Question

A biotech company builds a RAG assistant over 50,000 internal research papers dense with genomics terminology and chemical nomenclature. Using Amazon Bedrock Knowledge Bases with a general-purpose embedding model, retrieval quality is poor: relevant papers containing the exact scientific terms in user queries are frequently not returned in the top results, even though keyword matches clearly exist in the corpus. Chunking and vector dimensions have already been tuned. What is the MOST effective way to improve retrieval accuracy for this specialized domain?

Accepted Answer

Enable hybrid search in the Knowledge Base to combine dense semantic vectors with sparse keyword-based retrieval so exact domain-term matches are captured alongside semantic similarity.

Answer

Increase the number of chunks returned (topK) to a much larger value so more candidate documents are always included in the context window.

Answer

Switch the response-generation foundation model to a larger model with a bigger context window to compensate for weak retrieval.

Answer

Disable semantic chunking and revert to fixed-size chunking to preserve more raw text per chunk.