Dynamic attribution and/or modification of responsive content that is generated using a retrieval augmented generation (rag) process

Topics: AI Mode, AIOverviews, Chunk Relevance, LLM Readability, LLMO / GEO, Passage based retrieval, Personalization, Probably in use, Retrieval Augmented Generation (RAG)

This patent describes a system and method developed by Google for dynamically determining whether content generated by a generative model (such as an LLM) through a Retrieval Augmented Generation (RAG) process needs to be modified or attributed to its original sources. When a user submits a query, the system retrieves search result documents, generates responsive content using a generative model, and then compares segments of that generated content against the retrieved documents and the model’s training data to detect matches. Depending on the type and source of the matching content — such as public domain webpages, licensed content, or non-public domain material — the system applies different rules to modify the output, which may include adding source links, truncating content, omitting segments, or regenerating portions entirely. The approach uses a progressive, tiered search strategy that first checks against retrieved documents (a smaller search space) before expanding to the much larger training dataset, thereby conserving computational resources and reducing latency.