Implementation and Integration

Preprocessing documents to expand abbreviations during ingestion ensures the vector embeddings capture the full semantic meaning. Hybrid search (combining semantic vector search with keyword/lexical search) provides the best of both worlds—semantic understanding plus exact term matching for abbreviations. The 0.7/0.3 weighting favors semantic search while still catching exact matches. This approach maintains low latency since preprocessing happens once at ingestion time, not at query time.

Step Functions provides built-in error handling, retry logic, and the ability to catch timeouts gracefully. You can implement a Map state with error catching that returns partial results when some database queries succeed but others timeout. The workflow can aggregate available results and return them to the agent, while the agent's orchestration continues. This maintains the user experience even during database slowdowns.

This option meets all HIPAA requirements efficiently: KMS customer-managed keys provide encryption at rest with audit trails; Lex and Bedrock both encrypt data in transit by default; CloudTrail data events provide comprehensive audit logs; using base model inference (not fine-tuning) ensures conversation data is never used for training, which is documented and provable for compliance. S3 with SSE-KMS provides secure, cost-effective log storage. This is the standard AWS-recommended pattern for HIPAA workloads with managed services, minimizing operational overhead.

Claude's prompt caching works by caching the prefix of the prompt—content must appear in order with cached content coming first. The brand guidelines (15K tokens) should be wrapped in cache control tags at the start, then variable customer data (500 tokens) appended after. With a 5-minute TTL and 100K requests/day (~69 requests/minute), request batching at 30-second intervals ensures multiple requests benefit from the same cached content, maximizing cache hit rates. This is the documented pattern for Claude prompt caching and provides maximum savings on the expensive 15K token prefix.

This approach provides surgical precision: the Lambda function identifies SKUs using regex (XX-1234-YY pattern), replaces them with safe placeholders (e.g., [SKU_1], [SKU_2]) before sending to Guardrails, then Guardrails PII filters work normally blocking actual credit cards, and finally the Lambda unmasks SKUs in the response. This preserves all Guardrails PII protections while preventing false positives on SKUs. It's a common pattern for handling domain-specific entities that superficially resemble PII.

This is the correct pattern for supervisor-subordinate multi-agent systems in Bedrock. The supervisor agent uses action groups to invoke specialized agents via Lambda functions, which can call the specialized agents' APIs. The return control invocation pattern allows the supervisor to receive results from subordinate agents, process them (aggregate, handle failures), and continue orchestration. Native session state handles context automatically across the conversation. This uses Bedrock's designed features for multi-agent orchestration and scales to 200 concurrent sessions efficiently.