Grounding

A grounded response is backed by a retrievable source. RAG is the dominant grounding technique · retrieve relevant documents and condition the answer on them. Citations ("according to document X") let users verify. Tool-grounded responses (database query, web search) base answers on real-time external data. Ungrounded output is fluent but potentially fabricated.

Grounding spectrum: full RAG (retrieve + cite everything), partial RAG (retrieve for facts, LLM for synthesis), tool-grounded (model queries structured sources), and reasoning-grounded (model justifies each claim). Quality of grounding depends on retrieval recall and source authority. Enterprise AI increasingly demands citation-level grounding · "show me which document this came from" · for auditability and legal compliance. Perplexity popularized the citation-first UX pattern.

Groundedness metrics: TruLens's groundedness score (does the output's claims appear in retrieved context?), RAGAS faithfulness, Anthropic's context adherence. Target: 95%+ grounding on factual responses. Techniques: extractive grounding (quote source verbatim), abstractive grounding (paraphrase with citation), hybrid (cite key claims, synthesize between). Failure modes: source misattribution, partial grounding (cited but distorted), and ungrounded inferences presented as grounded. Production pipeline: retrieve → rerank → generate with citation → grounded-check → return.