Foundation Model Integration, Data Management, and Compliance
Vector Stores, Retrieval, and RAG at Scale
Design high-quality RAG systems with suitable embeddings, chunking, metadata, vector indexes, hybrid retrieval, reranking, access filters, citations, and performance controls.
Concepts
- Chunk boundaries, embedding choice, metadata, and retrieval parameters jointly determine recall and precision.
- Hybrid retrieval combines semantic similarity with lexical signals and can outperform either technique alone.
- Metadata filtering must enforce tenant and document permissions before retrieved context reaches the model.
- Measure retrieval separately from answer generation using recall, precision, relevance, faithfulness, and latency.
- Decompose compound questions when one embedding cannot reliably retrieve every required evidence set, then preserve provenance through reranking and synthesis.
Exam tips
- Knowledge Bases for Amazon Bedrock provides managed ingestion and retrieval orchestration.
- OpenSearch Serverless vector collections support scalable semantic and hybrid retrieval.
- Changing an embedding model normally requires re-embedding the corpus with compatible dimensions.