The Citation Bug a Better Prompt Could Not Fix

Text Labels Are Weak Identifiers

If the prompt gives the model several passages and asks it to reproduce their human-readable references, the answer depends on copying accuracy. Similar labels, ranges, transliteration, and collection names make errors easy to introduce and difficult to validate.

Stable internal IDs change the task. The model can cite a constrained token such as a source index or identifier. The application then resolves that token to the canonical reference, Arabic, translation, grade, and URL already held in the retrieval response.

Validation Belongs Outside the Model

A prompt can ask the model not to invent citations. It cannot prove that the model complied. The server can. Every citation returned by the model should map to an allowed source ID from the current request. Unknown IDs should fail validation rather than reaching the user as plausible-looking references.

The same check can verify that cited source blocks exist and that no citation points to material omitted from the answer payload. This does not prove the prose interpreted the source correctly, but it removes an entire class of reference fabrication.

Prompts can request citation discipline. Systems have to enforce it.

Sentence-Level Grounding Is a Separate Problem

A valid source ID only proves that the source was retrieved. It does not prove that the adjacent sentence follows from it. Models can attach a real citation to an unsupported claim, especially when several sources discuss nearby themes.

Teams can reduce this risk by asking for claim-and-source structures, keeping answers concise, and reviewing whether each substantive sentence has direct support. Automated entailment checks may help, but sensitive religious interpretation still requires human judgment and clear scope limits.

The Prompt Becomes Smaller After the System Improves

Once retrieval supplies stable source blocks and the server validates citation IDs, the prompt no longer needs to carry the entire safety architecture in prose. It can focus on tone, scope, and how to use the provided evidence.

This is a useful engineering signal. When a prompt keeps growing to compensate for missing data constraints, move the responsibility into types, validators, and UI. Models are good at language. The application should remain responsible for identity.