AWS has published a reference design for querying insurance claim files in ordinary language while linking each answer to source evidence. The walkthrough uses Bedrock Knowledge Bases to index PDFs, Word documents and text notes stored in S3, along with metadata for fields such as claim type, status and filing date.
The AgenticRetrieveStream API can split a multi-part question into subqueries and repeat retrieval until it has enough evidence or reaches an iteration limit. It streams the answer, citations and trace events that expose the retrieval plan. Metadata filters can restrict a query to one claim, customer, region or range of amounts.
A contextual-grounding guardrail checks whether the generated response is supported by the retrieved records. That is especially important when revised estimates replace earlier versions or provisional payments are later reversed. Citations let contact-center staff verify a claim before repeating it and give supervisors an audit trail.
The sample uses synthetic records and does not describe a live insurer deployment. AWS warns against putting real personal or health information into the setup without required controls and approvals. Teams would also need access rules that prevent one policyholder’s session from retrieving another person’s files. The design demonstrates the retrieval plumbing, not compliance or answer accuracy for a production claims operation.