AI securitysrc: curated AI glossary

Grounding

Grounding is like giving an AI a textbook to look at while it answers your questions, so it has to stick to the facts in the book instead of just guessing or making things up.

Grounding is the process of connecting an LLM to trusted external data sources, typically through Retrieval-Augmented Generation (RAG), to ensure that generated responses are constrained by verifiable evidence and minimize the risk of hallucinations.

Tying a model's output to verifiable external sources or data (often via retrieval) so claims are supported by evidence rather than fabricated, reducing hallucination.

evolution

2020-05 · history
GPT-3 and In-Context Learning
The release of GPT-3 demonstrated that providing examples in the prompt could steer model outputs, laying the groundwork for retrieval-based grounding.
2020-07 · history
RAG Introduction
Lewis et al. introduced Retrieval-Augmented Generation (RAG), formalizing the method of grounding LLMs by querying external knowledge bases.
2023-02 · history
Bing Chat Launch
Microsoft integrated RAG into a consumer search product, popularizing the use of citations to ground AI responses in real-time web data.
2023-09 · history
Enterprise RAG Standardization
Major cloud providers began offering managed RAG services, establishing grounding as a standard security requirement to mitigate hallucinations in enterprise AI.

← all terms