Question

A financial services company is building a RAG application using Amazon Bedrock Knowledge Bases with Anthropic Claude as the generation model. During testing, they notice that when the RetrieveAndGenerate API returns answers for broad questions, the responses are frequently truncated and sometimes omit relevant information that engineers confirmed exists in the ingested documents. The knowledge base is configured to retrieve the top 15 chunks per query, and each chunk is roughly 500 tokens. The team wants complete, accurate answers without re-architecting the ingestion pipeline. What is the MOST effective adjustment to resolve this issue?

Accepted Answer

Reduce the number of retrieved results (numberOfResults) and enable a reranking model so only the most relevant chunks fill the model's context window

Answer

Increase the chunk size during ingestion to 1,500 tokens so fewer chunks are needed to answer each query

Answer

Switch the embedding model to one with higher vector dimensions to improve retrieval recall

Answer

Enable hybrid search combining semantic and keyword matching to retrieve more candidate chunks