Putting relevant and up-to-date facts into the prompt so that the model bases its answer on them, which cuts fabrication and makes citations possible. It is central to RAG.
Also called grounding data.
Read more: Microsoft Learn
In the Ultra Transcenders books
Each book explains Grounding in context, with comparison tables and the common traps.
Terms in this definition
- Fabrication
Convincing yet wrong content produced by an LLM, which does not check facts. Grounding lowers how often it happens but can't remove it entirely.
- RAG
Technique for answering from private or recent information: relevant content is fetched from your own data, typically a search index, and inserted into the prompt as grounding so the model can respond with citations.
Related terms
- Azure OpenAI On Your Data
Feature of Azure OpenAI, now deprecated, for grounding chat answers in your own content retrieved from a search index. Its replacement for new work is Foundry Agent Service paired with Azure AI Search or Foundry IQ.
- Foundry IQ
A managed knowledge layer for Foundry agents, made up of knowledge bases and knowledge sources and built on agentic retrieval in Azure AI Search. It provides grounding that respects permissions and includes citations.
- GPT-4 vision-preview
Preview model adding vision to GPT-4, with OCR, object grounding and video enhancements. Now retired, it gave way to gpt-4 turbo-2024-04-09 and later GPT-4o, so it should not be chosen for new image scenarios.
- Metaprompt and grounding mitigation layer
Layer of generative AI mitigation at the application level, in which grounding data and the system message guide the model away from output that is ungrounded or harmful.
- Microsoft 365 Copilot and Copilot Chat location
A Purview DLP location that can only be chosen in a Custom policy and cannot be combined with other locations. It can keep Copilot from processing emails and files carrying selected sensitivity labels, block prompts that include selected sensitive information types, or stop external email being used for grounding.
- Prompt Shields for documents
Prompt Shields check that defends against indirect (cross-prompt) injection by scanning third-party material, such as emails, grounding data, OCR text or tool output, for planted instructions.
- Training cutoff
The point in time beyond which a model knows nothing. Grounding with RAG at request time, not retraining, is how recent events are covered.