Relevance and confirmation answer different questions
A search result can answer, “Where is the policy for supplier invoices?” A source-record query can answer, “Which supplier belongs to this identifier in this company?” A post-write read can answer, “Does the draft now contain the intended values?” All three may be necessary, but they are not interchangeable.
The distinction is about the claim and evidence path, not whether a component is labeled retrieval. A retrieval system may access an authoritative source directly. Conversely, a database query may read a stale replica or the wrong tenant. The label alone establishes neither authority nor freshness.
A perfectly retrieved document can be insufficient
Imagine a search returns yesterday’s purchase-order export with the exact line items referenced by an invoice. The excerpt is relevant, accurately quoted, and internally consistent. This morning the order was amended. The old export cannot establish the current quantity available to bill.
A citation helps identify the evidence used; it does not upgrade the evidence’s freshness. The workflow needs an explicit policy for which source and version can authorize the action. If only the old export is available, the limitation belongs in the result.
Confirming a write requires a new observation
The context used to construct a bill cannot prove that the later write succeeded. Confirmation needs evidence from the operation itself or an authoritative observation of its effect. An assistant repeating the intended values is not that observation.
Read-back design also needs to account for consistency and identity. The integration should retrieve the record associated with the original operation and compare the relevant fields. Finding a similar bill by text search is weaker than confirming the exact saved record.
Use retrieval without asking it to prove too much
Retrieval remains valuable for discovering context, resolving terminology, locating policies, and assembling evidence candidates. The next step is to qualify those candidates: identify their source, version, access scope, observation time, and applicability to the contract.
This adds data-model work and sometimes an extra system query. The benefit is a receipt that explains the chain from source to check to effect. The result can remain useful even when it is incomplete, because the operator can see exactly which claim lacks sufficient evidence.
Missing a detail or found a problem?
Send a documentation question →