A useful answer is easier to trust when people can inspect where it came from and what the system did to produce it. For a retrieval-augmented generation (RAG) system, transparency means attaching the sources behind a response. For an agent, it can also mean showing intermediate actions, such as a planned weather-service lookup. These details let users check an answer and spot when the system needs correction. section 3.3.5 section 6.4.4
Make the relevant evidence visible at the point of use:
- Return source information with each generated answer so users and developers can fact-check it and identify retrieval gaps. section 3.3.5
- Show understandable agent steps, including which tool is being used and why, so users can check whether the action matches their expectations. section 6.4.4
- Where possible, provide links to original documents or a way to view them in full. section 4.5.4
Policy answers with sources
A health-care policy assistant retrieves relevant policy chunks, places them in the LLM’s context, and asks for an answer based only on that context. Its response includes both the answer and the policy sources. If retrieval finds nothing, it reports that it could not find specific information rather than presenting an unsupported policy answer. A low temperature of 0.2 encourages answers tied to the retrieved information; it does not make the sources complete or the answer automatically correct. section 4.5.4 section 3.3.5
Figure 6.7 Example: Perplexity displays the steps the agent used to answer the question as well as citations.
The interface example shows agent steps alongside citations, making both actions and supporting sources available for inspection. section 6.4.4
Transparency is a way to inspect and challenge a result, not a guarantee of accuracy. Documents may be outdated or incomplete, retrieval may miss a key document, and an LLM may hallucinate despite correct context. Keep oversight, guardrails, and evaluation in place. In production, transparency also belongs alongside monitoring, feedback loops, fairness, privacy, and governance. section 3.3.5 section 1.4.3
