bedrock-rag

RAG on Bedrock Knowledge Bases

Activate this skill when building retrieval-augmented answering on AWS. For the framework-agnostic, self-hosted-vector-DB flavour, see the data-ai preset's RAG material; this skill is the AWS-native path.

When to Use

• A chatbot must answer from internal documents with citations
• Choosing a chunking strategy, embedding model, or vector store
• Wiring retrieval into a Strands agent as a tool
• Diagnosing "the answer is wrong" in a RAG system

The Pipeline

S3 source → StartIngestionJob → chunk → embed → vector store
                                                     ↓
user query → (optional rewrite) → Retrieve → rerank → prompt + context
                                                     ↓
                          generate → citations → output guardrail → user

Most RAG failures are retrieval failures, not generation failures. Before touching the prompt, check whether the correct chunk was retrieved at all. If it was not, no amount of prompt engineering fixes it.

Chunking

The single highest-leverage config choice.

Preserve structure in the chunk text — a heading path prepended to each chunk materially improves retrieval on technical corpora. Attach metadata (source document, section, last-modified, access tier) at ingestion; metadata filtering at query time is how you enforce per-user document access without a second index.

Retrieval and Generation

Retrieve when the agent should reason over the results:

resp = bedrock_agent_runtime.retrieve(
    knowledgeBaseId=KB_ID,
    retrievalQuery={"text": query},
    retrievalConfiguration={"vectorSearchConfiguration": {
        "numberOfResults": 10,
        "overrideSearchType": "HYBRID",   # semantic + keyword
        "filter": {"equals": {"key": "access_tier", "value": user_tier}},
    }},
)