IDP Accelerator - Monitor (Beta)

Once an IDP Accelerator pipeline is in production, the questions change. Is throughput keeping up? Which document type is failing, and why? What did yesterday cost, and which configuration version or model drove it? Which documents came back with low confidence and are sitting in human review? Answering these from raw CloudWatch logs, Step Functions histories and reporting tables is slow, and it gets slower as volume grows. IDP Monitor puts the answers on one screen. It installs into your existing IDP Accelerator stack as a subscription feature and adds a Monitoring tab to the IDP Web UI, secured by the same Cognito user pool as the rest of the application.

The dashboard is built from a library of widgets you can turn on and arrange per dashboard. Operations widgets cover live in-progress status, completed volume over time, per-stage latency with percentiles, a failures table with error details, throttle and rate-limit events, workflow concurrency and capacity, low-confidence alerts, human-in-the-loop queue tracking, document type distribution, and the configuration versions in use across recent documents. Cost widgets show spend over time, cost by document type, by configuration version, by pipeline stage, and by Bedrock model, alongside the cost of the pipeline's own Lambda and control-plane components and a ranked list of your most expensive documents. Time ranges run from the last hour to the last 30 days or a custom window, and every chart drills into the underlying documents. Data comes from hourly and daily rollups queried through Amazon Athena over the accelerator's reporting bucket, with DynamoDB for in-flight state, so the dashboard stays fast on large volumes without adding load to your processing path. A collapsible AI Summary widget uses Amazon Bedrock to write a health narrative from the current metrics; it stays collapsed and costs nothing until you expand it, and its model, categories, level of detail and custom instructions can be tuned per dashboard.

The AI Assistant turns the dashboard into a conversation. Ask a question in plain language, such as why costs rose last week or which document types are throttling, and a purpose-built agent running on Amazon Bedrock queries your Athena tables, answers in text, and assembles an AI Insights view from the widgets most relevant to the question, with suggested follow-ups. Views you find useful can be saved as named dashboards. Any question can also be scheduled: the assistant runs it on a cron schedule you choose, stores the results, and raises an in-app notification when they are ready, so a daily cost check or a Monday-morning failure review happens without anyone logging in. Dashboard configurations are versioned, with private and shared versions, per-user active selection, and side-by-side comparison. Everything runs inside your own AWS account and region against your own stack's data. No document content, metrics or configuration leaves it.