Support Bot

@agent-express/preset-support — a single function call that wires up everything you need for a production customer support agent.

Install

npm install @agent-express/preset-support

Optional adapters (pick what you need):

# Knowledge base search (RAG)
npm install @agent-express/search-llamaindex @agent-express/embed-openai

# Web search
npm install @agent-express/search-brave

# Session persistence
npm install @agent-express/session-sqlite

Usage

import { Agent, search, memory, tools } from "agent-express"
import { supportBot } from "@agent-express/preset-support"
import { llamaindexRetriever } from "@agent-express/search-llamaindex"
import { openaiEmbed } from "@agent-express/embed-openai"
import { braveProvider } from "@agent-express/search-brave"
import { sqliteStore } from "@agent-express/session-sqlite"
import { z } from "zod"

const agent = new Agent({
  name: "support",
  model: "anthropic/claude-sonnet-4-6",
  instructions: "You are a customer support agent for Acme Corp.",
})

agent.use(supportBot({
  fileSearch: search.file({
    retrieve: llamaindexRetriever({
      sources: ["./knowledge-base"],
      embed: openaiEmbed(),
    }),
  }),
  webSearch: search.web({ provider: braveProvider() }),
  sessionStore: memory.store({ backend: sqliteStore() }),
  escalation: tools.function({
    name: "escalate_to_human",
    description: "Transfer customer to human support agent",
    schema: z.object({ reason: z.string() }),
    execute: async ({ reason }) => `Connecting you with a human agent. Reason: ${reason}`,
  }),
}))

What’s included by default

Capability	Middleware	Default
Budget cap	`guard.budget()`	$0.50/session
Turn timeout	`guard.timeout()`	30 seconds
PII redaction	`guard.piiRedact()`	All types (email, phone, CC, SSN, IP)
Tone enforcement	`guard.tone()`	”friendly-professional”
Escalation safety net	Escalation counter	After 5 unproductive turns
Rate limiting	`guard.rateLimit()`	60/min per session

All of these are created automatically inside supportBot(). You just pass config to override defaults.

What you provide

Option	Middleware	Purpose
`fileSearch`	`search.file()`	Knowledge base / document RAG
`webSearch`	`search.web()`	Real-time web search
`sessionStore`	`memory.store()`	Conversation persistence
`escalation`	`tools.function()`	Escalation tool for the model to call

These are optional. Without them, the bot still works — it just won’t have RAG, web search, persistence, or model-driven escalation.

Customization

Every default is overridable. Set any option to false to disable:

supportBot({
  budget: 2.00,                    // override budget
  pii: false,                      // disable PII redaction
  tone: "formal",                  // change tone style
  escalationAfter: 3,              // escalate faster
  rateLimit: { maxPerMinute: 30 }, // stricter rate limit
})

Escalation

Two mechanisms work together:

Model-driven (primary) — your escalation tool. The model decides when to call it based on user intent, frustration, or inability to help. This is the recommended path.
Safety net (fallback) — a turn counter that force-escalates after N unproductive turns (default: 5) if the model doesn’t act. Resets whenever the model calls any tool.

Without an escalation tool, only the safety net works.

Tone Styles

Style	Best for
`"friendly-professional"` (default)	General customer support
`"formal"`	Enterprise, legal, finance
`"casual"`	Consumer apps, social
`"empathetic"`	Complaints, fraud, sensitive issues
`"concise"`	Technical support, quick answers
`"educational"`	Onboarding, tutorials

Example: NeoBank

See the fintech support bot demo for a complete working example with fraud detection, card blocking, RAG from knowledge base, and SQLite session persistence.