Anchoring answers in evidence
Grounding means generating an answer from retrieved source documents rather than from the model's fuzzy memory. Citation attaches those sources to the claims, so a reader can verify them.
The pipeline
- Retrieve relevant passages for the query, often with a vector search.
- Place them in the context and instruct the model to answer only from the provided sources.
- Ask the model to attach the source for each claim it makes.
Why this helps safety
- It shifts the model from recall to reading comprehension, reducing hallucination.
- Citations make answers auditable, so errors are traceable to a source.
- When no source supports a claim, a grounded system can abstain.
Pitfalls
- The model may still hallucinate a citation or misattribute a claim, so citations must be checked.
- Bad retrieval poisons the answer. Garbage in, confident garbage out.
- Faithfulness, whether the answer actually follows the sources, must be measured separately from relevance.
Key idea
Grounding answers in retrieved sources with verifiable citations turns recall into reading comprehension, reducing hallucination and making answers auditable, provided retrieval quality and citation faithfulness are checked.