An AI team is organizing which challenge each Bedrock feature should be applied to. Which is the correct mapping for how to use RAG (knowledge bases), agents, guardrails, and model evaluation?
A question about choosing the way to use Bedrock features by purpose.
RAG for referencing internal information, agents for multi-step processing, guardrails for safety control
Correct. Features are used according to their purpose. RAG is suited to answers based on the latest internal information, agents to multi-step processing that uses external APIs, guardrails to safety control, and model evaluation to quality measurement.
Guardrails for referencing internal information, RAG for safety control, agents for quality measurement
All the mappings are swapped. Referencing internal information is the role of RAG (knowledge bases), safety control is the role of guardrails, and quality measurement is the role of model evaluation.
Model evaluation for multi-step processing, agents for quality measurement, RAG for safety control
The mappings are swapped. Multi-step processing is the role of agents, quality measurement is the role of model evaluation, and safety control is the role of guardrails.
Knowledge bases for safety control, model evaluation for referencing internal information, guardrails for multi-step processing
The mappings are swapped. Safety control is the role of guardrails, referencing internal information is the role of knowledge bases (RAG), and multi-step processing is the role of agents.
The basic idea is to use Bedrock features by purpose.
- Answers based on the latest internal information → RAG (knowledge bases)
- Multi-step processing that uses external APIs → agents
- Safety control over harmful output → guardrails
- Measuring model quality → model evaluation
These can also be combined. "Always only guardrails" and "you can use only one" are both wrong.