Claim answers can be scattered across adjuster diaries, repair estimates, police reports and payment ledgers. AWS’s conversational claims assistant guide uses Amazon Bedrock Knowledge Bases to retrieve information from those files and return natural-language answers with citations. Its implementation combines documents in Amazon S3, metadata filters, multi-turn follow-ups and the AgenticRetrieveStream API. AWS’s guide describes the pattern.
For developers and cloud architects, the useful detail is how those pieces work together. Document ingestion makes the records searchable; metadata narrows the records considered; agentic retrieval plans searches around the question; citations connect the response to its evidence. The supplied research supports that architecture, while the reported evaluation remains limited to a synthetic corpus and automated grading.
Start with the claim documents
AWS describes Amazon Bedrock Knowledge Bases as a managed retrieval-augmented generation capability for documents. Its guidance says Bedrock handles parsing, chunking, embeddings and vector storage, with claim documents indexed from Amazon S3. This gives the conversational application a retrieval layer over the files from which it must answer. AWS’s technical guidance sets out these responsibilities.
The documents matter because a claim is represented by several kinds of record. AI Briefs’ account of the guide lists diary entries, repair estimates, police reports, payment ledgers and scanned attachments. It gives example questions about whether a claim was approved and which open auto claims exceed $10,000. One question concerns a particular claim’s history; the other combines several conditions across claims.
According to Hello Marvis AI Today’s summary, the setup handles PDFs, Word files and text notes. It pairs each document with a metadata sidecar file. That pairing supplies attributes for filtering while leaving the document’s narrative available for retrieval.
Make metadata comparable
Metadata filters can restrict documents by claim ID, claim type, status, amount and date filed, according to Hello Marvis AI Today. For a question about open auto claims above a monetary threshold, these attributes provide the structured conditions that narrow the search. The document text supplies the detail needed to explain the result.
AWS’s metadata guidance includes three concrete constraints. Sidecar files are limited to 10 KB. Dates should be stored as YYYYMMDD integers. The amount field should contain only one comparable monetary value. The AWS guidance also says reserves, payments and holds should remain in document text so their labels stay clear.
Those rules make field definitions an implementation decision. An amount filter needs a consistent meaning across the documents it compares. Combining different financial concepts in that field would make the resulting selection difficult to interpret. Keeping the separately labelled values in the text preserves the context an answer needs when explaining a claim’s financial position.
The practical sequence is to settle those attribute definitions, pair documents with their metadata and then use filters to scope retrieval. This is an architectural sequence rather than a runnable configuration recipe: the available evidence does not supply a complete request schema, sidecar naming convention or API code example. Developers should not infer those details from the field names alone.
What AgenticRetrieveStream adds
AWS describes AgenticRetrieveStream as planning an answer, breaking a multi-part question into sub-queries and running one or more retrieval passes. It checks whether the evidence is sufficient before generating a response. The API streams trace events, answer text and citations. AWS’s description supports these capabilities.
This makes the retrieval process part of answering the question. When a question has several parts, the system can search for the evidence each part requires rather than depend on a single retrieval pass. The sufficiency check is another step before generation. It addresses whether the retrieved material supports a response, although the supplied evidence does not establish that the check succeeds on every question.
Adjusters and policyholders can use the cited claim documents to check the evidence behind an answer.
AWS’s guide also covers multi-turn follow-ups. The evidence does not specify session-management syntax or context-retention behavior, so it supports the conversational capability without establishing a particular implementation of conversation state.
Check grounding
Contextual grounding guardrails are another component of the AWS pattern. The supplied research does not specify their behavior or demonstrate that they eliminate fabricated answers in every case.
Read the evaluation within its limits
According to Amazon, as reported by Hello Marvis AI Today, its Bedrock Knowledge Bases setup with AgenticRetrieveStream achieved 90.5% retrieval recall and 81.2% citation recall on 40 test questions. The evaluation used a 30-document synthetic corpus and automated grading rather than human review. The evaluation summary supplies the figures and these limitations.
The same source reports Amazon’s findings as directional and says human review on a real corpus is needed before exposing the assistant to policyholders. For an implementation team, that sets a concrete next step: review retrieval and citation quality against its own claim files. The guide supplies a document-querying architecture; the reported synthetic evaluation leaves its suitability for a particular insurer’s records to be established.
