AI AgentsAI / Retrieval-Augmented Generation
RAG Document Assistant
Staff were repeatedly answering the same questions from a large, scattered document set, and a black-box chatbot would not be trusted. 3 Dices built a retrieval-augmented assistant on Amazon Bedrock that grounds every answer in source documents with citations, resolving most routine queries without a human.
Most routine queries resolved without a human
Cited
Every answer sourced
Bedrock
Claude 3.5 Sonnet
Fewer
Repetitive lookups
Auth
Access controlled
The challenge
Staff spent time answering the same questions from a large, scattered document set. Answers needed to be trustworthy, so a black-box chatbot would not do.
- The same questions answered repeatedly by staff
- A large, scattered document set with no single search surface
- Trust requirement: answers had to be verifiable, not black-box
- Access to source documents had to stay controlled
Our solution
Retrieval Pipeline
- Retrieval-augmented generation pipeline on Amazon Bedrock (Claude 3.5 Sonnet)
- Document set indexed for semantic retrieval
- Answers grounded in retrieved passages
Trust & Verifiability
- Every answer cites the source documents it came from
- No ungrounded, black-box responses
Delivery & Access
- FastAPI backend with authentication and rate limiting
- Cognito for identity and access control
- DynamoDB for session and metadata storage
The results
Efficiency
- Most routine queries resolved without human involvement
- Staff freed from repetitive lookups
Trust
- Answers cite their source documents
- Responses are verifiable, not black-box
Control
- Authenticated access with rate limiting
- Identity managed through Cognito
Technologies & AWS services
- Amazon Bedrock
- Claude 3.5 Sonnet
- FastAPI
- DynamoDB
- Cognito
Want Results Like These?
Tell us what you're building. We'll tell you honestly how we'd approach it.