May 22

With the advent of generative text AI systems many potentials to improve efficiency in technical documentation could be leveraged. The writing process is often supported with Large Language Models, with proofreading of the generated materials by humans. Using a Large Language Model with a human-in-the-loop for proofreading and checking against the required standards and often (unwritten) rules, the efficiency of the creation process may be further improved.

Good prompt engineering is key for generating novel texts and making these fit into a company’s language. When working with LLMs for proofreading, the quality of a good prompt and the provided additional material is getting even more critical to achieve the results that are expected. During the presentation, Thomas will provide an outline of what makes good input-data good and how to use the available data to make a Large Language Model an efficient proofreader that provides valuable insights. Using a human-in-the-loop approach, the LLM provides both an increased efficiency and also the required quality.

The Huawei network product information experience team is required to deliver tens of thousands of product documents to partners each year. How can the application of large language models help users efficiently and promptly retrieve accurate content from vast amounts of information, thereby assisting them in independently completing project delivery? We have carried out related practices by building a unified intelligent knowledge base and developing AI applications such as intelligent Q&A and virtual assistants. This approach effectively addresses key experience issues such as low information retrieval efficiency and inaccurate content generation, significantly improving partner satisfaction with the information experience.

Based on multimodal deep learning technology, Hikvision’s self-developed LanZhi large model provides intelligent Q&A services and customized knowledge base services for all internal users. It also offers a variety of customized applications, including intelligent programming, insights, testing, project management, knowledge Q&A, and documentation.
In documentation writing, Hikvision faces challenges such as high content redundancy, heavy workloads, and high costs. By leveraging the powerful language capabilities of the LanZhi model, the team has explored all aspects of the entire process of assisting with document writing and application, as demonstrated by:

1）Providing outlines and text polishing in the initial stage of documentation writing;
2）Quickly and intelligently extracting key parameters for repetitive and complex documents;
3）Inspecting content with AI to rapidly verify document compliance with standards and check for typos, formatting, and terminology;
4）Translating and converting documents into multiple languages;
5）Implementing intelligent Q&A based on internal terminology and a knowledge base formed from technical documents in terms of documentation application.

To reduce the unpredictability of AI-generated content, Hikvision has implemented fine-tuning training and built small models for proprietary domains to enhance the reliability of generated content. This documentation application can be extended to all Hikvision subsidiaries and significantly improve the writing efficiency and quality of operational manuals, user guides, troubleshooting manuals, and more.