A city grants portal must answer eligibility questions from published rulebooks, not from unverified model guesses. Which pattern should the team implement?

RAG so eligibility answers are grounded in published municipal rulebooks is the pattern the city grants portal should implement. Eligibility guidance drives resident decisions about whether and how to apply; ungrounded model guesses can falsely encourage applications or wrongly discourage qualified households. Retrieving from published rulebooks before generation keeps answers aligned with the city's authoritative criteria and supports quotation or citation of governing text staff can defend.

Fine-tuning the model on the municipal rulebooks to embed eligibility knowledge bakes policy into static weights that cannot cite the governing passage and go stale when the city updates the rulebooks. Summarizing the municipal rulebooks to answer eligibility from cached summaries compresses away critical conditions and exceptions and risks serving stale summaries instead of grounding the answer in the current published rulebook text. Widening the context window to cite eligibility answers from municipal rulebooks only increases input capacity and does nothing to retrieve the correct governing passages, so answers remain ungrounded without retrieval before generation.

Exam caveat: high-stakes municipal Q&A—eligibility, benefits, legal criteria—maps to retrieval-augmented generation over published corpora, not fine-tuning on the municipal rulebooks to embed eligibility knowledge.

Operational check: seed the index with current rulebooks, ask eligibility questions tied to specific clauses, and verify answers cite retrieved passages; update a rule and confirm the bot reflects the change after reindexing rather than clinging to stale parametric guesses.