Guide for Creating a Multilingual Query Resolution System similar to Zepto from the ground up.
Zepto, a leading technology company, has unveiled a novel multilingual query resolution system that effectively addresses misspellings and enhances search quality across various languages. The system combines the power of Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) to deliver a robust and efficient search experience.
The process can be broken down into three stages: input and query, processing, and final query refinement and search. In the first stage, the system takes misspelled or vernacular queries and converts them into multilingual embeddings using a multilingual embedding model. This initial embedding serves as a foundation for the subsequent stages.
The second stage involves processing, where the system uses Instruct fine-tuning, a lightweight tuning method that employs stepwise prompts and role-playing instructions. The system then utilises FAISS, Facebook's similarity-search engine, to find the top K brand and product names that are closest in the embedding space to the query.
The noisy query and the retrieved names are then fed to an LLM prompt, which outputs a clean, corrected query. Zepto employs Meta's Llama3-8B, hosted on Databricks for cost control and performance, for this purpose.
The final stage involves parsing the JSON output from the LLM, extracting the corrected query, and rerunning the similarity search on the Vector DB to find the details of the searched product. The system, when tested with challenging queries, can correct the raw and noisy user query with the exact and corrected brand or product name.
Key features of the system include its robustness, scalability, and its ability to improve user experience and search conversion rates. The system can correct misspellings and slang with high accuracy, understand multilingual queries, disambiguate queries by using retrieved context to infer user intent, and provide structured, auditable outputs, showing not just the correction but also the reasoning behind it.
In summary, Zepto's system uses LLMs to interpret and correct user queries while leveraging RAG to incorporate external knowledge retrieval, enabling a robust multilingual search system that handles misspellings and improves search relevance and quality. This approach offers a powerful combination of natural language understanding and up-to-date information retrieval crucial in complex search tasks.
Although direct technical details about Zepto’s exact multilingual query resolution system are not fully detailed in the search results, this explanation fits typical industry approaches combining LLMs and RAG for such purposes.
- In the realm of data science, Zepto's novel query resolution system applies Artificial Intelligence techniques to correct misspellings and enhance search quality across various languages, particularly in home-and-garden, lifestyle, data-and-cloud-computing, and technology domains.
- During the processing stage of the system, the utilisation of technology like Instruct fine-tuning, FAISS, and Meta's Llama3-8B (hosted on Databricks) showcases the impact of AI and data science on improving the efficiency and robustness of search engines.
- The system's final stage, refinement and search, extends the capabilities of artificial intelligence and data-and-cloud-computing by providing structured and auditable outputs, demonstrating both the corrected query and the reasoning behind it, thus offering an enhanced user experience and increased search conversion rates.
