AWS Publishes Metadata Workflow That Escalates Ambiguous Fixes to Bedrock LLMs

AWS has published a deployable metadata-harmonization workflow that separates two jobs often blurred together: interpreting a dataset’s schema and fixing individual field values. The sample uses Bedrock’s LLMs for semantic schema alignment, while its field-correction layer begins with rules and similarity techniques before escalating uncertain cases to LLM resolution. AWS also documents both a review-led workflow and an MCP-connected agent that can autonomously apply corrections.

The code and deployment instructions are available through an AWS Samples repository. The stack uses Bedrock for LLM-powered alignment and recommendations, S3 for storage, DynamoDB for job tracking, Cognito for authentication, and ECS for compute. Its CDK deployment can provision the core storage, tracking, authentication, and Fargate resources.

Different jobs for rules and models

Schema alignment checks whether incoming columns correspond to the expected structure. AWS says its LLM approach can recognize domain-specific synonyms, infer meaning from surrounding columns, and identify cases where a source column should be split or combined. Field validation then checks required values, controlled vocabularies, and regular-expression format rules, producing error reports that identify each problem’s location, type, and nature.

The field-correction ladder

• Fuzzy matching handles typographical, spacing, and punctuation inconsistencies.
• Embeddings compare semantic similarity, while contextual inference uses nearest-neighbor similarity, TF-IDF representations, and co-occurrence patterns within the uploaded dataset.
• LLM-based resolution is a fallback when earlier methods do not meet a sufficient confidence threshold.

The agent route changes the approval question

The review flow sends uploaded metadata through parallel schema-alignment and field-validation streams before recommendations reach the user. AWS’s agent instructions instead run an MCP-connected command against a live API server. An adopter can use the described review workflow or the documented autonomous agent path; AWS documents the paths without prescribing a handoff between them, so approval rules need to be established for the route being deployed.

Architecture is published; operating results remain local

AWS says it evaluated several embedding models and selected Amazon Titan for this implementation, citing its performance on general and biomedical metadata tasks, commercial availability, and Bedrock compatibility. The post provides no measured accuracy, throughput, cost, or escalation-rate benchmarks. For prospective users, performance on their metadata and under their chosen controls remains a deployment test rather than a published result.