An answer can be incomplete without being obviously wrong
A policy may state a general rule in one section and an exception in another. A newer revision may replace both. Retrieving only the paragraph with the closest wording can leave an answer missing the condition that matters.
Preserve source documents, passage identifiers, versions, and dates. These fields let the application connect a useful sentence to the larger context in which it should be interpreted.
Give source relationships a meaning
A source can support a claim, qualify it, or supersede an older source. Treat these as distinct relationships when your workflow records them. An undifferentiated link loses information that a reader needs to understand the evidence.
The model does not have to cover every possible argument. Start with the source relationships your application already knows and make them available to retrieve and inspect.
Retrieve the surrounding evidence
AI Search finds passages relevant to the question. Graph retrieval can include linked exceptions or revisions where those relationships are represented. The application decides which scope and source versions are eligible.
Pass source identifiers through to answer generation and presentation. A citation should lead to the passage supporting the statement, not merely to a document that happens to discuss the same broad subject.
Keep disagreement visible
If sources conflict or a required fact is missing, preserve that uncertainty in the application’s response. Retrieval provides evidence; it does not turn an incomplete source set into certainty.
Evaluate questions that require an exception or revision, not just straightforward lookups. Inspect whether the returned context contains the qualifying passage before judging how fluently a model summarizes it.
Explore the next step
Continue with the interactive workflow. For the current setup and API contract, use the Polygres documentation.