Turning Technical Handbooks into Blog Posts with Generative AI
Introduction
Creating technical documentation and crafting blog posts to communicate the same solutions to a wider audience require distinct skill sets. While an engineering firm may excel at the former, it may lack expertise in the latter. Fortunately, recent advances in generative AI offer a solution that doesn’t require hiring a dedicated content agency. Our proposed solution leverages the company’s knowledge base through a set of technical documents and the company’s desired communication style using previously written blog posts or textual descriptions. This enables us to produce new posts entirely generated and quality-checked by generative AI models, which have honed their writing skills from a vast quantity of training texts.
Challenges in Converting Technical Handbooks
The aim of a technical handbook is to convey its contents to experts. For this reason, it often employs various visual and layout techniques to arrange the information in a limited space. Textual AI models lack the ability to interpret visual cues and require careful reformatting of the input document. Strategies like fixed-rule layout parsing or visual AI models like Google Document OCR address the challenge.
Generative AI for Content Generation
Generating useful content is more than simply instructing a Large Language Model (LLM) to write a blog post based on a technical document. Our solution comprises multiple steps, each executed through a series of distinct LLM prompts. When provided with a technical document as a reference for the blog, we first analyze the company’s previous posts to determine its writing style and possibly adapt it to match the desired tone for the current blog post. This style serves as the foundation for all subsequent prompts. Then we generate a suitable table of contents for the blog post and proceed to fill each section with relevant technical information. Once the blog post is complete, we conduct a stylistic and content review, which guides us in rewriting the post in its entirety.
Evaluating the Quality of Generated Content
Anyone who has used generative AI is aware that the generated text is not always correct, coherent, or well-written, especially when the assignment is complex. For this reason, it is essential to have the capability to evaluate the AI-generated content, identifying mistakes, contradictions, and stylistic or semantic inconsistencies. We have found that, with the appropriate prompts, Large Language Models (LLMs) can provide a significant initial distinction between good and poor-quality posts. Our primary evaluation criteria include:
- Ensuring strict adherence to the desired writing style
- Verifying that the content in the technical document has been rephrased and not simply copied and pasted
- Assessing various textual coherence parameters, such as logical, grammatical, and structural coherence
- Avoiding excessive repetitions of the same sentence structures
- Confirming the inclusion of all key entities from the technical documents in the final post
We have found that, with the appropriate prompts, an LLM can provide a significant initial distinction between good and poor-quality posts.
