Control · C-048 · Model & System Lifecycle
Retrieval grounding and citation
The system answers only from retrieved, approved sources, cites the source for each statement, and declines where the sources do not support an answer.
- Control type
- Preventive
- Family
- Risks addressed
- 8
- Audience
- Practitioner
- Frameworks
- Confabulation · LLM07
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-10
What it does
It replaces the model's memory with the organisation's current documents as the basis for answers, which reduces invention and keeps answers up to date. Citation lets the user check. Engineering owns the mechanism; the business owner decides which sources are authoritative and keeps them current.
How to implement it
- Restrict retrieval to a curated set of sources with named owners.
- Instruct and constrain the model to answer from the retrieved passages only.
- Attach to each statement a reference to the passage that supports it.
- Decline, or refer to a person, when retrieval finds nothing adequate.
- Test that citations actually support the statements they are attached to.
- Review and refresh sources on a schedule, and remove superseded documents.
Evidence it operates
Design: a list of approved sources with owners and review dates. Operation: sampled answers in which each statement is supported by its citation; the rate of declined answers; evidence that a superseded document was removed and is no longer cited.
Where it fits
Built at design and running on every answer. It is a primary control for misinformation and confabulation and for stale model knowledge, and it reduces the load on factuality verification.
Limits
Grounding makes the sources the point of failure: an error or a planted document in the corpus is repeated with a citation. Models still misread or over-generalise from a passage, and can attach a citation to a statement it does not support. A cited answer invites more trust than an uncited one, so citation accuracy has to be measured.
Risks this mitigates
Primary mappings
- Inaccurate personal data in outputs High
- Confines statements to retrieved, approved records. A generative model states false information about an identifiable person as though it were fact.
- Misinformation and confabulation High
- Constrains answers to approved sources and withholds unsupported ones. A generative model states false or fabricated information with the same fluency and confidence as accurate information, and the output is relied on.
- Stale model knowledge Medium
- Supplies current information at answer time. A model answers from knowledge that ended at its training cut-off, on matters such as rates, rules or products that have since changed.
- Ungrounded or misattributed answers Medium
- Requires every answer to be supported by a retrieved source and cites it. A retrieval-based system cites a source that does not support its answer, or presents content the model generated as though it came from a source.
Supporting mappings
- Defamation by AI output High
- Limits statements to sourced material. A system publishes a false statement about an identifiable person or organisation that damages their reputation.
- Degraded information environment Low
- Confines answers to approved sources. The growing volume of synthetic content lowers the reliability of the public and internal information that the organisation and its models depend on.
- Retrieval corpus poisoning High
- Restricts retrieval to approved sources. Content placed in a document store, website or other retrieved source is written to mislead the model, so that authoritative-looking answers carry an attacker's content.
- Vector and embedding weaknesses Medium
- Limits answers to approved retrieved sources. The store of embeddings behind a retrieval system lacks the access control, validation and separation applied to other data stores, so its contents can be read, inferred or altered.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| NIST Generative AI Profile (NIST AI 600-1) | Confabulation | Confabulation Reduces confabulation. |
| OWASP Top 10 for LLM Applications | LLM07 | Misinformation Misinformation. |
Related terms
How to cite
Axiom Verity (2026). Retrieval grounding and citation. Reference Library, C-048. https://axiom-verity.com/controls/retrieval-grounding-and-citation. Reviewed 2026-11-10.
Found an error or a better source? Report a correction.