PRODUCTION SCENARIO
A Generative AI Engineer builds an assistant over 4,000 HR policy documents that the benefits team edits most weekdays. Answers must quote the wording in force on the day the question is asked, and the policies are prose rather than fields.
Which architecture should the engineer choose?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Index the policy files and retrieve the sections that match each question
Retrieval reads the indexed files at query time, so a morning edit reaches the afternoon's answer and the model quotes the passage it was given. A nightly adapter refresh still lags every edit made during the working day, and fine-tuning is unreliable for verbatim recall. A cached system prompt cannot hold 4,000 documents, and invalidating it on every edit destroys the cache it depends on. Delta table columns suit fields and dates, not the prose the answer has to quote.