• workshop_dense_100#slide 48

In RAIDT, a retrieval query is the recorded search request issued by the retrieval component during a specific run in order to obtain external context for generation. It is not identical to the user prompt. The user prompt expresses the task presented to the system; the retrieval query records how the system translated that task into a search over a defined corpus or index. Foundations makes this explicit by stating that, for retrieval augmentation, the run record should capture query or corpus identifiers and preserve a snapshot of retrieved passages with document identifiers and hashes. Evidence Review likewise argues that reconstruction of a grounded run requires retrieval queries, retrieved sources, and the state of retrieval indexes.

An example in RAIDT terms would be a public-service assistant that receives a staff question about eligibility and issues a query against the approved policy corpus to retrieve the governing clauses for that case. Recorded properly, the query becomes part of the run-level evidence pack alongside the corpus identifier, retrieved passage IDs, and hashes. It matters because RAIDT uses a score profile to judge whether governance claims are supported by inspectable evidence. The retrieval query shows what source space was searched and how the system sought grounding for that specific output. That directly supports Auditability and Traceability within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), and indirectly supports Interpretability because reviewers can see whether the answer rests on a sensible search strategy. In anchors 1=missing / 3=partial / 5=audit-ready terms, the retrieval query is one of the fields that prevents grounding from collapsing into an unverifiable citation claim.

An HR team uses a GenAI assistant to draft a manager's performance appraisal, drawing on internal policy before the manager signs off the text. The employee later disputes a sentence that appears to invoke policy. If the run-level evidence pack contains only the generated paragraph and a vague citation to the handbook, reviewers still cannot tell how the policy text was found or whether the search framed the issue appropriately.

If the recorded retrieval query shows that the system searched the appraisal policy for conduct thresholds, review periods, and appeal wording, the HR reviewer can inspect whether the system searched the correct policy space and whether the search itself may have steered the output. That makes the use of retrieval reviewable rather than assumed.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10