@operor/core

Core framework for Operor — provides the agent runtime, message routing, intent classification, and extensibility points.

Knowledge Base Integration

Agents can opt into knowledge base (KB) powered responses via configuration:

• OperorConfig.knowledgeStore — configure the KB backend (SQLite path, embedding provider/model, dimensions)
• AgentConfig.knowledgeBase: true — opt-in per agent to use KB for context retrieval

const os = new Operor({
  knowledgeStore: {
    dbPath: './my-kb.db',
    embeddingProvider: 'openai',
    embeddingModel: 'text-embedding-3-small',
    dimensions: 1536,
  },
});

os.createAgent({
  name: 'support',
  triggers: ['*'],
  knowledgeBase: true, // this agent uses KB
});

Training Mode

Training mode allows whitelisted phone numbers to manage KB content via chat commands.

const os = new Operor({
  trainingMode: {
    enabled: true,
    whitelist: ['+1234567890'],
  },
});

Available commands (sent as chat messages by whitelisted users):

| Command | Description | |---------|-------------| | /teach <content> | Add knowledge to the KB | | /faq list | List all FAQ entries | | /faq delete <id> | Delete an FAQ entry | | /faq search <query> | Search FAQ entries | | /test <question> | Test KB response | | /status | Show KB status | | /help | Show available commands |

Training commands emit training:command events for external handlers to implement the actual KB operations.

Guardrails

Per-agent guardrails control what messages are accepted and what responses are allowed.

os.createAgent({
  name: 'support',
  triggers: ['*'],
  guardrails: {
    systemRules: ['Always be polite', 'Never share internal pricing'],
    blockedTopics: ['politics', 'religion'],
    escalationTriggers: ['speak to a human', 'talk to manager'],
    maxResponseLength: 2000,
  },
});

• systemRules — injected into the LLM system prompt
• blockedTopics — messages containing these topics are refused (case-insensitive substring match)
• escalationTriggers — phrases that trigger human handoff
• maxResponseLength — caps response character length

The GuardrailEngine class provides checkInput() and checkOutput() methods for use in agent processing pipelines.