfortably falling below production service level agreements. Notably, these achievements are further reinforced by positive customer engagement, transactional and search page metrics, including a significant reduction in instances of search results page with low or almost zero recall. Moreover, we have also extended our efforts into fine-tuning a multilingual model, which, notably, exhibits substantial accuracy enhancements, amounting to a minimum of 16%, across four distinct European languages and English.