Concepts overview
Understand the eight concepts that make Amarsia reliable, flexible, and fast to iterate in production.
Overview
Amarsia is built around eight concepts that cover the full lifecycle of an assistant: building its behavior, grounding it in your own knowledge, giving it real-world tools, proving it works before you ship, and securing and monitoring it in production.
The goal is practical: less engineering overhead, fewer blind spots, and faster iteration when your product needs change.
Core concepts
Assistants
Versioned AI behavior with prompts, tools, knowledge, testing, and deployment in one workflow.
Knowledge base
Your assistant's context layer for documents and business memory, processed for high-answer quality.
Actions
Isolated tools your assistant can execute, built from templates or imported via MCP.
Test cases
Saved scenarios you can replay on every change to catch regressions before users do.
Evaluations
Score every run against the metrics you care about and learn what is regressing.
Usage
Detailed logs, traces, and token metrics so you can debug, improve, and control costs.
API key
The credential used to authenticate calls to assistants that require authentication.
Security
Choose between API-key authentication and a domain allowlist for public, in-browser assistants.
Outcome-driven view
| What you need | Concept that solves it |
|---|---|
| Reliable AI behavior you can update without app rewrites | Assistants |
| Better factual responses from your own content | Knowledge base |
| Real-world actions like API calls, database queries, and notifications | Actions |
| Confidence that changes didn't break existing behavior | Test cases |
| Measurable quality scores for every production run | Evaluations |
| Visibility into failures, reasoning, and tool runs | Usage |
| Secure access from your product or backend | API key |
| Safe public access from browser apps you own | Security |