Grounding

Process evaluation published

Also known as: Grounded Generation, Source Attribution

Definition

Constraining LLM outputs to be based on and traceable to specific source material, rather than generated from the model's parametric knowledge alone. Grounding connects generated text to verifiable sources, enabling fact-checking and reducing hallucination. Implemented through RAG, citations, and attribution.

What this is NOT

Not just RAG (grounding includes attribution and verification)
Not preventing all errors (grounding reduces certain types)
Not the same as truth (sources can be wrong too)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

Techniques to ensure model outputs are derived from provided context or retrieved documents, with mechanisms to trace claims back to sources. Often involves RAG plus citation generation.

Sources: Google grounding documentation, RAG with citations patterns, LlamaIndex citation features

Examples

RAG system that cites document chunks for each claim
Model output: 'According to [source], the revenue was $X'
Google Vertex AI grounding with Search or custom data
Refusing to answer when no sources support the query

Counterexamples

Things that might seem like Grounding but are not:

Model generating from parametric knowledge alone
Model making claims without citation
Model hallucinating a citation

Relations

inTensionWith hallucination (Grounding reduces hallucination)
overlapsWith retrieval-augmented-generation (RAG is a grounding technique)
overlapsWith faithfulness (Grounded outputs should be faithful)

Implementations

Tools and frameworks that implement this concept:

Google Gemini secondary
RAGAS secondary