Grounding
A process or system-level mechanism that connects generation or answers to external evidence, retrieved information, or other materials so that claims can be based on that information.
Example
Before generating an answer, a system retrieves relevant documents and supplies them to the model as reference information.
Distinctions and cautions
In this catalog, Grounding is a process or system-side concept, whereas Groundedness is a property of the output. Retrieval-Augmented Generation (RAG) can be one means of grounding, but the two are not synonymous. Retrieval alone does not ensure well-supported output if the model ignores or misinterprets what was retrieved.
Terminology note
The distinction between Grounding and Groundedness is an editorial convention adopted by this Catalog, not a standardized terminology system for the research field as a whole.
Evidence
View evidence →
Overview of the TREC 2025 Retrieval Augmented Generation Track
- Title
- Overview of the TREC 2025 Retrieval Augmented Generation Track
- Author / organization
- NIST TREC 2025
- Year
- 2025
- Source type / publication status
- Official evaluation proceedings; published
- Supported concepts / claim context
- Grounding, completeness, attribution, retrieval/relevance, RAG evaluation.
- Scope and limitations
- Operational definitions for a shared evaluation task, not standardized definitions.
- Access / version note
- Official proceedings page.
- Recorded official URL / DOI / identifier
- Official URL / recorded link
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
- Title
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
- Author / organization
- Lewis et al.
- Year
- 2020
- Source type / publication status
- NeurIPS peer-reviewed research; published
- Supported concepts / claim context
- RAG definition; retrieval plus parametric generation; factuality benefit.
- Scope and limitations
- Describes the original RAG architecture, not every present-day RAG variation.
- Access / version note
- Published NeurIPS proceedings.
- Recorded official URL / DOI / identifier
- As recorded:
NeurIPS 2020 proceedings URL (exact URL not recorded here)(no URL inferred)
FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation
- Title
- FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation
- Author / organization
- Zhang et al.
- Year
- 2025
- Source type / publication status
- ACL peer-reviewed research; published
- Supported concepts / claim context
- Grounding, Faithfulness, Knowledge Conflict, RAG limitations.
- Scope and limitations
- Conflict-focused RAG conditions; strict adherence to context can also have failure modes.
- Access / version note
- Published ACL.
- Recorded official URL / DOI / identifier
- DOI:
10.18653/v1/2025.acl-long.1062(DOI resolver)
Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models
- Title
- Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models
- Author / organization
- Wang et al.
- Year
- 2025
- Source type / publication status
- ACL peer-reviewed research; published
- Supported concepts / claim context
- Retrieval Failure, RAG, Knowledge Conflict.
- Scope and limitations
- Controlled RAG analyses; no universal extrapolation to all retrievers or domains.
- Access / version note
- Published ACL.
- Recorded official URL / DOI / identifier
- DOI:
10.18653/v1/2025.acl-long.1476(DOI resolver)