A company plans to use Amazon Bedrock LLMs to build a chat interface that references product manuals stored as PDF files. Which approach is the most cost-effective?

Uploading PDF documents to an Amazon Bedrock knowledge base and using it for Retrieval Augmented Generation (RAG) is the most cost-effective approach. Bedrock knowledge bases automatically handle document ingestion, chunking, embedding, and retrieval, providing relevant context to the LLM without requiring the user to manage these complex steps or incur high token costs for passing large amounts of raw text with each prompt. Including a single PDF or all PDFs directly in the prompt context is highly inefficient and costly. LLMs have token limits, and passing entire documents with every prompt would quickly exceed these limits, incur high token usage fees, and degrade performance. Fine-tuning a model is expensive and time-consuming, requiring significant data preparation and training costs, and is generally overkill for simply referencing existing documents; RAG with a knowledge base is more suitable for this use case.