Walter Heck (Helixiora): Smarter Search: How RAG Pipelines Help Navigate Complex Documentation

This talk was presented at the AI The Docs 2025 online conference. We are thrilled to share the recording and the summary with you. 

Visit the talk summary page to see all of the presentations from the conference.

presentation

Walter Heck (Co-Founder and CTO at Helixiora )

Summary

How can a RAG system effectively handle diverse user queries, distinguishing between questions about user data, product capabilities, or general conversational interactions, to provide a seamless user experience? 

In his talk, Walter Heck (Co-founder of Helixiora BV) presented how his team implemented a RAG-based documentation search. The system enables users to ask questions about both their own data and the product’s capabilities. 

Core challenge: RAG alone cannot distinguish between different kinds of user queries (e.g., casual greetings, product questions, data lookups). To provide a seamless experience, Helixiora added an LLM-based classification step that sorts each query before retrieval. 

How diverse queries are handled: 

• Conversational: greetings or small talk → answered directly.
• Error: gibberish or unclear input → prompts clarification.
• Meta Model: questions about the product itself → routed to product documentation.
• Metadata: questions about data indexing details → handled by metadata queries.
• Informational: actual user data questions → go through the RAG pipeline. 

Implementation highlights: 

• Indexing & Querying: Documents (Google Drive, Slack) embedded into Pinecone; queries embedded and matched with top results before LLM answer generation.
• Documentation Management: Markdown + MKDocs, stored in GitHub, selectively re-indexed; recency checks ensure the newest versions are used.
• Seamless experience: Classification ensures users feel the system “understands” the intent of their query and responds appropriately. 

Key takeaways: 

• RAG reduces hallucinations and provides source-backed answers.
• Automated indexing keeps content fresh.
• Classification is essential for diverse query handling, bridging product docs, user data, and conversational input.
• Future work: scaling to more complex, multi-document queries.