liveBook · Manning

Grounding

As explained in LLM Evaluation and Alignment, The Foundational Ideas

A grounding instruction directs a model to base its response on specified source information rather than unsupported knowledge.

Grounding directs a model to base responses on specified source information instead of unsupported knowledge. A grounded response cites relevant sources, acknowledges missing information, and avoids plausible-sounding inference. section 8.5.1

Implement grounding with explicit, testable requirements:

  • Cite the source document and location for every numerical claim. section 8.5.1
  • State explicitly when retrieved context does not contain the requested information. section 8.5.1
  • Use only information from provided documents and avoid substituting parametric knowledge. section 6.4.1
  • Preserve entity attribution and calibrate confidence to the strength of the retrieved evidence. section 8.5.1

Financial research assistant

A domain specification can require citations for every numerical claim, prohibit uncited claims, and require hedging such as “the report indicates” when retrieved context supports the statement indirectly. section 8.5.1

Table 6.4 Interpreting metric combinations for RAG diagnosis
Faithfulness
Answer relevance
Recall
Precision
Diagnosis
High
High
High
High
Working well
Low
High
High
High
Hallucination beyond retrieved context
High
Low
High
High
Off-topic generation. Misaligned.
High
High
Low
High
Retrieval missing relevant documents
High
High
Low
Low
Retrieval is fundamentally misaligned with the query
Low
Low
Low
Low
Systemic failure

Use metric combinations to distinguish generation failures from retrieval failures; low faithfulness with strong retrieval points toward stronger grounding instructions. section 6.4.1

How LLM Evaluation and Alignment, The Foundational Ideas covers it

How deep it goes
Defines testable response clauses and diagnostic prompt interventions, but does not detail a complete grounding implementation.
Examples
Uses financial numerical claims, source paragraphs, named-entity attribution, hedged language, and explicit document-only prompt constraints.
What you can do after
Operate/configure system prompts that require citations, document-only answers, explicit gaps, faithful attribution, and calibrated confidence.
How it connects
Connects grounding with retrieval-augmented generation, retrieved context, hallucination, faithfulness, citation, and confidence calibration.

Also explained in

Compare how 13 books explain grounding →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Evaluation and Alignment, The Foundational Ideas. The section links point to the parts of the book it draws on.