The most cost effective approach is Implement Retrieval-Augmented Generation by indexing the PDFs in an Amazon Bedrock knowledge base and have the chatbot retrieve relevant passages at runtime.

This approach uses embeddings and a vector index to fetch only the most relevant passages at inference time which minimizes token consumption and latency. It keeps answers current because you update the index or documents instead of retraining the foundation model which avoids repeated fine tuning costs and reduces operational overhead.

Fine-tune the foundation model using text extracted from the brochure PDFs is costlier and slower to update because every brochure change would require retraining which adds compute expense and delays in reflecting new information.

Include the combined contents of all brochure PDFs in every prompt so the model sees everything is impractical because it dramatically increases token usage and latency and it risks exceeding model context windows which becomes expensive and unreliable at scale.

Place Amazon S3 pre-signed URLs to the PDFs inside the prompt so the model can fetch them when needed does not work because models cannot browse or fetch external URLs during inference so the prompt will not actually provide the document content and you will not gain useful context.

Full AWS Practitioner Certification Question