When a user asks whether a product can still be returned, the application may need a public update, a controlled policy document, and the identity of the order mentioned one turn earlier. Calling all three capabilities “more knowledge” hides different data paths and different risks.
Search handles current public information
Web search fits public material that changes often. Keep the result URL, retrieval time, and supporting passage. A high-ranked page may be stale or incomplete, so financial, legal, and security decisions still need source checks.
Model Studio's enable_search option is a platform extension. With the OpenAI Python SDK it is passed through extra_body; it is not a portable OpenAI API field.
RAG retrieves from controlled sources
Retrieval-augmented generation searches a selected knowledge source and provides relevant passages to the model. Internal policies and product manuals fit this design because the application can control scope and updates.
A usable system must track document versions, enforce user access before passages reach the model, evaluate retrieval relevance, cite the original location, and refuse to invent an answer when evidence is absent. Embeddings and vector search are implementation steps, not the acceptance criterion.
Memory maintains user continuity
Conversation history resolves references such as “that order” and preserves user choices. The application normally resends relevant messages or recalls a summary from storage. It also decides what may be saved, how long it remains, and how a user can remove it.
Unlimited history increases token use and false associations. Keep information that supports the current task and discard unrelated material.
Choose by data responsibility
Ask four questions: Is the source public or controlled? How often does it change? Do users have different access? Must the answer cite evidence, or only preserve conversational continuity?
Public announcements point to search. The currently approved refund policy points to RAG. The order selected one turn ago points to memory. One request may use all three, but each item should carry its source, authorization, and lifetime.
Test retrieval, permissions, references, and deliberate no-answer cases separately. A fluent response is not proof that the correct data path was used.