This post talks about how enriching DITA with iiRDS and a trusted knowledge base yields the most reliable GenAI outputs, which is especially critical for generating safety-related content.
While headlines trumpet the revolutionary potential of large language models (LLMs) and GenAI, a critical question remains unanswered in technical documentation circles: can we trust AI-generated content when safety and precision are non-negotiable? AI systems promise sweeping solutions at a fraction of traditional metadata management costs, but their reliability deserves careful evaluation, particularly when it comes to safety-critical information.
Assessing AI Reliability for Technical Content Creation
Standard LLMs show significant shortcomings such as lack of traceability and explainability of results, sequence handling limitations, inability to prioritize content importance, and so on. But when safety instructions guide maintenance workers or operating procedures direct technicians, even small errors can have serious consequences.
In their article iiRDS for Trusted AI, Harald Stadlbauer, Markus Kronfellner, and Helmut Nagy look at three approaches to generating technical information:
- Standard LLM implementation
- LLM with vector-based retrieval augmented generation (RAG)
- LLM with RAG optimization based on iiRDS (Intelligent Information Request and Delivery Standard)
As hydraulic systems present many safety challenges, they chose to examine the documentation for a SANDVIK bolter miner for their use case. Hydraulic system issues aren’t easily diagnosed because similar symptoms can stem from multiple causes, so accessing the correct safety warnings at the right time is critical.
For assessing the third approach, they enriched DITA (Darwin Information Typing Architecture) topics and maps with the SANDVIK taxonomy, a mining industry-specific domain taxonomy, and the iiRDS core and machinery ontology. Consequently, the iiRDS-RAG system was able to use corporate knowledge as context and select the most relevant and specific content, tagged by iiRDS, to feed the LLM.
The results of the assessment reveal a clear winner for organizations that can’t afford to compromise on accuracy. The iiRDS-based RAG — enhanced with trusted knowledge — significantly outperformed both the standard LLM and the simple chunking vector-based RAG approaches. As the system drew from contextualized information rather than generalized language patterns, its retrieval accuracy was superior.
Implementing Trusted AI in Documentation Workflows
Currently, iiRDS is the only standardized schema, specifically designed for technical information retrieval. The above assessment of the three approaches to generating technical information conclusively demonstrates that combining iiRDS with the language capabilities of LLMs provides the best quality GenAI output.
By focusing on knowledge-first approaches instead of raw AI capabilities, organizations can ensure that they don’t just generate more technical documentation content faster, but can deliver trusted information exactly when and where it’s needed.