agent

@savoir/agent

Framework-agnostic agent core for Knowledge Agent Template. Handles question routing, prompt management, agent creation, and AI tool definitions.

See also: Main README, Architecture, Customization

Overview

This package provides the AI agent infrastructure used by the chat interface and bots. It is designed to be consumed by apps/app but can also be used independently.

Components

Router

The router classifies incoming questions by complexity and selects the appropriate model and step count.

Complexity	Max Steps	Model
trivial	4	gemini-3-flash
simple	8	gemini-3-flash
moderate	15	claude-sonnet-4.6
complex	25	claude-opus-4.6

import { routeQuestion } from '@savoir/agent'

const config = await routeQuestion(question)
// { complexity: 'moderate', maxSteps: 15, model: 'claude-sonnet-4.6', reasoning: '...' }

Prompts

System prompts for different contexts:

File	Function	Purpose
router.ts	ROUTER_SYSTEM_PROMPT	Question classification
chat.ts	buildChatSystemPrompt()	Chat interface
chat.ts	buildAdminSystemPrompt()	Admin assistant
bot.ts	buildBotSystemPrompt()	Bot responses
shared.ts	applyAgentConfig()	Config overlay (style, language, citations)

Agents

Pre-configured agent factories:

• createAgent() — Base agent for chat and bot responses
• createAdminAgent() — Admin assistant with stats/management tools
• createSourceAgent() — Source-specific agent

Tools

• webSearchTool — Web search tool for finding information not in the sandbox

Types

Key types exported:

import type {
  AgentConfigData,
  ThreadContext,
  RoutingResult,
  AgentExecutionContext,
  CreateAgentOptions,
  AgentCallOptions,
  AgentConfig,
} from '@savoir/agent'

Customization

To customize AI behavior:

1. Admin UI — Change response style, language, temperature, etc. at /admin/agent
2. Prompts — Edit files in src/prompts/ for deeper customization
3. Router — Modify src/prompts/router.ts to adjust complexity classification
4. Tools — Add new tools in src/tools/

License

MIT