Comparisons
Retrieval-augmented generation vs a system of record
Retrieval supplies context for an answer; a system of record governs which values may be treated as settled.
The distinction
Retrieval-augmented generation (RAG) selects material relevant to a question and places it in a model's context before generation. It is useful when a person needs a concise answer across a changing collection of documents. A retrieved passage can support an answer, but retrieval rank does not establish whether its assertion is current, authoritative, or consistent with other passages. The answer is an output of a particular query and model run, not necessarily a durable organizational decision.
A system of record maintains the accepted state of a subject under a defined scope and governance policy. It must make corrections, effective dates, conflicting evidence, and responsibility for a decision intelligible to later users. It is appropriate when an organization needs the same value to survive beyond one question and be defensible in a subsequent review. The two approaches can connect: RAG can explain or find recorded decisions, while the record supplies a controlled answer and its supporting evidence.
By dimension
| Dimension | Retrieval-augmented generation | System of record |
|---|---|---|
| Primary output | A generated response to a query. | A durable, governed value or state. |
| Selection criterion | Retrieval relevance to the query. | Acceptance under a stated decision policy. |
| Conflicting sources | Contradictions may appear in retrieved context and need explicit handling. | Conflicts remain visible alongside the accepted state or an unresolved status. |
| Time | Results depend on the indexed corpus and retrieval time. | Effective and recorded times can be retained for reconstruction. |
| Repeatability | A changed index or model can change an answer. | A dated state can be queried and reviewed again. |
| Failure mode | A plausible answer can exceed what retrieved evidence supports. | A stale decision can persist if source changes are not reviewed. |
| Best unit of review | The answer and passages retrieved for that query. | The accepted value, its evidence, and its history. |
Choose Retrieval-augmented generation when
- A reader needs exploratory answers across a large document set.
- Questions vary more quickly than a governed record can be curated.
- The output can be checked by a person before use.
Choose System of record when
- A value controls a repeatable operational decision.
- Reviewers must reconstruct what was accepted at an earlier date.
- Conflicting evidence needs a durable disposition.
Using both
Retrieve passages to help users investigate and explain a recorded value; let the governed record, with its own citations and dates, remain the source for decisions.
A worked example
A fictional medical-affairs team at Halden Therapeutics AG wants to know the labelled storage temperature of one of its products. With retrieval-augmented generation, an analyst asks the question in plain language, the system retrieves passages from the label, a supplier sheet and an old internal memo, and a model writes an answer citing whichever passages ranked highest. The answer reads well. If the memo ranks first and predates a label change, the answer is current in tone and out of date in fact.
With a system of record, the storage temperature is a recorded value for that product, attached to the label sentence it was read from, with a reading date and with the earlier memo value shown as superseded. The analyst gets one answer, the evidence for it and the history. Retrieval remains useful to explain the label or to find where a change was announced, but the value used in a decision comes from the record.
Where each one fails
Retrieval-augmented generation fails quietly. Retrieval can miss the decisive passage, rank an outdated one above a current one, or return several copies of one error. The model can then add an inference the passages do not support. Research on long inputs shows that models use information unevenly depending on where it sits in the context, and surveys of retrieval-augmented systems treat retrieval, augmentation and generation as stages that each need their own evaluation. The result is that a confident answer can be wrong in ways that are hard to see from the answer alone.
A system of record fails differently. Its scope is limited to what has been curated, so a question outside that scope has no answer. Its values can lag behind the world if the sources change and no one rereads them. And a poorly governed record can carry a wrong value with a respectable citation. The failure modes are slower and more visible, because a stale value has a reading date that shows its age.
Questions to settle before choosing
Does the answer feed a decision that someone will later need to defend? If so, the value needs a stable home. Will the same question be asked again in the same way? A repeated question is a sign that the answer should be recorded rather than regenerated. Who is the reader: a person exploring, who can read and judge the passages, or a system and a committee that act on the value? How often do the sources change, and how will the record learn of it? Finally, what happens when the answer is unknown? A retrieval system will try to answer. A record can say that nothing was found, and when it looked.
How the two meet in a review
A reviewer who finds a questionable generated answer should be able to compare it with the record: does the recorded value agree, and if not, which is right? The comparison is also an evaluation method. Questions with recorded answers make a test set for the retrieval system, and disagreement between the two points to either a retrieval failure or a record that needs updating. Neither is a substitute for the other, and the combination is stronger than either alone.
See also
- Long context vs retrieval
- Data catalog vs system of record
- Auditing an LLM answer against its sources
- Citing at the sentence
- Grounding
- Context engineering
Further reading in the articles
References
- Lewis et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS.
- Gao et al. (2023). Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv:2312.10997.
- Gao et al. (2023). Enabling Large Language Models to Generate Text with Citations. arXiv:2305.14627.
- Es et al. (2023). Ragas: Automated Evaluation of Retrieval Augmented Generation. arXiv:2309.15217.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Which do you run today: Retrieval-augmented generation or System of record?
Describe how your organization handles this now and one entity where it falls short. We reply with where a first record would sit alongside it and what it would need from you.
Or write to [email protected] in your own words.