December 2, 2023 7:00 PM PST
This meeting focused on the Amazon Q - AIGC Assistant, a tool designed to access and summarize internal company information across various resources such as ERP, CRM, and customer service systems. The presentation covered system design, implementation details, and comparisons with other services.
Presenter: G, Engineering Manager
System Design Presentation - Amazon Q
- Most search and chat products access only public information.
- Amazon Q accesses information internal to a company, summarizing answers across different resources.
Key Features
- Users can create multiple chatbots, each specializing in one task.
- Application setup includes service role creation and adding data sources.
- Data sources can be uploaded documents, SharePoint, and other systems.
- Secure access: only customers can access their data; Amazon employees cannot.
- Users can look up the original source through a link and upload files on the fly for querying.
Data Source Management
- Adding data sources involves:
- Uploading documents.
- Adding secrets for authentication.
- Setting up retrieval methods.
- Configuring VPC (Virtual Private Cloud).
- Defining sync scope, sync mode, and sync schedule.
- Sync now feature allows immediate synchronization.
Security and Access Control
- Access control settings determine document visibility:
- Turning off access control exposes all documents to all employees.
- Turning on access control restricts access based on permissions.
- Implementation of web crawlers and proxies is discussed.
Implementation Details
- No fine-tuning based on customer data is required.
- The system can handle new data without retraining the model.
- Comparison with Kendra:
- Kendra is an enterprise search service.
- The difference between Kendra and Q is likened to Bing vs. ChatGPT.
Additional Considerations
- Cloud Watch can be used to monitor activities and respect robots.txt.
- External website support is available but not recommended.
- Consolidation of multiple data sources is possible.
- Document size limitations and retrieval augmented generation capabilities were briefly mentioned.