Generative AI Patient Document Chatbot for Healthcare Providers

Healthcare organizations store vast volumes of patient documentation, referral letters, discharge summaries, lab reports, as unstructured files, often governed by strict, per-file access controls. Clinicians need fast answers from these records, but conventional search and chatbot solutions rarely respect the existing file-level permissions that regulate who may see which patient's data, creating both a usability gap and a compliance risk.

This offering delivers a production-grade, permissions-aware RAG chatbot built entirely on AWS. Clinicians ask questions in a web interface and receive answers drawn only from the patient documents they are individually authorized to access. Authorization is enforced natively: Windows ACLs set on files in Amazon FSx for NetApp ONTAP (FSxN) are captured as metadata during ingestion and matched against each doctor's Security Identifier (SID) at query time, so retrieval is filtered before any content reaches the language model.

The architecture is organized into three layers. An asynchronous ingestion layer runs an embeddings container on Amazon EC2 that periodically reads file content and ACLs from FSxN, chunks and embeds the text using Amazon Bedrock Titan Text Embeddings, and stores vectors plus ACL metadata in Amazon OpenSearch Serverless (AOSS). A real-time retrieval layer fronts a Streamlit UI with an Application Load Balancer, routes requests through Amazon API Gateway to a Retrieval Lambda that performs SID-filtered vector search, augments the prompt with authorized context, and generates answers using Anthropic Claude models on Amazon Bedrock, while persisting conversation history in Amazon DynamoDB. A storage layer built on FSxN, integrated with Managed Active Directory, holds the patient documents and their ACLs as the single source of truth for both content and permissions.

Because access control is inherited directly from existing Windows ACLs, administrators manage permissions using familiar tools; updated ACLs are reflected in the chatbot's behavior after the next embedding sync, with no re-architecting required. The solution was developed and validated using synthetic patient documents, enabling rigorous testing without exposing real patient data, and is designed to support regulated healthcare environments.

Engagements delivered through this offering include architecture design and deployment via Terraform, Active Directory and FSxN configuration, embedding and retrieval pipeline setup, and validation of ACL-enforced access, taking clients from proof of concept to an auditable, production-grade deployment.