In-context learning adapts a model through information placed in its prompt rather than by changing internal weights. In retrieval-augmented generation, retrieved passages become that information: search selects relevant content, and prompt construction supplies it to the generator. section 1.3.4
A retrieved passage changes the answer
A BART-based generator receives a question together with a retrieved document, combined simply by concatenating the passage before the question. The model itself remains unchanged: replacing the passage can change the answer, while updating the corpus updates the knowledge available at query time. section 4.1.2
Use in-context learning when the system needs to connect a general model with changing or specialized information:
- Prefer retrieved, high-value context over placing an entire large collection in the prompt; excessive context can obscure relevant facts and produce lost-in-the-middle effects. section 1.3.4
- Treat retrieval, ranking, filters, and task decomposition as explicit reasoning structure supplied by the developer rather than as changes to the model. section 1.3.4
- Distinguish this approach from fine-tuning: in-context learning changes the prompt contents, whereas fine-tuning changes model parameters. section 1.3.4


