[2026 Latest] Automated Generation of Preventive Measures via Cross-Referencing Past Incidents Using RAG (Retrieval-Augmented Generation)

In manufacturing and construction sites, daily "daily reports" and "near-miss reports" are valuable information assets for preventing industrial accidents. However, in many companies, this data is buried as massive amounts of text, making it difficult to instantly reference similar past incidents and take effective measures. As of 2026, with the introduction of a semantic search infrastructure utilizing RAG (Retrieval-Augmented Generation), systems that instantly cross-reference new reports against vast past databases to automatically generate specific preventive measures have been put into practical use. This article explains the advancement of safety management through AI and its specific implementation approaches.

1. The Mechanism of RAG: Turning Tacit Knowledge from the Field into Explicit Knowledge

Traditionally, "identifying similar incidents" relied on the empirical rules of veteran managers, but this is now being digitized through the introduction of RAG. By converting unstructured descriptions in daily reports into vector representations (embeddings) and storing them in a vector database, context-based semantic search—going beyond simple keyword matching—becomes possible.

For example, for a report stating "the scaffolding was slippery," the system instantly extracts semantically similar past cases such as "fall risks after rain" or "accidents due to oil adhesion." This allows small signs (near-misses) occurring on-site to be immediately linked to lessons learned from major past disasters.

2. Improving Accuracy through Cross-Referencing with Past Cases

The greatest advantage of RAG lies in its ability to dynamically inject company-specific "past incident data" into the general knowledge held by LLMs (Large Language Models). In the cross-referencing process, the top few past cases similar to the new report are included in the prompt, allowing the LLM to formulate proposed measures (In-Context Learning).

According to data, sites that have introduced automated generation systems using RAG have seen a reduction of approximately 85% in the time required to identify similar incidents compared to manual searches, and the specificity of proposed measures has significantly improved. The graph below shows the efficiency of information extraction in conventional search versus RAG-based search.

3. Automatic Summarization of Preventive Measures and Automation of Internal Sharing

Based on the retrieved past cases, the LLM automatically summarizes preventive measures in formats such as "Three key points to watch out for in this specific case." A key feature is that these are not mere copies of the past, but are generated as dynamic advice that considers current site conditions (weather, worker skill levels, types of equipment).

The generated summaries are immediately sent via push notifications to site supervisors' tablets or internal chat tools. This "real-time nature from incident occurrence to measure presentation" is a decisive factor in preventing secondary disasters. The transition from person-dependent safety management to data-driven, automated safety management has also become an important indicator in corporate ESG management.