Emil Soerensen (kapa.ai): Optimizing Technical Documentation for LLMs

Emil Soerensen shared insights on how to optimize technical documentation for large language models (LLMs). He highlighted Kappa's work with companies like OpenAI, Prisma, and Mapbox to deploy LLMs that assist in answering technical questions within developer communities. 

Key Takeaways 

• Introduction to LLMs in Technical Documentation 
  • Overview: Emil explained how LLMs work in technical documentation, using Kappa’s system as an example. The process involves connecting the LLM to various technical knowledge sources, deploying the LLM, and utilizing the insights from user interactions. 
• Optimizing Technical Documentation 
  • Page Structure and Hierarchy: Proper structuring with clear headings and subheadings helps LLMs navigate and understand the content better. An example from Temporal demonstrated effective hierarchy. 
  • Segment Documentation by Sub-Products: Separating documentation for different sub-products prevents confusion and enhances context understanding. Prisma’s segmented documentation serves as a model. 
  • Include Troubleshooting FAQs: FAQs in a Q&A format align with how users interact with LLMs and improve their effectiveness. OpenAI’s use of technical FAQs is highlighted as a best practice. 
  • Provide Example Code Snippets: Self-contained code snippets with descriptions and comments help users and LLMs understand complex SDks and APIs. Mixpanel’s documentation is a notable example. 
  • Build a Community Forum: Forums offer a valuable source of information for both users and LLMs. Kundas’s active community forum illustrates the benefits of this approach. Filtering and proper attribution of forum content are important considerations. 
• Future of Documentation Platforms 
  • Adoption of AI Technologies: Within a year, more companies are expected to adopt LLMs for their documentation due to their proven benefits. The rapid advancement of AI technology makes it difficult to predict long-term changes. 
• Balancing Writing for LLMs and Humans 
  • Consistency in Quality: Good documentation for humans usually translates well for LLMs. Writing for LLMs might require simplifying content slightly, but the fundamental principles of clear and structured technical writing remain the same. 
• Defensive Measures Against Spam 
  • Preventing Manipulation: Implementing technical defenses is essential to prevent bots from skewing user interaction data and affecting the roadmap. 

Emil concluded by inviting attendees to explore Kappa’s LLM systems and mentioned a related blog post on Hacker News, underscoring the overlap between writing good documentation for humans and LLMs.