A LLM based customer service tool in complex engineering environments

A LLM based customer service tool in complex engineering environments

The Challenge

World’s leading specialist in water treatment needs to unlock the knowledge within a large repository of technical specifications, user manuals and product sheets to support their customers and customer service representatives/account managers in supporting issues with a customer installations. The objective is to enable new recruits and experienced account managers access contextually relevant information quickly to support diagnosis and implementing solutions, to reduce down time at their customer installations.

Hypothesis

  • The RAG based LLM will be able to be sensitive to technical abbreviations and provide relevant answers.
  • Complex layouts of technical user manuals and product sheets can be understood and appropriately interpreted by the LLM and incorporated in different contexts.

Execution

  • Abbreviation Handling System: Specially designed for technical terms, ensuring accurate comprehension and processing by Large Language Models (LLMs).
  • Processes tabular data through a vision model, enhancing data identification and summarization for improved accuracy and usability.
  • Implements various strategies to enhance the precision and relevance of data chunk retrieval, ensuring high-quality information delivery.
  • A robust validation system assesses the quality of answers generated by Retrieval-Augmented Generation (RAG). In cases of suboptimal responses, it leverages public GPT knowledge to enhance accuracy and completeness.

Outcomes

  • LLM deployed in production that manages queries in multiple languages and with varying intents, providing accurate and contextually relevant responses.
  • The LLM filters knowledge base dynamically based on user roles, permissions, and regional data compliance, ensuring secure and relevant access.
  • Custom-built Serverless Pipeline: Efficiently reads, chunks, and constructs a robust knowledge base from various data sources.

Project Highlights

  • Dramatically Reducing Diagnosis and Response Time to Clinet Installation Issues